Ecology

These authors contributed equally: Gábor Pozsgai, Ibtissem Ben Fekih.State Key Laboratory of Ecological Pest Control for Fujian and Taiwan Crops, Institute of Applied Ecology, Fujian Agriculture and Forestry University, Fuzhou, 350002, ChinaGábor Pozsgai, Ibtissem Ben Fekih, Jie Zhang & Minsheng YouJoint international Research Laboratory of Ecological Pest Control, Ministry of Education, Fuzhou, 350002, ChinaGábor Pozsgai, Gábor L. Lövei & Minsheng YouCE3C – Centre for Ecology, Evolution and Environmental Changes, Azorean Biodiversity Group and Universidade dos Açores, Angra do Heroísmo, 9700-042, Azores, PortugalGábor PozsgaiInstitute of Environmental Microbiology, College of Resources and Environment, Fujian Agriculture and Forestry University, Fuzhou, 350002, ChinaIbtissem Ben Fekih & Christopher RensingBasic Forestry and Proteomics Research Center, College of Life Science, Fujian Provincial Key Laboratory of Haixia Applied Plant Systems Biology, Fujian Agriculture and Forestry University, Fuzhou, 350002, ChinaMarkus V. KohnenLaboratoire de Biologie et de Physiologie des Organismes, Faculté des Sciences Biologiques, Université des Sciences et de la Technologie Houari Boumediène, BP 32 El Alia, Alger, 16111, AlgeriaSaid AmraniDuna-Ipoly National Park Directorate, Költő u. 21, H-1121, Budapest, HungarySándor BércesJuhász-Nagy Pál Doctoral School, University of Debrecen, Egyetem tér 1, H-4032, Debrecen, HungarySándor BércesDepartment of Zoology, Plant Protection Institute, Centre for Agricultural Research, Nagykovácsi út 26-30, H-1029, Budapest, HungaryDávid FülöpFujian University Key Laboratory for Plant-Microbe Interaction, College of Plant Protection, Fujian Agriculture and Forestry University, Fuzhou, 350002, ChinaMohammed Y. M. JaberDepartment of Plant and Environmental Sciences, University of Copenhagen, Thorvaldsensvej 40, 1871, Frederiksberg C, DenmarkNicolai Vitt MeylingDepartment of Algology and Mycology Faculty of Biology and Environmental Protection, University of Łódź, Banacha 12/16, PL-90-237, Łódź, PolandMalgorzata Ruszkiewicz-MichalskaDepartment of Molecular Biotechnology and Microbiology, University of Debrecen, Egyetem tér 1, Debrecen, H-4032, HungaryWalter P. PflieglerFujian Provincial Key Laboratory of Insect Ecology, College of Plant Protection, Fujian Agriculture and Forestry University, Fuzhou, 350002, ChinaFrancisco Javier Sánchez-GarcíaÁrea de Biología Animal, Departamento de Zoología y Antropología Física, Facultad de Veterinaria, Universidad de Murcia, Murcia, 30100, SpainFrancisco Javier Sánchez-GarcíaDepartment of Agroecology, Aarhus University, Flakkebjerg Research Centre, Forsøgsvej 1, DK-4200, Slagelse, DenmarkGábor L. Lövei More

