Divergent abiotic spectral pathways unravel pathogen stress signals across species
Airborne hyperspectral and thermal image acquisitionWe scanned over one million olive and almond trees between 2011 and 2019 with an airborne imaging spectroscopy and thermal imaging facility targeting infected and healthy trees in seven different regions located in Apulia (Italy), Majorca (Balearic Islands, Spain), Alicante, Cordoba and Seville (mainland Spain). In olive groves, over 200,000 and 372,000 trees were imaged from Xf and Vd outbreaks, respectively. In almond groves, we scanned over 132,000 trees from Xf outbreaks in Alicante and Majorca. To evaluate the effects induced by abiotic stress on spectral plant traits, we surveyed over 370,000 healthy trees (outside the outbreak areas) comprising olive and almond species subjected to a wide range of water stress conditions.We surveyed these areas with airborne hyperspectral and thermal cameras on board a manned aircraft flying at 500 m altitude above ground, yielding 40 cm and 60 cm spatial resolution, respectively. We used a hyperspectral camera (VNIR model, Headwall Photonics, Fitchburg, MA, USA) collecting 260 bands in the 400–885 nm region at 1.85 nm/pixel and 12-bit radiometric resolution with a frame rate of 50 Hz. With this spectral configuration, we captured imagery at 6.4 nm full-width at half-maximum (FWHM) bandwidth and obtained an instantaneous field of view (IFOV) of 0.93 mrad and an angular field of view (FOV) of 49.82 deg with an 8 mm focal length lens. The hyperspectral sensor was radiometrically calibrated in the laboratory using an integrating sphere (CSTM-USS-2000C Uniform Source System, LabSphere, North Sutton, NH, USA). At the time of flight, we measured aerosol optical thickness at 550 nm using a Sunphotometer (Microtops II S model 540, Solar LIGHT Co., Philadelphia, PA, USA). We then applied the resulting atmospheric correction of the calibrated radiance imagery with the SMARTS model51 to derive surface reflectance spectra. We carried out ortho-rectification of the hyperspectral imagery (PARGE, ReSe Applications Schläpfer, Wil, Switzerland) with readings acquired by the inertial measuring unit on board the airborne platform (IG500 model, SBG Systems, France). We applied spatial binning through object-based image analysis, thus increasing the signal-to-noise ratio (SNR) of the instrument. Additionally, we conducted spectral binning to reduce the number of spectral bands (260 bands at 1.85 nm resolution). SNR reached >300:1 after binning. We acquired high-resolution tree-crown temperature images with a thermal camera (FLIR SC655, FLIR Systems, USA) at 640 × 480 pixels resolution using a 24.6 mm f/1.0 lens, sensitive to the 7.5–14 μm spectral range and sensitivity below 50 mK.We identified individual trees in the high-resolution hyperspectral and thermal images using object-based crown detection and segmentation methods52. We then calculated the mean hyperspectral radiance, reflectance and temperature for each pure tree crown within every orchard under evaluation. We based our image segmentation methods on Niblack53 and Sauvola and Pietikäinen54, which allowed the isolation of tree crowns from the soil and shadow components. The segmentation of each tree crown was assessed visually to ensure a minimum number of pure vegetation pixels were selected within each tree crown and also spectrally to evaluate the purity of the reflectance extracted from the crown to avoid spectral mixture with soil, shadows and background components24,35.Collection of Xf and Vd biotic stress field dataField assessments of Xf- and Vd-infected trees were carried out from outbreaks affecting olive and almond species in the indicated regions of Italy and Spain between 2011 and 201924,35,52. During these campaigns, we performed quantitative PCR (qPCR)55 for Xf in olive and almond (Alicante), recombinase-polymerase-amplification (RPA) using the AmplifyRP XRT + test (Agdia®, Inc., Elkhart, IN)56 for Xf in almond (Majorca) or conventional PCR57 assays for Vd, as well as visual assessments in individual trees of disease incidence (DI) and disease severity (DS). A sample was considered positive if Ct values were ≤36 and amplification curves were exponential. PCR/qPCR data for model analysis were transformed to 0 and 1, for negative and positive results, respectively, and Ct values were not used in the analysis (see Supplementary Table 2 for the PCR