Choice of biowaste treatment plants and sample identifiersCompost samples were collected from four central municipal biowaste treatment plants (denominated as #1 to #4) in Baden-Wurttemberg, Germany (Table 1). All plants used a state-of-the-art two-stage biowaste treatment process comprising of (a) anaerobic digestion/biogas production and (b) subsequent composting of the solid digestate to produce a high-quality mature compost sold for direct use as fertilizer in agriculture. The composts were regularly analyzed by an independent laboratory for quality and residual contamination and consistently fulfilled the quality requirements of the label RAL-GZ 251 Gütezeichen Kompost of the German Bundesgütegemeinschaft Kompost e.V. (www.gz-kompost.de). Plants #1 and #3 produce in addition a liquid fertilizer, which is separated from the solid digestate at the end of stage a) by press filtration and which is also intended for direct use on agricultural soil (replacement of liquid manure). In case of plants #1, #3, and #4 up to 25 wt% of shrub/tree cuttings were added to the solid digestate for composting. All plants used sieving (typically with a 12 or a 20 mm mesh) at the end of the process to assure the necessary purity of their finished composts. Whenever technically possible, we as well took samples of the pre-compost immediately before this final sieving step to evaluate its contribution to the removal of residual BPD fragments. For analysis, composts were passed consecutively through two sieves with mesh sizes of 5 mm and 1 mm, yielding two fragment preparations for IR-analysis namely a > 5 mm fraction corresponding to the contamination by residual “macroplastic” (5 mm is a commonly used upper size limit for “microplastic”, anything larger is macroplastic) and a 1–5 mm fraction corresponding to the regulatory relevant residual contamination by microplastic. The lower limit of 1 mm rather than 2 mm was chosen in anticipation of the expected changes in regulation, where the replacement of the 2 mm limit by a 1 mm limit is imminent.Table 1 Technical data of the investigated plants and incidence of BDP fragments in the sampled composts.Full size tableOccurrence of plastic fragments  > 1 mm in the sampled compostsComposting times of 5–9 weeks were used in the investigated plants (Table 1), which is shorter than the 12 weeks indicated in EN 13432 for the 90% disintegration of a compostable plastic material, but a realistic time span for state-of-the-art technical waste treatment. Since we were not in a position to estimate the quantity of BDP entering the plants, since for technical reasons we were unable to obtain a representative sample, we cannot say, whether any residual BDP detected by us in the finished composts was due to a yet incomplete disintegration process or whether it corresponds to the 10% material still permissible by EN 13432 even after the full composting step. However, in 7 out of the 12 sampled composts and pre-composts fragments with chemical signatures corresponding to the BDPs poly (lactic acid) (PLA) and poly (butylene-adipate-co-terephthalate) (PBAT) were identified in the > 5 mm and/or the 1–5 mm sieving fractions using FTIR analysis3 (Fig. 1; Table 1). All recovered fragments appeared to stem from foils, bags or packaging, since they were thin compared to their length and width (see Suppl Figure S1 for typical examples). Fragments with overlapping signatures, most likely PBAT/PLA mixtures or blends, were also found (see Suppl Figure S2 for the interpretation of the spectra). In addition, the recorded BDP fragment spectra (Fig. 1A) showed high similarity to the FTIR spectra of commercial compostable bags sold in the vicinity of the biowaste treatment plants (Fig. 1B), which together with the geometry of the recovered fragments led us to assuming that the majority of the BDP entered the biowaste in the form of such bags.Figure 1FTIR spectra of BDP fragments from composts and commercial bags. (A) BDP fragments recovered from the composts and (B) the commercial compostable bags. Fragments were coded as follows: p or f for pre-compost or finished compost, followed by the plant number (#1 to #4), an indication of the size fraction ( > 5 mm or 1–5 mm) in which the fragment was found, and finally, the fragment number. Fragment F#1_5mm_4 therefore represents the 4th fragment collected in the  > 5 mm size fraction from the finished compost of plant number 1. Bags were arbitrarily numbered 1–10, see Suppl Table S1 for supplier information. The spectra (in grey) of the reference materials for PLA and PBAT are given as basis for the interpretation. Spectra in red refer to test samples consisting only of PBAT, while those in blue indicate samples composed of PBAT/PLA mixtures.Full size imageThe BDP fragments were found alongside fragments of commodity plastics (mostly PE) in all cases. Finished composts tended to contain fewer and smaller fragments than the corresponding pre-composts. The final sieving of the pre-composts to prepare the finished composts hence appears to be quite effective in removing such fragments, in particular those from the > 5 mm size fraction (Table 1) and for that reason has become state-of-the-art in preparing quality composts (contamination by plastic fragments > 2 mm of less than 0.1 wt%). Given that the size of the fragments is a crucial factor regarding ecological risk, we analyzed the sizes (length Î width) of the BDP fragments in comparison to that of the plastic fragments with signatures of commodity plastics such as PE (Fig. 2). BDP fragments found in a given compost sample tended to be smaller than the fragments stemming from non-BDP materials, which may indicate that BDPs degrade faster or tend to disintegrate into tinier particles than commodity plastics. This may also explain why in the compost from plant #2, no BDP fragments were found in the particle fraction retained by the 5 mm sieve ( > 5 mm fraction), while 19 such particles were found in the fraction then retained by the 1 mm sieve (1–5 mm fraction). Interestingly, plant #2 is the only one included in our study that uses no mechanical breakdown of the incoming biowaste. This reduces the mechanical stress on the incoming material. Mechanical stress can alter the properties of plastic foils such as the crystallinity whereby crystallinity has been shown to influence the biological degradation of BDP such as PLA7.Figure 2Size distribution of plastic fragments  > 1 mm. (A) Fragments found in the finished compost from plant #1, (B) in the finished compost from plant #2, and (C) in the pre-compost from plant #3. For reasons of statistical relevance, only samples containing more than 20 BDP fragments per kg of compost were included in the analysis.Full size imageMaterial characteristics of BDP fragments in comparison to those of commercial biodegradable bagsIn order to verify whether the BDP fragments recovered from the composts differed from the compostable bags in any parameter with possible relevance for biodegradation and environmental impact16, the physico-chemical properties of bags and fragments were studied in detail. Since we wanted to have a maximum of information of the BDP fragments, size/weight was a limiting factor in selecting fragments for analysis. Fragments of at least 1 mg were required for the FT-IR analysis. 5 mg-fragments could be analyzed in addition by 1H-NMR, while the full set of analytics (FT-IR, 1H-NMR, and DSC) required at least 10 mg of sample.For insight into the chemical composition, 1H-NMR spectra of the commercial bags and all suitable BDP fragments were compared (Fig. 3). In case of material mixtures and blends, the 1H-NMR analysis allows quantification of the PBAT/PLA weight ratio in the materials and also of the ratio of the butylene terephthalate (BT) and butylene adipate (BA) units in the involved PBAT polyesters.Figure 31H NMR spectra of BDP fragments from composts and commercial bags. (A) BDP fragments recovered from the composts and (B) the commercial compostable bags. Fragments were coded as follows: p or f for pre-compost or finished compost, followed by the plant number (#1 to #4), an indication of the size fraction ( > 5 mm or 1–5 mm) in which the fragment was found, and finally, the fragment number. Bags were arbitrarily numbered 1–10, see Suppl Table S1 for supplier information. The spectra (in grey) of the reference materials for PLA and PBAT are given as basis for the interpretation. Spectra in red refer to test samples consisting only of PBAT, while those in blue indicate samples composed of PBAT/PLA mixtures. (C) Chemical structures of PLA and PBAT, chemical shifts of the protons are assigned as indicated in the reference spectra in (B).Full size imageThe 1H-NMR spectra corroborate the FTIR measurements in that all investigated commercial bags were made from PBAT/PLA mixtures of varied composition (Table 2). By comparison, some of the fragments, for instance, f#1_5mm_4, appeared to consist of only PBAT. Other fragments, e.g., f#1_1mm_9, were mixtures of PLA and PBAT (Table 2). However, even in the case of PBAT/PLA mixtures, the average PBAT content tended to be higher in the fragments than in the bags, while the BT/BA monomer ratio in the respective PBATs, was also significantly higher in the fragments than in the bags. If we assume the fragments to stem from similar compostable bags as the ones included in our comparison, this would mean that during composting of such a bag, the PLA degrades more quickly than the PBAT, whereas within a given PBAT polyester, the BA unit is more easily degraded than the BT unit. Evidence can indeed be found in the pertinent literature that PLA has faster biodegradation kinetics than PBAT, while BT is more resistant to mineralization than BA17,18.Table 2 Composition of commercial compostable bags and BDP fragments recovered from the composts as analyzed by 1H-NMR.Full size tableNext, differential scanning calorimetry (DSC) was used to analyze fragments compared to commercial bags in regard to the presence of amorphous vs. crystalline domains, a parameter expected to affect biodegradation kinetics and therefore the putative environmental impact of the produced microplastic16 upon release into the environment with the composts. Whereas amorphous domains show glass transition, crystalline domains show melting, both of which can be discerned by the respective phase transition enthalpy in the DSC curves (Fig. 4).Figure 4DSC curves of BDP fragments and compostable bags #1 and #7. Curves for the reference materials (in grey) for PLA and PBAT are given for comparison. Curves were recorded during the first heating run (temperature range: − 50 °C to 200 °C, heating rate: 10 °C min−1). (A) and (B) curves in red refer to test samples consisting only of PBAT, while those in blue indicate samples composed of PBAT/PLA mixtures. Fragments were coded as follows: p or f for pre-compost or finished compost, followed by the plant number (#1 to #4), an indication of the size fraction ( > 5 mm or 1–5 mm) in which the fragment was found, and finally, the fragment number.Full size imageThe curve for the reference PBAT shows a glass transition temperature (Tg) of − 29 °C and a broad melting range between 100 and 140 °C for the crystalline domains, while that of the PLA reference shows a glass transition temperature of 58 °C and a narrower melting peak between 144 °C and 162 °C. The curve for commercial bag #1, which had a comparatively high PLA content, shows a pronounced melting peak in the expected range; the same is the case for fragment p#3_5mm_1 and to a lesser extent for fragment p#3_5mm_9, two fragments, which also have high PLA contents. The DSC curves of the other fragments and bag #1 are undefined in comparison, which is due to their high PBAT content. According to the DSC curves, most of the investigated materials are semicrystalline, i.e., contain both amorphous (glass transition) and crystalline (melting) domains. However, the DCS data alone allow only a qualitative discussion of the differences between fragments and bags.To obtain quantitative data on the crystallinity differences, wide angle X-ray scattering (WAXS) spectra were recorded. WAXS requires fragments at least 3 cm long, which restricted the number of fragment samples to three, all of which were found in pre-compost samples. The corresponding curves are shown in Fig. 5A–C. The spectra of the commercial biodegradable bags are shown in Suppl Figure S3. Foils were in addition prepared by heat pressing from the reference materials for PLA and PBAT in order to include them into the WAXS measurements (Fig. 5D). While the foils produced from the PBAT reference material produced crystallinity peaks at 16.2°, 17.3°, 20.4°, 23.2°, and 24.8°, the foil prepared from the PLA reference material showed only an amorphous halo at 15.5° and 31.5°, which is in accordance with values published in the literature19. A more pronounced crystallinity peak was obtained in the case of an additionally annealed PLA foil.Figure 5WAXS curves with Lorenz fitting for (A) fragment p#3_5mm_1, (B) fragment p#3_5mm_9, and (C) fragment p#4_5mm_2. (D) WAXS curves for foils produced from the PBAT and PLA reference materials; the percent values indicate the crystallinity. The dash lines are the fitting peak curves for the XRD spectrum. Crystallinity can be obtained by dividing the integration area of the fitted peaks by the integration area of the entire spectrum. Fragments were coded as follows: p or f for pre-compost or finished compost, followed by the plant number (#1 to #4), an indication of the size fraction ( > 5 mm or 1–5 mm) in which the fragment was found, and finally, the fragment number.Full size imageIn case of the fragments and bags, the peaks of PLA and PBAT overlapped to some extent in the WAXS spectra, but by conducting Lorenz fitting using Origin software, the overall crystallinity could be calculated as follows:$$chi = { 1}00% , *{text{ Aa}}/left( {{text{Aa }} + {text{ Ac}}} right)$$where χ is the crystallinity and Aa and Ac represent the areas of the amorphous and crystalline peaks.Using this equation, crystallinities of 55% (fragments p#3_5mm_1), 34% (p#3_5mm_9), and 34% (p#4_5mm_2) were calculated for the fragments. The foils prepared in house for the reference materials had similar crystallinities (43% in case of the annealed PLA foil and 26% of the PBAT foil), while the simple PLA foil was amorphous. By comparison, for eight of the commercial bags, crystallinities in the range from 1% to 7% were calculated, whereas these values were 14% and 15% for the remaining two bag types (Suppl Figure S3).The high crystallinity of the larger fragments recovered from the pre-compost samples suggests that crystalline domains of BDP materials may indeed disintegrate more slowly than the amorphous ones, as prior studies on microbial biodegradation have suggested7,8. Admittedly, such large fragments per se would not enter the environment, since the final sieving step used to prepare the finished composts is quite efficient at removing them. However, it is tempting to extrapolate that residual BDP in general are remnants of the more crystal domains of the original material, even though experimental proof of this assumption is at present not possible. 10 wt% of a BDP bag is allowed to remain after standard composting. It is usually assumed that any such residues continue to degrade with comparable speed. However, should these residues correspond to the more crystalline domains, rather than degrading with similar speed as the bulk material, the more crystalline fragments can be expected to persist for a much longer and at present unpredictable length of time in the environment, e.g. when applied to the soil with the composts; in particular, when they are also enriched in PBAT and BT units as suggested by our analysis of the chemical composition. Data from the use of biodegradable foils in agriculture show that the degradation in the environment may take years20. Altogether this may have unforeseen economic and environmental consequences, especially when considering the high fraction of BDP fragments < 5 mm. Putative consequences include changes in soil properties, the soil microbiome and therefore in plant performance21, a factor indispensable for worldwide nutrition.Residues of BDP fragments  1 mm were found in the collected LF samples. This is hardly surprising, given that the LF is produced by press filtration of the digestate after the anaerobic stage. Such a filtration step can be expected to retain fragments > 1 mm in the produced filter cake, which goes into the composting step, leaving the filtrate, i.e. the LF, essentially free of such particles. Anaerobic digestion is currently not assumed to contribute significantly to the degradation of BDP17,22, but the process conditions (mixing, pumping) may promote breakdown of larger fragments, particularly when additives such as plasticizers23 leach out of the material.Since the residual solids content of the LF is low (plant #1: 8.6 wt%, plant #3: 5.8 wt%), a combination of enzymatic-oxidative treatment and µFTIR imaging originally developed for environmental samples from aqueous systems24,25 could be adapted for the analysis (size and chemical signature) of particles in the LF down to a size of 10 µm. The corresponding data are compiled in Table 3. In all cases, residual fragments from PBAT-based polymers represented the dominant plastic fraction in the investigated samples; i.e. approximately 53% of all plastic particles in the LF from plant #1 (11,520 BDP particles per liter) and 65% in the case of plant #3 (12,480 BDP particles per liter). Liquid manure is applied several times a year to fields at a concentration of 2–3 L m−2. According to our analysis > 20,000 BDP microparticles of a size ranging from 10 µm to 500 µm enter each m2 of agricultural soil whenever LF is applied on agricultural surfaces.Table 3 Microplastic fragments (BDP/all) found per liter of liquid fertilizer.Full size tableDue to the complexity of the matrix, a similar analysis of individual plastic fragments  1 mm. Six compost samples representing the more contaminated ones based on the content of fragments > 1 mm, namely, f#1, f#2, p#3, f#3, p#4 and f#4 (nomenclature: f or p for finished or pre-compost, followed by plant number), were extracted with a 90/10 vol% chloroform/methanol mixture. The amounts of PBAT and PLA in the obtained extracts were then quantified via 1H-NMR (Table 4). Briefly, the intensity of characteristic signals in the extract spectra of the compost samples (see Suppl Figure S4) were compared to peak intensities produced by calibration standards of the pure polymer dissolved at a known concentration in the chloroform/methanol. All samples and standards were normalized using the 1,2-dichloroethan signal at 3.73 ppm as internal standard. See also Suppl Figure S5 for an exemplification of the quantification of the PBAT/PLA ratios. Based on the amounts of PBAT and PLA extracted from a known amount of compost, the total mass concentration (wt% dry weight) of these polymers in the composts was calculated.Table 4 Evidence of PBAT and PLA residues caused by fragments  2 mm. Moreover, residues of PBAT and PLA were found in all investigated compost samples, including the finished compost from plant #4, which had shown no contamination by larger BPD fragments (Table 1). The pre-compost from that plant had shown a few contaminating BDP fragments in the > 5 mm fraction. However, in regard to the fragments More

Department of Environmental Science, Policy and Management, UC Berkeley, Berkeley, CA, USAT. F. Keenan, X. Luo, Y. Zhang & S. ZhouClimate and Ecosystem Sciences Division, Lawrence Berkeley National Laboratory, Berkeley, CA, USAT. F. Keenan, X. Luo, Y. Zhang & S. ZhouDepartment of Geography, National University of, Singapore, SingaporeX. LuoARC Centre of Excellence for Climate Extremes, Sydney, New South Wales, AustraliaM. G. De KauweClimate Change Research Centre, University of New South Wales, Sydney, New South Wales, AustraliaM. G. De KauweSchool of Biological Sciences, University of Bristol, Bristol, UKM. G. De KauweHawkesbury Institute for the Environment, Western Sydney University, Penrith, New South Wales, AustraliaB. E. MedlynDepartment of Life Sciences, Imperial College London, Ascot, UKI. C. PrenticeDepartment of Biological Sciences, Macquarie University, North Ryde, New South Wales, AustraliaI. C. PrenticeDepartment of Earth System Science, Tsinghua University, Haidian, Beijing, ChinaI. C. Prentice & H. WangDepartment of Environmental Systems Science, ETH, Zurich, SwitzerlandB. D. StockerSwiss Federal Institute for Forest, Snow and Landscape Research WSL, Birmensdorf, SwitzerlandB. D. StockerDepartment of Biological Sciences, Texas Tech University, Lubbock, TX, USAN. G. SmithPhysical and Life Sciences Directorate, Lawrence Livermore National Laboratory, Livermore, CA, USAC. TerrerDepartment of Civil and Environmental Engineering, Massachusetts Institute of Technology, Boston, MA, USAC. TerrerSino-French Institute for Earth System Science, College of Urban and Environmental Sciences, Peking University, Beijing, ChinaY. ZhangLamont-Doherty Earth Observatory of Columbia University, Palisades, NY, USAS. ZhouEarth Institute, Columbia University, New York, NY, USAS. ZhouDepartment of Earth and Environmental Engineering, Columbia University, New York, NY, USAS. ZhouState Key Laboratory of Earth Surface Processes and Resources Ecology, Faculty of Geographical Science, Beijing Normal University, Beijing, ChinaS. Zhou More

Rendell, L. & Whitehead, H. Culture in whales and dolphins. Behav. Brain Sci. 24, 309–324 (2001).CAS 
Article 

Google Scholar 
Krützen, M. et al. Cultural transmission of tool use in bottlenose dolphins. Proc. Natl. Acad. Sci. U.S.A. 102, 8939–8943 (2005).ADS 
Article 

Google Scholar 
Kawai, M. Newly-acquired pre-cultural behavior of the natural troop of Japanese monkeys on Koshima Islet. Primates 6, 1–30 (1965).Article 

Google Scholar 
Slater, P. The cultural transmission of bird song. Trends Ecol. Evol. 1, 94–97 (1986).CAS 
Article 

Google Scholar 
Whitehead, H. & Rendell, L. The cultural lives of whales and dolphins. (University of Chicago Press, 2014).Whiten, A. The identification and differentiation of culture in chimpanzees and other animals: from natural history to diffusion experiments. The question of animal culture, 99–124 (2009).Allen, J. A. Community through culture: from insects to whales: How social learning and culture manifest across diverse animal communities. BioEssays 41, 1900060 (2019).Article 

Google Scholar 
Allen, J., Weinrich, M., Hoppitt, W. & Rendell, L. Network-based diffusion analysis reveals cultural transmission of lobtail feeding in humpback whales. Science 340, 485–488 (2013).ADS 
CAS 
Article 

Google Scholar 
Baker, C. Migratory movement and population structure of humpback whales (Megaptera novaeangliae) in the central and eastern North Pacific. Mar Ecol Prog Ser 31, 105–119 (1986).ADS 
Article 

Google Scholar 
Garrigue, C. et al. Movement of individual humpback whales between wintering grounds of Oceania (South Pacific), 1999 to 2004. J. Cetacean Res. Manage 3, 275–281 (2011).
Google Scholar 
Rosenbaum, H. C. et al. First circumglobal assessment of Southern Hemisphere humpback whale mitochondrial genetic variation and implications for management. Endang. Spec. Res. 32, 551–567 (2017).Article 

Google Scholar 
Noad, M. J., Cato, D. H., Bryden, M. M., Jenner, M. N. & Jenner, K. C. Cultural revolution in whale songs. Nature 408, 537. https://doi.org/10.1038/35046199 (2000).ADS 
CAS 
Article 
PubMed 

Google Scholar 
Payne, R. S. & McVay, S. Songs of humpback whales. Science 173, 585–597 (1971).ADS 
CAS 
Article 

Google Scholar 
Garland, E. C. et al. Dynamic horizontal cultural transmission of humpback whale song at the ocean basin scale. Curr. Biol. 21, 687–691. https://doi.org/10.1016/j.cub.2011.03.019 (2011).CAS 
Article 
PubMed 

Google Scholar 
Payne, R. & Guinee, L. N. Humpback whale (Megaptera novaeangliae) songs as an indicator of “stocks”. Communication and behavior of whales, 333–358 (1983).Garrigue, C. et al. Movements of humpback whales in Oceania, South Pacific. J. Cetac. Res. Manage. 4, 255–260 (2002).
Google Scholar 
Derville, S., Torres, L. G., Zerbini, A. N., Oremus, M. & Garrigue, C. Horizontal and vertical movements of humpback whales inform the use of critical pelagic habitats in the western South Pacific. Sci. Rep. 10, 1–13 (2020).Article 

Google Scholar 
Garrigue, C. et al. First assessment of interchange of humpback whales between Oceania and the East coast of Australia. J. Cetac. Res. Manage. 3, 269–274 (2011).
Google Scholar 
Steel, D. et al. Migratory connections between humpback whales from South Pacific breeding grounds and Antarctic feeding areas based on genotype matching. Int. Whal. Comm. (2008).Constantine, R., Russell, K., Gibbs, N., Childerhouse, S. & Baker, C. S. Photo-identification of humpback whales (Megaptera novaeangliae) in New Zealand waters and their migratory connections to breeding grounds of Oceania. Mar. Mam. Sci. 23, 715–720 (2007).Article 

Google Scholar 
Garland, E. C. et al. Humpback whale song on the southern ocean feeding grounds: implications for cultural transmission. PLoS ONE 8, e79422 (2013).ADS 
Article 

Google Scholar 
Garland, E. C. et al. Population structure of humpback whales in the western and central South Pacific Ocean as determined by vocal exchange among populations. Conserv. Biol. 29, 1198–1207 (2015).Article 

Google Scholar 
Cholewiak, D. M., Sousa-Lima, R. S. & Cerchio, S. Humpback whale song hierarchical structure: Historical context and discussion of current classification issues. Mar. Mam. Sci. 29, E312–E332. https://doi.org/10.1111/mms.12005 (2013).Article 

Google Scholar 
Payne, K., Tyack, P. & Payne, R. Progressive changes in the songs of humpback whales (Megaptera novaeangliae): a detailed analysis of two seasons in Hawaii. Communication and behavior of whales, 9–57 (1983).Payne, K. & Payne, R. Large scale changes over 19 years in songs of humpback whales in Bermuda. Ethology 68, 89–114 (1985).
Google Scholar 
Allen, J. A., Garland, E. C., Dunlop, R. A. & Noad, M. J. Cultural revolutions reduce complexity in the songs of humpback whales. Proc. R. Soc. B: Biol. Sci. 285, 20182088. https://doi.org/10.1098/rspb.2018.2088 (2018).Article 

Google Scholar 
Allen, J. A., Garland, E. C., Murray, A., Noad, M. J. & Dunlop, R. Using self-organizing maps to classify humpback whale song units and quantify their similarity. J. Acoust. Soc. Am. 142, 1943–1952 (2017).ADS 
Article 

Google Scholar 
Murray, A., Dunlop, R. A., Noad, M. J. & Goldizen, A. W. Stereotypic and complex phrase types provide structural evidence for a multi-message display in humpback whales (Megaptera novaeangliae). J Acoust Soc Am. 143, 980–994 (2018).ADS 
Article 

Google Scholar 
Garland, E. C. et al. Redefining western and central South Pacific humpback whale population structure based on vocal cultural exchange. (2013).Rekdahl, M. Humpback whale vocal communication: Use and stability of social calls and revolutions in the songs of east Australian whales. (2012).Templeton, C. N., Laland, K. N. & Boogert, N. J. Does song complexity correlate with problem-solving performance in flocks of zebra finches?. Anim. Behav. 92, 63–71 (2014).Article 

Google Scholar 
Boogert, N. J., Giraldeau, L.-A. & Lefebvre, L. Song complexity correlates with learning ability in zebra finch males. Anim. Behav. 76, 1735–1741 (2008).Article 

Google Scholar 
Winn, H. & Winn, L. The song of the humpback whale Megaptera novaeangliae in the West Indies. Mar. Biol. 47, 97–114 (1978).Article 

Google Scholar 
Girola, E., Noad, M. J., Dunlop, R. A. & Cato, D. H. Source levels of humpback whales decrease with frequency suggesting an air-filled resonator is used in sound production. The Journal of the Acoustical Society of America (In Review).Warren, V. E., Constantine, R., Noad, M., Garrigue, C. & Garland, E. C. Migratory insights from singing humpback whales recorded around central New Zealand. R. Soc. Open Sci. 7, 201084 (2020).ADS 
Article 

Google Scholar 
Garland, E. C. et al. Quantifying humpback whale song sequences to understand the dynamics of song exchange at the ocean basin scale. J. Acoust. Soc. Am. 133, 560–569. https://doi.org/10.1121/1.4770232 (2013).ADS 
Article 
PubMed 

Google Scholar 
Owen, C. et al. Migratory convergence facilitates cultural transmission of humpback whale song. R. Soc. Open Sci. 6, 190337 (2019).ADS 
Article 

Google Scholar 
Garland, E. C., Rendell, L., Lamoni, L., Poole, M. M. & Noad, M. Song hybridization events during revolutionary song change provide insights into cultural transmission in humpback whales. Proc. Natl. Acad. Sci. 114, 7822–7829 (2017).CAS 
Article 

Google Scholar 
Noad, M. & Cato, D. A combined acoustic and visual survey of humpback whales off southeast Queensland. Mem. Qld. Mus. 47, 507–523 (2001).
Google Scholar 
Spierings, M., de Weger, A. & Ten Cate, C. Pauses enhance chunk recognition in song element strings by zebra finches. Anim. Cogn. 18, 867–874 (2015).Article 

Google Scholar 
Doupe, A. J. & Kuhl, P. K. Birdsong and human speech: common themes and mechanisms. Annu. Rev. Neurosci. 22, 567–631 (1999).CAS 
Article 

Google Scholar 
Allen, J. A., Garland, E. C., Dunlop, R. A. & Noad, M. J. Network analysis reveals underlying syntactic features in a vocally learnt mammalian display, humpback whale song. Proc. R. Soc. B 286, 20192014 (2019).Article 

Google Scholar 
Barón Birchenall, L. Animal communication and human language: An overview. International Journal of Comparative Psychology 29 (2016).Noad, M. J. The use of song by humpback whales (Megaptera novaeangliae) during migration off the east coast of Australia (doctoral dissertation) Doctor of Philosophy thesis, University of Sydney, (2002).Catchpole, C. Song and female choice: good genes and big brains?. Trends Ecol. Evol. 11, 358–360. https://doi.org/10.1016/0169-5347(96)30042-6 (1996).CAS 
Article 
PubMed 

Google Scholar 
Nowicki, S., Hasselquist, D., Bensch, S. & Peters, S. Nestling growth and song repertoire size in great reed warblers: evidence for song learning as an indicator mechanism in mate choice. Proc. R. Soc. London B: Biol. Sci. 267, 2419–2424 (2000).CAS 
Article 

Google Scholar 
NOAA. Vol. 81 (ed National Marine Fisheries Service) 62260–62320 (Department of Commerce, Federal Register, 2016).Noad, M. J., Kniest, E. & Dunlop, R. A. Boom to bust? Implications for the continued rapid growth of the eastern Australian humpback whale population despite recovery. Popul. Ecol. 61(2), 198–209 (2019).Article 

Google Scholar 
Garrigue, C., Albertson, R. & Jackson, J. A. An anomalous increase in the New Caledonia humpback whales breeding sub-stock E2. Scientific Committee of the International Whaling Commission, Paper (2012).Garland, E. C. & McGregor, P. K. Cultural transmission, evolution, and revolution in vocal displays: Insights from bird and whale song. Front. Psychol. 11, 2387 (2020).Article 

Google Scholar 
Zandberg, L., Lachlan, R. F., Lamoni, L. & Garland, E. C. Global cultural evolutionary model of humpback whale song. Philos. Trans. R. Soc. B 376, 20200242 (2021).Article 

Google Scholar 
Crates, R. et al. Loss of vocal culture and fitness costs in a critically endangered songbird. Proc. R. Soc. B 288, 20210225 (2021).Article 

Google Scholar 
Garland, E. C., Garrigue, C. & Noad, M. J. When does cultural evolution become cumulative culture? A case study of humpback whale song. Philos. Trans. R. Soc. B 377, 20200313 (2022).Article 

Google Scholar 
Garland, E. C. et al. Improved versions of the Levenshtein distance method for comparing sequence information in animals’ vocalisations: tests using humpback whale song. Behaviour 149, 1413–1441 (2012).Article 

Google Scholar 
Garland, E. C. et al. The devil is in the detail: quantifying vocal variation in a complex, multileveled, and rapidly evolving display. J. Acoust. Soc. Am. 142, 460–472 (2017).ADS 
Article 

Google Scholar 
Rekdahl, M. L. et al. Culturally transmitted song exchange between humpback whales (Megaptera novaeangliae) in the southeast Atlantic and southwest Indian Ocean basins. R. Soc. Open Sci. 5, 172305 (2018).ADS 
Article 

Google Scholar 
Suzuki, R. & Shimodaira, H. Pvclust: an R package for assessing the uncertainty in hierarchical clustering. Bioinformatics 22, 1540–1542 (2006).CAS 
Article 

Google Scholar 
Sokal, R. R. & Rohlf, F. J. The comparison of dendrograms by objective methods. Taxon 11(2), 33–40 (1962).Article 

Google Scholar 
R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. 2015. URL http (R Core Development Team, 2016). More

Friedlingstein, P. et al. Global carbon budget 2020. Earth Syst. Sci. Data 12, 3269–3340 (2020).
Google Scholar 
Myneni, R. B. et al. A large carbon sink in the woody biomass of northern forests. Proc. Natl Acad. Sci. USA 98, 14784–14789 (2001).CAS 

Google Scholar 
Sitch, S. et al. Recent trends and drivers of regional sources and sinks of carbon dioxide. Biogeosciences 12, 653–679 (2015).
Google Scholar 
Kauppi, P. E. et al. Large impacts of climatic warming on growth of boreal forests since 1960. PLoS ONE 9, e111340 (2014).
Google Scholar 
Zhu, Z. C. et al. Greening of the Earth and its drivers. Nat. Clim. Change 6, 791–795 (2016).CAS 

Google Scholar 
Piao, S. L. et al. Characteristics, drivers and feedbacks of global greening. Nat. Rev. Earth Environ. 1, 14–27 (2020).
Google Scholar 
IPCC Climate Change 2013: The Physical Science Basis (eds Stocker, T. F. et al.) (Cambridge Univ. Press, 2013).Penuelas, J. et al. Shifting from a fertilization-dominated to warming-dominated period. Nat. Ecol. Evol. 1, 1438–1445 (2017).
Google Scholar 
D’Arrigo, R., Wilson, R., Liepert, B. & Cherubini, P. On the ‘divergence problem’ in northern forests: a review of the tree-ring evidence and possible causes. Glob. Planet. Change 60, 289–305 (2008).
Google Scholar 
Beck, P. S. A. et al. Changes in forest productivity across Alaska consistent with biome shift. Ecol. Lett. 14, 373–379 (2011).
Google Scholar 
Vickers, H. et al. Changes in greening in the High Arctic: insights from a 30-year AVHRR max NDVI dataset for Svalbard. Environ. Res. Lett. 11, 105004 (2016).
Google Scholar 
Piao, S. L. et al. Evidence for a weakening relationship between interannual temperature variability and northern vegetation activity. Nat. Commun. 5, 5018 (2014).CAS 

Google Scholar 
Duffy, K. A. et al. How close are we to the temperature tipping point of the terrestrial biosphere? Sci. Adv. 7, eaay1052 (2021).CAS 

Google Scholar 
Liu, Y. W. et al. Seasonal responses of terrestrial carbon cycle to climate variations in CMIP5 models: evaluation and projection. J. Clim. 30, 6481–6503 (2017).
Google Scholar 
Huang, M. T. et al. Air temperature optima of vegetation productivity across global biomes. Nat. Ecol. Evol. 3, 772–779 (2019).
Google Scholar 
Park, T. et al. Changes in timing of seasonal peak photosynthetic activity in northern ecosystems. Glob. Change Biol. 25, 2382–2395 (2019).
Google Scholar 
Keeling, C. D., Chin, J. F. S. & Whorf, T. P. Increased activity of northern vegetation inferred from atmospheric CO2 measurements. Nature 382, 146–149 (1996).CAS 

Google Scholar 
Piao, S. L., Friedlingstein, P., Ciais, P., Viovy, N. & Demarty, J. Growing season extension and its impact on terrestrial carbon cycle in the Northern Hemisphere over the past 2 decades. Glob. Biogeochem. Cycles 3, GB3018 (2007).
Google Scholar 
Keenan, T. F. et al. Net carbon uptake has increased through warming-induced changes in temperate forest phenology. Nat. Clim. Change 4, 598–604 (2014).CAS 

Google Scholar 
Xia, J. Y. et al. Joint control of terrestrial gross primary productivity by plant phenology and physiology. Proc. Natl Acad. Sci. USA 112, 2788–2793 (2015).CAS 

Google Scholar 
Eyring, V. et al. Overview of the Coupled Model Intercomparison Project Phase 6 (CMIP6) experimental design and organization. Geosci. Model Dev. 9, 1937–1958 (2016).
Google Scholar 
Meehl, G. A. & Tebaldi, C. More intense, more frequent, and longer lasting heat waves in the 21st century. Science 305, 994–997 (2004).CAS 

Google Scholar 
Yamori, W., Hikosaka, K. & Way, D. A. Temperature response of photosynthesis in C3, C4, and CAM plants: temperature acclimation and temperature adaptation. Photosynth. Res. 119, 101–117 (2014).CAS 

Google Scholar 
Berry, J. & Bjorkman, O. Photosynthetic response and adaptation to temperature in higher plants. Annu. Rev. Plant Physiol. Plant Mol. Biol. 31, 491–543 (1980).
Google Scholar 
Chen, A., Huang, L., Liu, Q. & Piao, S. L. Optimal temperature of vegetation productivity and its linkage with climate and elevation on the Tibetan Plateau. Glob. Change Biol. 27, 1942–1951 (2021).
Google Scholar 
Smith, N. G., Lombardozzi, D., Tawfik, A., Bonan, G. & Dukes, J. S. Biophysical consequences of photosynthetic temperature acclimation for climate. J. Adv. Model. Earth Syst. 9, 536–547 (2017).
Google Scholar 
Chen, M. & Zhuang, Q. L. Modelling temperature acclimation effects on the carbon dynamics of forest ecosystems in the conterminous United States. Tellus B 65, 19156 (2013).
Google Scholar 
Crous, K. Y. Plant responses to climate warming: physiological adjustments and implications for plant functioning in a future, warmer world. Botany 106, 1049–1051 (2019).
Google Scholar 
Conley, M. M. et al. CO2 enrichment increases water-use efficiency in sorghum. New Phytol. 151, 407–412 (2001).
Google Scholar 
Norby, R. J. & Zak, D. R. Ecological lessons from free-air CO2 enrichment (FACE) experiments. Annu. Rev. Ecol. Evol. Syst. 42, 181–203 (2011).
Google Scholar 
Huang, M. T. et al. Change in terrestrial ecosystem water-use efficiency over the last three decades. Glob. Change Biol. 21, 2366–2378 (2015).
Google Scholar 
Gonsamo, A. et al. Greening drylands despite warming consistent with carbon dioxide fertilization effect. Glob. Change Biol. 7, 3336–3349 (2021).
Google Scholar 
Keenan, T. F. et al. Increase in forest water-use efficiency as atmospheric carbon dioxide concentrations rise. Nature 499, 324–327 (2013).CAS 

Google Scholar 
Lemordant, L. et al. Modification of land–atmosphere interactions by CO2 effects: implications for summer dryness and heat wave amplitude. Geophys. Res. Lett. 43, 10240–10248 (2016).CAS 

Google Scholar 
Lian, X. et al. Summer soil drying exacerbated by earlier spring greening of northern vegetation. Sci. Adv. 6, eaax0255 (2020).
Google Scholar 
Lian, X. et al. Multifaceted characteristics of dryland aridity changes in a warming world. Nat. Rev. Earth Environ. 4, 232–250 (2021).
Google Scholar 
Druel, A., Ciais, P., Krinner, G. & Peylin, P. Modeling the vegetation dynamics of northern shrubs and mosses in the ORCHIDEE land surface model. J. Adv. Model. Earth Syst. 11, 2020–2035 (2019).
Google Scholar 
Scheiter, S., Langan, L. & Higgins, S. I. Next-generation dynamic global vegetation models: learning from community ecology. New Phytol. 198, 957–969 (2013).
Google Scholar 
Mod, H. K. & Luoto, M. Arctic shrubification mediates the impacts of warming climate on changes to tundra vegetation. Environ. Res. Lett. 12, 124028 (2016).
Google Scholar 
Zhang, Y., Commane, R., Zhou, S., Williams, A. P. & Gentine, P. Light limitation regulates the response of autumn terrestrial carbon uptake to warming. Nat. Clim. Change 10, 739–743 (2020).CAS 

Google Scholar 
Bauerle, W. L. et al. Photoperiodic regulation of the seasonal pattern of photosynthetic capacity and the implications for carbon cycling. Proc. Natl Acad. Sci. USA 109, 8612–8617 (2012).CAS 

Google Scholar 
Zhang, Y., Parazoo, N. C., Williams, A. P., Zhou, S. & Gentine, P. Large and projected strengthening moisture limitation on end-of-season photosynthesis. Proc. Natl Acad. Sci. USA 117, 9216–9222 (2020).CAS 

Google Scholar 
Fritz, M. et al. Brief communication: future avenues for permafrost science from the perspective of early career researchers. Cryosphere 9, 1715–1720 (2015).
Google Scholar 
Jin, X. Y. et al. Impacts of climate-induced permafrost degradation on vegetation: a review. Adv. Clim. Change Res. 12, 29–47 (2021).
Google Scholar 
Tramontana, G. et al. Predicting carbon dioxide and energy fluxes across global FLUXNET sites with regression algorithms. Biogeosciences 13, 4291–4313 (2016).CAS 

Google Scholar 
Zhang, Y., Joiner, J., Alemohammad, S. H., Zhou, S. & Gentine, P. A global spatially contiguous solar-induced fluorescence (CSIF) dataset using neural networks. Biogeosciences 15, 5779–5800 (2018).CAS 

Google Scholar 
Navarro-Racines, C., Tarapues, J., Thornton, P., Jarvis, A. & Ramirez-Villegas, J. High-resolution and bias-corrected CMIP5 projections for climate change impact assessments. Sci. Data 7, 7 (2020).
Google Scholar  More

Illustration of the compound event indicesBuilding on earlier studies24,25, we develop two univariate indices to model concurrent climate conditions, i.e., a CWD index that varies from compound cold-wet conditions to CWD conditions, and a CCD index that varies from compound warm-wet conditions to CCD conditions (see “Methods”). The two indices incorporate the dependence between temperature and precipitation and are a measure of how warm/cold and dry a point is relative to the distribution of climate conditions at a given location. We illustrate the two indices on two grid points that have strong but opposite temperature-precipitation correlation. In the case where temperature and precipitation are strongly negatively correlated, the CWD index is well aligned with the primary axis of the bivariate distribution (Fig. 1a). In the case where temperature and precipitation are strongly positively correlated, the same holds for the CCD index (Fig. 1d). As illustrated for several concurrent hot-dry and cold-dry events that occurred around the globe, the two indices well capture these events (Supplementary Figs. 1 and 2).Fig. 1: The relationship between precipitation and temperature and compound indices.a Scatter plot of summer precipitation and temperature anomalies (z-score) with corresponding CWD index in color (see “Methods”). The location is at 97.25°W and 33.75°N from 1901 to 2018. b The same as a but for spring at 84.75°E and 66.75°N. c Same distribution as in a but colored based on the CCD index. d Same distribution as in b but colored based on the CCD index.Full size imageNotably, in the case where precipitation and temperature are strongly positively correlated, the CWD index indicates the relative anomalies of bivariate joint distribution, and some counterintuitive situations might occur relative to the univariate marginals (Fig. 1b). For instance, points might be labeled as strong CWD events (CWD index > 1.5) even though temperature is anomalously cold (temperature anomalies < 0, red dots in lower left quadrant of Fig. 1b). The CCD index exhibits similar behavior (Fig. 1c). This indicates an interesting property of the compound indices to identify strong compound conditions relative to bivariate distribution that are not necessarily extreme from a univariate perspective3,24,26,27.Widespread direct and lagged impacts of springtime compound climate conditionsTo evaluate the lagged summer vegetation responses to spring compound climate conditions, we compute partial correlation between CWD (CCD) spring and subsequent summer vegetation variation by controlling for the influence of summer compound climate conditions on these correlations (see “Methods”). Results show widespread negative associations between CWD spring and subsequent summer vegetation in the mid-latitudes (50°N).a–c The average standardized anomalies (z-score) of GPP during CWD spring but subsequent non-CWD summer (a), non-CWD spring but subsequent CWD summer (b), and consecutive CWD spring and summer (c) for areas in Fig. 2a where summer vegetation responds positively (r ≥ 0.22) to spring CWD climate conditions. d–f The same as a–c, but for soil moisture. g–i The same as a–c, but for runoff. The bar plots with dash lines (without dash line) indicate the average anomalies of multiple observation-based (model-based) products, and the circles indicate the average anomalies of each product. GLASS, LUE, NIRv, Flux-CRU, and Flux-ERA5 are observation-based GPP products, while model simulations are taken from TRENDYv6. GLEAM is observation-based soil moisture. GRUN represents observation-based runoff. GLDAS-VIC, GLDAS-Noah, GLDAS-Catchment, and FLDAS indicate assimilatory soil moisture and runoff that incorporate satellite- and ground-based observational products.Full size imageFig. 4: The responses of vegetation productivity and hydrological variables to CWD events in mid-latitudes (23.5–50°N/S).a–c The average standardized anomalies (z-score) of GPP during CWD spring but subsequent non-CWD summer (a), non-CWD spring but subsequent CWD summer (b), and consecutive CWD spring and summer (c) for areas in Fig. 2a where summer vegetation responds negatively (r ≤ −0.22) to spring CWD climate conditions. d–f The same as a–c, but for soil moisture. g–i The same as a–c, but for runoff. The bar plots with dash lines (without dash line) indicate the average anomalies of multiple observation-based (model-based) products, and the circles indicate the average anomalies of each product. For details on data see Fig. 3.Full size imageFig. 5: The effects of CCD events on vegetation productivity and hydrological variables in mid-to-high latitudes.a–c The average standardized anomalies (z-score) of GPP during CCD spring but subsequent non-CCD summer (a), non-CCD spring but subsequent CCD summer (b), and consecutive CCD spring and summer (c) for areas in Fig. 2b where summer vegetation responds negatively (r ≤ −0.22) to spring CCD climate conditions. d–f The same as a–c, but for soil moisture. g–i The same as a–c, but for runoff. The bar plots with dash lines (without dash line) indicate the average anomalies of multiple observation-based (model-based) products, and the circles indicate the average anomalies of each product. For details on data see Fig. 3.Full size imageCWD events increase vegetation productivity in high latitudesWe first analyze the direct responses of productivity to springtime and summertime CWD events across high latitudes ( >50°N, Fig. 3). Productivity increases during CWD spring and summer (Fig. 3a–c), which is consistent with vegetation responses (Supplementary Fig. 8a–c). Despite elevated spring greenness, spring water overall shows positive anomalies during CWD spring (Fig. 3d, f, g, i). This result indicates that spring greenness during CWD conditions is not associated with dry spring across high latitudes, which is further confirmed by similar anomalies in springtime TWS (Supplementary Fig. 8d, f). In contrast, severe water reduction is found in CWD summer (Fig. 3e, f, h, i). This suggests that despite the beneficial effects of CWD events on productivity in summer, they are associated with summer water deficit.Next, to analyze the lagged effects of springtime CWD events, we investigate the productivity anomalies in summer under three cases, namely CWD spring but non-CWD summer, non-CWD spring but CWD summer, and consecutive CWD spring and summer. Our results indicate that springtime CWD events have positive lagged effects on summer productivity across high latitudes (Fig. 3). Specifically, we find that during non-CWD summer (that is not favorable for summer vegetation growth) preceded by CWD spring, positive anomalies are still found in summer productivity (Fig. 3a). In contrast, during CWD summer (preceded by non-CWD spring), some models and observation-based products exhibit a reduction in summer productivity (Fig. 3b). We further find that summer productivity highly increases during consecutive events (Fig. 3c). Vegetation anomalies show similar behaviors (Supplementary Fig. 8a–c). Regarding the lagged responses of hydrological variables, CWD springs followed by non-CWD summers do not lead to water dryness, despite increased vegetation greenness (Fig. 3d, g). The magnitude of summer water deficit is similar for both cases that include CWD summer (Fig. 3e, f, h, i) and is consistent with summer TWS anomalies (Supplementary Fig. 8e, f). These results imply that in high latitudes, summer water reductions characterized by TWS, soil moisture, and runoff are not associated with increased spring greenness but are primarily caused by summer precipitation deficit.The productivity responses to compound climate conditions may be stronger than that to individual events through the synergistic effects of temperature and precipitation28. To investigate this, we compute the average anomalies in GPP and soil moisture associated with univariate events across the focus areas, which are then compared with the effects of CWD and CCD events in high latitudes (see “Methods”). Warm events can not only directly increase productivity but also show positive lagged effects (Supplementary Fig. 9a, b). In contrast, dry events reduce productivity (Supplementary Fig. 9e, f). This indicates that the direct and lagged positive effects of CWD events across high latitudes are mainly dominated by the warm component, while dry conditions have negative effects. Therefore, the warm-induced increase in productivity slightly exceeds that associated with CWD events (Supplementary Fig. 9b). Soil moisture under warm springs shows positive anomalies (Supplementary Fig. 9c, d), while they slightly decline during dry spring (Supplementary Fig. 9g, h). This suggests that the positive anomalies in soil water during CWD spring are driven by the warm component.CWD events reduce vegetation productivity in mid-latitudesHere, we first investigate the direct effects of springtime and summertime CWD events across mid-latitudes (23.5–50°N/S). Springtime productivity exhibits little changes during CWD spring (Fig. 4a, c), despite dry spring (Fig. 4d, f, g, i). When considering the direct effects of CWD events in summer, the results are similar, whereas the negative magnitude of productivity in summer is larger than that in spring (Fig. 4b, c). This difference suggests CWD conditions in summer show more adverse effects on productivity than that in spring in mid-latitudes. The anomalies in vegetation and TWS are consistent (Supplementary Fig. 10).Next, the lagged effects of springtime CWD events in mid-latitudes are assessed. In cases with CWD spring but non-CWD summer, summer productivity exhibits slight anomalies (Fig. 4a), with slightly decreased summer water (Fig. 4d, g). Summer productivity and water show much higher reductions in case with consecutive events (Fig. 4c, f, i) than for the case with only CWD summer (Fig. 4b, e, h). These results are supported by the responses of vegetation indices and TWS (Supplementary Fig. 10), revealing that springtime CWD events in mid-latitudes have negative lagged effects on summer productivity and water availability.The direct and lagged effects of individual events are finally compared with that of CWD events in mid-latitudes. Dry conditions in spring and summer directly decrease productivity and cause soil water dryness (Supplementary Fig. 11a–d). Moreover, dry spring depletes soil moisture earlier, which, in turn, causes dry summer and reduction in productivity during non-dry summer (Supplementary Fig. 11a, c). This indicates that dry springs have negative lagged effects on summer productivity. In contrast, productivity and soil water show positive anomalies during warm springs, while they show negative anomalies in summer (Supplementary Fig. 11e–h). These results suggest that the direct and lagged negative effects of CWD springs are dominated by the dry component in mid-latitudes, while the warm component mitigates the negative effects of the dry component in spring. Accordingly, the decline in productivity due to dry conditions thus exceeds that triggered by CWD events (Supplementary Fig. 11b).Decreased vegetation productivity due to the negative synergistic effects of CCD eventsHere, we first investigate the direct effects of CCD events across mid-to-high latitudes. Productivity reductions are found during springtime and summertime CCD events (Fig. 5a–c) concurrent with water reductions (Fig. 5). Vegetation and TWS show similar behaviors during CCD spring and summer (Supplementary Fig. 12). These results reveal that CCD events in spring and summer can impose direct adverse impacts on productivity and soil water across mid-to-high latitudes. The productivity reductions in spring and summer are similar in magnitude (Fig. 5a, b), indicating that CCD events between spring and summer can cause similar damage to productivity.We then analyze the lagged effects of springtime CCD events. Our results indicate that springtime CCD events show negative lagged effects on summer productivity and cause summer water reductions in mid-to-high latitudes (Fig. 5). Specifically, we find that in cases with CCD spring but non-CCD summer, summer productivity and water exhibit strongly negative anomalies (Fig. 5a, d, g). Moreover, summer anomalies are higher during consecutive events (Fig. 5c, f, i) than the cases including only CCD summer (Fig. 5b, e, h). Vegetation indices and TWS show similar responses (Supplementary Fig. 12). Our results further indicate that CCD spring has more severe negative lagged effects on productivity than CWD spring. That is, we find that in comparison to cases with preceding CWD spring and consecutive CWD events, summer productivity shows higher reduction in cases with preceding CCD spring and consecutive CCD events (Fig. 4a, c versus Fig. 5a, c). Moreover, in cases with CCD spring but non-CCD summer (Fig. 5a, d, g), summer anomalies are close to those in scenarios with non-CCD spring but CCD summer (Fig. 5b, e, h). The vegetation and TWS anomalies further confirm this situation (Supplementary Fig. 12a, b, d, e). These results suggest that the lagged effects of CCD spring can be of similar magnitude as their direct adverse effects.We finally compare the direct and lagged effects of individual events with that of CCD events in mid-to-high latitudes. Cold conditions in spring and summer directly reduce productivity but show weak effects on soil moisture (Supplementary Fig. 13a–d), and cold spring shows negative lagged effects on summer productivity (Supplementary Fig. 13a). Dry events show direct and lagged negative effects on productivity and soil moisture (Supplementary Fig. 13e–h). These results imply that the negative lagged effects of CCD springs are dominated by both cold and dry components. The effects of CCD events on productivity mostly exceeds the individual dry or cold impacts (Supplementary Fig. 13a, b, e, f). More

OverviewIn this study, we used VIIRS active fire detections to track the dynamic evolution of all fires in California from 2012 to 2020 (Fig. 1). We developed an approach that has the following steps. First, after reading the satellite fire pixel data at each 12-hour time step, the new fire pixels are aggregated into multiple clusters using the fire pixel locations and an automatic clustering algorithm. These clusters are then spatially compared to existing fire objects. If a cluster is not close to any existing active fire object, we use all fire pixels within the cluster to form a new fire object. If a cluster is located near an existing fire object which is still active, we view the cluster as an extension of the existing fire. In this case, we append all pixels within the cluster to the corresponding existing fire object, allowing the existing object to grow. When a fire expands and gets close enough (within a pre-defined distance threshold) to an existing active fire object, we merge the two objects. For each time step (12 hours in this case for the two overpasses), we derive or update a suite of attributes and status indicators associated with each fire event, including pixel-level attributes of fire and surface properties, vector geometries related to the fire shape, and meta-attributes characterizing the entire fire object.Data inputSatellite remote sensing instruments provide active fire detections with accurate geographical location and broad spatial coverage. The primary data for this fire tracking system are active fire locations and the fire radiative power (FRP) recorded by the VIIRS instrument aboard the Suomi National Polar-orbiting Partnership (Suomi-NPP) satellite24. VIIRS observes Earth’s surface twice each day in low and mid latitude regions, with local overpass times of approximately 1:30 am and 1:30 pm. Compared to its predecessor, the MODIS sensors on the Terra and Aqua satellites, VIIRS has a higher spatial resolution and can detect smaller and cooler fires24. Also, the VIIRS instrument provides a more consistent pixel area across the image swath25, resulting in more accurate estimates of active fire location. Therefore, compared with MODIS, the VIIRS active fire products can be used to map fire event progression with higher accuracy21. Two streams of VIIRS active fire data are operationally produced using a contextual fire detection algorithm24, drawing upon VIIRS moderate resolution band (M-band) and imaging band (I-band) reflectance and radiance data layers. In this fire tracking system, we used the Suomi-NPP VIIRS I-band fire location data product (VNP14IMGML, Collection 1 Version 4) that contains the centre location, FRP, scan angle, and other attribute fields associated with each pixel. The I-band fire detection product has a 375-m spatial resolution at nadir (the sub-satellite point) and an average resolution across the full swath of about 470 m. Theoretical estimates of fire detection efficiency for the VIIRS sensor indicate that during the day, VIIRS can detect 700 K fires with 50% probability that have a size of about 200 m2 (a 15 m × 15 m fire area)24. During night, the detection efficiency increases, and VIIRS can detect 700 K fires as small as 40 m2. From a fire spread tracking perspective, these detection efficiencies imply that in many instances, the area of a fire pixel that is covered with flaming fire combustion is several orders of magnitude smaller than the overall pixel size. The VNP14IMGML data, available from 2012 onwards, were downloaded from the University of Maryland VIIRS Active Fire website (https://viirsfire.geog.umd.edu/).Land cover data are an additional input in the system required to classify different fire types and determine the spatial connectivity threshold. Here we use the U.S. National Land Cover Database (NLCD 2016)26 that is available from the Multi-Resolution Land Characteristics (MRLC) Consortium website (https://www.mrlc.gov/national-land-cover-database-nlcd-2016). We aggregated the original 30-m data to match the spatial resolution of VIIRS active fire data, and merged the original 16 classes into several groups: ‘Water’, ‘Urban’, ‘Barren’, ‘Forest’, ‘Shrub’, ‘Grassland’, and ‘Agriculture’. We used the 1000-hour dead fuel moisture from the high-resolution (4 km) gridMET product27 for the purpose of separating wildfires and management fires. This gridMET dataset was computed from 7–day average conditions composed of day length, hours of rain, and daily temperature and humidity ranges. Regularly updated gridMET data are available from the Climatology Lab website (http://www.climatologylab.org/gridmet.html).Other ancillary and validation datasets used in this study included a shapefile of California borders and fire perimeters from the California Forestry and Fire Protection’s Fire and Resource Assessment Program (FRAP) dataset (https://frap.fire.ca.gov/mapping/maps/).Fire object hierarchyFire detections from VIIRS are dynamically tracked within the framework of a three-level object hierarchy (Fig. 1). The lowest level is the fire pixel object, which includes the geographical location (latitude and longitude), the FRP value, and the origin (first assigned fire object id). The second level is the fire object, which includes all attributes associated with each individual fire event at a particular time step (Table 2). Each fire object includes one or more fire pixel objects, a unique identification number (id), and a set of attributes associated with the whole fire. Two types of fire attributes are derived and recorded for each fire object. The first type encompasses temporal (e.g., ignition time, duration) and spatial (e.g., centroid, ignition location) characteristics of the object as well as general properties (e.g., size, type, active status). The second type is the geometric information related to the fire object, including the fire perimeter, the active fire front line, and the newly detected fire pixel locations (stored as vectors). All fire objects in the State of California are combined to form an allfires object, to characterize the whole-region fire situation at a specific time step. The allfires object comprises a list of fire objects, and also contains meta information representing the statistics of all fires and the records describing fire evolution. A full list of the attributes associated with the pixel object, the fire object, and the allfires object is presented in Table 2.Table 2 List of main attributes associated with pixel, fire and allfires objects.Full size tableFire event trackingThe fire records (locations and FRPs) from the monthly VIIRS active fire location products (VNP14IMGML) are read into the system at each half-daily time step (roughly 1:30 am and 1:30 pm local time). We apply spatial and temporal filters to the data to extract active fire pixels recorded in California during each 12-hour time interval. We also apply quality flag filters (thermal anomaly type of ‘0: presumed vegetation fire’ in VNP14IMGML)) to ensure the use of only pixels likely associated with vegetation fires. The fire location and FRP values are used to create fire pixel objects. To speed up the calculation, the newly detected active fire pixels after filtering are first aggregated to specific clusters using the distances between them and an automatic clustering algorithm. In this initial aggregation algorithm, a ball tree28 is created to partition all newly detected active fire pixels into a nested set of hyperspheres in a 2-D space (latitude and longitude). This space partitioning data structure can be used to expedite nearest neighbours search29 and allow for quick cluster grouping. Here we refer to a cluster as a collection of pixel objects that are recorded at the same time step and are also spatially nearby. In the following steps, all pixels within a cluster are considered as a whole for fire merging and creation.We define an extended area for every existing fire object as the fire vector perimeter (see the section of Calculating and recording fire attributes for detail) plus a radial buffer that depends on the fire type property of the object. The buffer is set to 5 km for forest fires and 1 km for other fire types (shrub, crop, urban), considering that the fire spread rate can differ across biomes13. We then evaluate the spatial distance between the perimeters of a newly classified cluster and all existing active fire objects (a fire object keeps an active status if one or more active fire pixels associated with it are detected during the past 5 days), and calculate the shortest distance. If the shortest distance is smaller than the buffer of the associated existing active fire (i.e., new cluster overlaps with the extended area of an existing fire object), we assume all fire pixels in the new cluster are associated with the growth of the existing fire object at the current time step (Fig. 2). The existing fire object is updated by appending all fire pixel objects within the new cluster. If a newly classified cluster does not overlap with the extended area of any existing active fire object, we assume this is a new fire. A new fire object (by assigning a new fire id) is created using all fire pixel objects in the cluster.With the addition of new fire pixels, an existing fire object may expand and touch the extended area of another existing active fire object. If this happens, we assume that these two existing fire objects merge into a single object at this time step. All fire pixels in the fire object with a higher id number (a later start date, termed as the ‘source fire’) are appended to the fire object with lower id number (earlier start date, termed as the ‘target fire’) in this case. We record the id of the target fire in a list of fire mergers, and update all attributes associated with this fire (Fig. 3). In order to avoid double counting, the source fire object (with all pixels being transferred to the target fire object) is flagged as invalid, and is excluded from statistical analysis of fire events.Fig. 3The time series of growth for the SCU Lightning Complex fire (2020). Panel (b) shows the fire size of the SCU fire (total area within the fire object perimeter) at half-daily time steps. A fraction of the fire growth (shown in orange) was due to the addition of newly detected fire pixels. Panel (a) shows the number of new fire pixels (associated with the SCU fire object) detected at each time step. The other part of the fire growth (shown in red) was due to the merging with existing fire objects. Panel (c) shows the number of fire pixels in the existing objects that were merged to the SCU fire object.Full size imageCalculating and recording fire attributesOther than individual fire pixels contained in a fire object, several core attributes (properties and geometries) are also dynamically updated at each time step and are used for fire tracking and characterization.Important time-related attributes include the fire ignition time (the time step at which the first fire pixel within the fire object was detected), the fire end time (the latest time step with an active fire observation), and the fire duration (the time difference between the ignition time and end time). If a fire object does not have new active fire pixels appended during 5 consecutive days (i.e., the fire end time is more than 5 days before the present time step), its status is set to inactive. Once inactive, a fire object is no longer evaluated for use in future clustering (i.e., new active fire detections later will form new fire objects, even if they are spatially close to the inactive fire object).Each fire object is assigned to a specific fire type. The fire type is identified using the major land cover type within the fire perimeter (Table 3). In an initial analysis, we found that prescribed fires, on average, have higher coarse fuel moisture levels than wildfires. Therefore, we also record the 1000-hour fuel moisture (fm1000) from the gridMET dataset27 for each fire object (corresponding to the ignition time step) and use this value to divide forest and shrub fires further to wildfire and prescribed types.Table 3 Classification of fire types based on dominant land cover type (from the US National Land Cover Database) within each fire perimeter and the 1000-hr fuel moisture (FM-1000, from gridMET dataset) at the time of ignition.Full size tableAn essential step in this object-based fire tracking system is to determine the vector shape of the fire perimeter. In this system, we use an alpha shape30 algorithm to derive bounding polygons containing fire pixels in a fire object. For an alpha shape, the radius of the disks forming the curves in the polygon is determined by the alpha parameter α. Compared with the commonly used convex hull, the alpha shape hull is able to capture the irregular shapes around the fire perimeter more accurately22.To identify the optimal values for the α parameter, we performed the following analysis. First, we derived the final fire perimeters for all large fires that occurred in California during the 2018 wildfire season using a set of α values ranging from 500 m to 10 km and compared the results with more refined fire perimeters from the Fire and Resource Assessment Program (FRAP) dataset (Fig. 4). Large magnitude α values tended to overestimate the total burned area, while small α values often fragmented a large fire event. We found that a value of α = 1 km was optimal in terms of balancing the ability of the hull to catch the boundary shape and to keep the integrity of a fire object. For each time step, we applied the alpha shape algorithm to all fire pixel locations associated with a fire object since the time of ignition. This processing step resulted in a concave hull with the shape of polygon or multipolygon. To account for the pixel size, we expanded the concave hull to the fire perimeter using a buffer size equal to half of the VIIRS nadir cross-track pixel width (187.5 m). The alpha shape algorithm does not work when the total number of fire pixels (npix) is less than 4. If npix equals 3, we used a convex hull algorithm and the same 187.5 m buffer to determine a polygon perimeter. If npix is 1 or 2, circles centered on the fire pixel location with radius of 187.5 m were used.Fig. 4Optimization of the alpha shape parameter (α). For all large fires (final size  > 4 km2) in California during 2018, fire perimeters were estimated using VIIRS active fires and different alpha parameters. By comparing (a) the burned area (BA) and (b) the number of fire objects with the FRAP data, an optimal alpha parameter of 1 km was identified for use in this study (shown in red). The vertical bars and lines show the mean and 1-std variability from all fires. The dashed blue lines indicate the ideal values when compared to FRAP. Panels (c)–(h) show the fire perimeters derived using different alpha shape parameters for two sample fires in 2018. The shapes with pink color are final FEDS fire perimeters derived from VIIRS active fires using the alpha shape algorithm. The blue shapes represent the corresponding fire perimeters from the FRAP dataset. Overlap between FRAP and FEDS is shown in purple.Full size imageWe also calculate the active front line for each fire object at each time step. The active fire front consists of the segments of the fire perimeter that are actively burning and releasing energy and emissions. The position of the active fire line is critical in evaluating the fire risk, estimating the fire emissions, and predicting fire spread. We derive the active portion of the fire perimeter as segments that are within a 500 m radius of newly detected fire pixel locations. We found that this threshold allowed for a continuous projection of the active fire front in rapidly expanding areas of large wildfires during the 2018 fire season; this threshold may be optimized in future work to maximize performance metrics for fire model forecasts. The resulting active line for each fire at each time step has the shape of a linestring (object representing a sequence of points and the line segments connecting them), a multi-linestring (a collection of multiple linestrings), or a linear ring (closed linestring). Figure 5 shows an example map of the fire perimeters and active fire front lines on September 8 during the 2020 wildfire season.Fig. 5An example map of fire perimeters and fire active fronts in California. The map was created using the fire event data suite (FEDS) as of the Suomi-NPP afternoon overpass (~1:30 pm local time) of Sep 8, 2020. The background is the Aqua MODIS Corrected Reflectance Imagery (true color) recorded at the same day (provided by the NASA Global Imagery Browse Services). The active front line of a fire is shown in yellow, active fire areas are shown in red, and the area of inactive (extinguished) fires are shown in dark red.Full size imageAdditional fire properties, such as the fire area and active fire line length, are also derived using these geometries of the fire object (see Table 2). Note this list can be easily expanded to include more user-defined properties with the help of the fire object core vector data.The allfires object contains a list of all existing fire objects at a time step. This object also records the ids of fire objects that have been modified (including fires newly formed, fires that expanded with new pixel additions, fires with pixels addition due to merging, and fires that just became invalid) at the current time step.Creating the fire event data suite (FEDS)By tracking the spatiotemporal evolution of all fire objects in California, we derived a complete dataset of fire events for each calendar year (Jan 1 am – Dec 31 pm) during the Suomi-NPP VIIRS era (2012–2020). The dataset contains four products that represent the fire information in California at multiple spatial scales and from different perspectives (Fig. 1 and Table 4), ranging from the most detailed and memory-intensive data format (Pickle) to the most high-level format (CSV).Table 4 Data structure of the FEDS.Full size tableThe first product is the direct serialization result of the allfires object at each time step (twice per day). The product is stored as a Pickle file31 which allows for analysis of the complex allfires object structure (including all attributes associated with all fire objects it contains). This file also serves as the restart file for continued fire tracking at any time step, which is essential for the operational mode using the near-real-time fire data. By restoring an exact copy of the previously pickled allfires object, any attribute in the allfires object can be deserialized from the saved files. The Pickle file is the most basic data product in the dataset, and is created at each half-day time step.The second product (Snapshot) represents a more accessible and self-explanatory variant of the Pickle serialization product. In this product, we tabulated important diagnostic attributes for each fire and saved them in GeoPackage32 data files. Each GeoPackage file includes three data layers: one contains the properties and the fire perimeter geometry, another contains the active fire line geometry, and a third contains the new fire pixel location geometry. This product, created at a half-daily time step, allows for a more straightforward interpretation of regional fire status at a particular time step. We also created a GeoPackage file that summarizes the final fire perimeters and attributes for all fires during the whole study period (2012–2020).The third product (Largefire) focuses on the temporal evolution of individual large fires with an area greater than 4 km2. At each time step, the time series of properties and geometries (fire perimeter, active fire line, and new fire pixel locations) for each of the large fires are extracted and saved to GeoPackage files. This product facilitates the visualization and analysis for an individual targeted fire (Fig. 6) and is particularly useful in the near-real-time evaluation, forecasting, and policy making.Fig. 6The spatiotemporal evolution of the Creek fire (2020). Contours and dots reflect the fire perimeters and newly detected fire pixels at each 12-hour time step. Data for the period of Sep 5 am–Nov 6 am, 2020 are shown.Full size imageThe fourth product (Summary), which is stored as NetCDF and CSV files and created at the end of a fire season, records the all-year time series of fire statistics (including major fire attributes such as number, size, duration, fire line length, etc.) over the whole State of California. This product provides a feasible regional summary of the temporal evolution of fires.Potential for near-real-time (NRT) fire event trackingWhile the main objective of this paper is to apply the object-based fire tracking system to historical VIIRS fire detections and create a retrospective multi-year FEDS, we note that this system has the potential to be used for tracking fire events in near-real-time, providing rich and valuable information for fire management and short-term risk assessment. We have experimented with the use of this system for NRT fire event tracking in California using the daily NRT Suomi-NPP VIIRS active fire detection product (VNP14IMGTDL, collection 6) as the main data source. The VNP14IMGTDL product is routinely produced and is publicly available at the NASA Fire Information for Resource Management System (FIRMS). Since the NRT product undergoes less rigorous quality assurance, we use only fires with ‘nominal’ or ‘high’ confidence levels from the NRT product for fire tracking. Some active fire detections from the NRT data are potentially associated with static non-vegetation fires (e.g., fires from gas flaring in oil and gas or landfill industries or false detections due to reflection from solar panels) and are not the main interest for vegetation fire studies. To avoid the unnecessary computation associated with these static fires, we record and evaluate the fire pixel density for each fire object at each time step. When a small fire ( 20 per km2), it is considered to be a static fire and subsequently labelled as invalid.Similar to the retrospective FEDS, we use the active fire detections to create an object serialization product, a regional snapshot GIS product, and a time series product of large fire evolution twice daily. This experimental NRT data will be available upon publication through a university hosted server. More

Study areaGrassland herbage samples are from Shaerqin base, institute of grassland research of CAAS (Chinese Academy of Agricultural Sciences). We obtained the permission of the institution to take HSI of the grassland sample. Our work did not cause damage to grassland. Researcher Weihong Yan of the institute provided us with relevant information about grassland. The land use type in the study area is mainly grassland, which is composed of forage species, most of which are representative species of typical grassland. We take this area as an example to conduct research on grass classification. By enriching the relevant recognition technology, it can also be used as a reference for the pastures of other grasslands. The grass species Grass1 for the experiment is shown in Table 1. The official introduction of plant materials is detailed in the flora of China15.Table 1 Samples information for Grass1 dataset.Full size tableThe field hyperspectral platformWe assemble a system for collecting HSI in the field: HyperSpec©PTU-D48E HSI instrument, high-precision scanning PTZ, tripod, data analysis software Hyperspec, etc. The light source is natural light. The imaging instrument is in line scanning mode. Table 2 shows the technical parameters.Table 2 Technical parameters of hyperspectral instrument.Full size tableData collectionIn July 2021, the data was collected during the lush grass growth period. Collect data from 11:00 a.m. to 2:00 p.m. every day. At this time, it is sunny, cloudless and the wind force does not exceed level 2. So as to ensure the consistency of the acquisition time line and avoid the influence of different degrees of light on the reflectivity as far as possible. The measuring points are arranged facing the sun and the opposite direction of the shadow. We collect data from different angles of the grassland, which is based on the growth of various types of forages, and selects relatively concentrated places within the study area. Each shot is a single category of grass. The image resolution is 1166 × 1004 pixels (Fig. 1). The imaging spectrometer is fixed with scanning head when shooting. Data acquisition and transmission are executed on Hyperspec software. Then save it as a BIL file. The ENVI5.3 software was used to extract the forage spectrum to establish the dataset Grass1. Well balanced regions with a clear image, uniform spectral distribution are selected for further segmentation. The average value of spectral reflectance of grass pixels was taken as the reflectance spectrum of a single type of grass.Figure 1True color map of grass samples.Full size imageMethodologyIn Fig. 2, we present the framework of visible-NIR hyperspectral classification of grass based on multivariate smooth mapping and extreme active learning (MSM–EAL). Specifically, we first introduce the proposed MSM algorithm for global enhanced spectral reconstruction, which utilizes smooth manifold projection technology to alleviate the problems of difficult feature selection and redundant data. Then, the EAL framework is proposed to address the matter of hyperspectral labeled samples and spectral classification. In the following, each step of this method will be presented in detail.Figure 2Proposed MSM–EAL framework for grass HSI classification.Full size imageThe proposed MSM algorithmIn the process of field HSI acquisition, on the one hand, the surface distribution of grass is uneven and the plant height is different, causing certain scattering effect and coverage spectrum change. On the other hand, HSI is easy to be disturbed by external natural factors such as light, wind and shadow, resulting in a certain degree of distortion. Multiplicative scatter correction (MSC) is a scattering correction effect, which helps to eliminate the scattering effect caused by the above reasons and enhance the spectral variability. The moving window smooth spectral matrix (Nirmaf) belongs to the smooth effect, which improve the signal-to-noise ratio of the spectrum and reduce the influence of random noise16,17. Preprocessing methods are different and related to each other. We design an enhanced preprocessing multivariate smooth (MS) method that fusing MSC and smooth Nirmaf to target grass spectral signal features. In the follow-up, a model will be established to verify the validity of MS.Most of the high-dimensional spatial data have the characteristics of being embedded in a manifold body, so the manifold learning isometric feature mapping (Isomap) based on spectral theory is adopted. Isomap preserves the global geometric features of the initial data and extracts features by reconstructing the underlying smooth manifold of HSI. It is nonlinear dimensionality reduction based on linear and multidimensional scaling transformation18. Isomap has been applied in image and HSI classification19,20, but there is no report on visible-NIR hyperspectral classification of grass.In view of the above, we proposed the multivariate smooth mapping (MSM) spectral reconstruction algorithm, which can be represented as follows:$$ MSM_{z} { } = { }frac{{left( {P_{j} – b_{j} } right)left( {2n + 1} right) + n_{j} cdot mathop sum nolimits_{j = – n}^{n} C_{j} P_{k + j} }}{{n_{j} left( {2n + 1} right)}} + V_{Z} F_{Z}^{frac{1}{2}} { } $$
(1)
where Pj, bj, and Cj represent the raw reflectance value of spectrum j, baseline shift amount, and weight factor, respectively, k and nj represent the polynomial degree and offset, respectively. MSMz is the feature cube reconstructed to Z dimension from the spectrum calculated by 2n + 1 moving window width, V eigenvector matrix and F eigenvalue matrix.In Isomap equidistant mapping, the shortest path of edge Pi Pj needs to be solved, and the representation matrix is:$$ D_{G} = [d_{G}^{2} (P_{i} ,P_{j} )]_{i,j = 1}^{n} $$
(2)
where d (Pi, Pj) is the weight of the edge Pi Pj calculated from the neighborhood graph G and its side Pi Pj.The proposed EAL frameworkLabeling hyperspectral samples is expensive in terms of time and cost, at the same time, the lower spatial resolution and more bands increase the difficulty of labeling. Active learning (AL) provides an efficient labeling strategy, which only needs to label a relatively small number of samples to learn a more accurate model21. The pool-based AL selects the most informative samples according to the query strategy for limited labeling through iteration, so as to facilitate model improvement. Commonly used query strategies are uncertainty criteria, such as least confidence22, the bayesian active learning disagreement (BALD), the entropy sampling23, etc.Due to there is still an over-fitting problem, different strategies such as hybrid prediction and regularization need to be used for non-recursive datasets24. The research25 proposed that extreme gradient boosting algorithm (XGBoost) based on gradient boosting. As a classification method, XGBoost has been successfully applied in Kaggle competition and other fields. Its most important feature for visible-NIR hyperspectral classification is that can easily and directly classify according to features, and the physical interpretation of features can help understand the electronic nature behind spectral classification. XGBoost is a machine learning algorithm based tree structure that integrates multiple weak classifiers to achieve flexible and high-precision classification. It is an upgraded version of gradient boosting decision tree. The optimization process of XGBoost entailed: (1) Expanding the objective function to the second order, and finds a new objective function for the new base model to improve the calculation accuracy. (2) L2 regularization term is added to the loss function to prevent over-fitting. (3) Using blocks storage structure realize automatic parallel computing26,27. The algorithm steps are as follows:The objective function:$$ Lleft( Phi right) = mathop sum limits_{i} lleft( {y^{i} ,widehat{{y^{i} }}} right) + mathop sum limits_{k} Omega left( {f_{k} } right) $$
(3)
In formula (3), the first and second terms are the loss function term and the regularization term, respectively. Where,$$ Omega left( {f_{k} } right) =upgamma {text{T}} + frac{1}{2}lambda left| w right|^{2} $$
(4)
γ and λ are regularization parameters which are used to adjust complexity of the tree.Next, second derivative Taylor expansion of the objective function. Where (g_{i}) and (h_{i}) are the first derivative and second derivative, respectively.$$ L^{left( t right)} = mathop sum limits_{i = 1}^{n} lleft( {y_{i} ,widehat{{y_{i}^{t – 1} }} + f_{t} left( {x_{i} } right)} right) + Omega left( {f_{t} } right) $$
(5)
$$ g_{i} = partial_{{hat{y}_{i} (t – 1)}} lleft( {y_{i} ,widehat{{y_{i}^{t – 1} }}} right) $$
(6)
$$ h_{i} = partial_{{widehat{{y_{i} }}(t – 1)}}^{2} lleft( {y_{i} ,widehat{{y_{i}^{t – 1} }}} right) $$
(7)
$$ {text{L}}^{left( t right)} approx mathop sum limits_{i = 1}^{n} left[ {lleft( {y_{i} ,widehat{{y_{i}^{t – 1} }}} right) + g_{i} f_{i} left( {x_{i} } right) + frac{1}{2}h_{i} f_{t}^{2} left( {x_{i} } right)} right] + Omega left( {f_{t} } right) $$
(8)
Final objective function:$$ {hat{text{L}}}^{ i} left( q right) = – frac{1}{2}mathop sum limits_{j = 1}^{T} frac{{(mathop sum nolimits_{{i in I_{j} }} g_{i} )^{2} }}{{mathop sum nolimits_{{i in I_{j} }} h_{i} + lambda }} + gamma T $$
(9)
Equation (9) can be used as the fraction of tree cotyledons, and the tree structure is directly proportional to the fraction. If the result after splitting is less than the maximum value of the given parameter, the cotyledon depth stops growing24,28.AL solves the problems of limited number and high cost of grass hyperspectral labeling samples. The default model of traditional AL is logistic regression, which is mostly studied on the ideal public dataset. However, the actual data has more uncertain noise, which still poses a certain challenge to AL. Consequently, we propose the extreme active learning (EAL) framework to minimize the classification cost of visible-NIR hyperspectral. The framework replaces the logistic regression model with XGBoost. Taking advantage of AL, XGBoost can improve performance with less training marker samples. By jointing of XGBoost and AL, EAL provides significantly better results than AL in field Grassl dataset recognition. Additionally, based on the characteristics of XGBoost, EAL more intuitively enhances the physical essence behind spectral classification than AL. Algorithm 1 summarizes the workflow of EAL framework.Random forest (RF) and decision tree (DT) were used to compare with EAL. RF and DT are frequently used in the field of grassland remote sensing9,29. Furthermore, RF, DT and XGBoost have the same point is that are learning algorithms based on tree structure. DT determines the direction by judging the conditions of the decision node12. RF is an integrated learning of multiple decision trees30. More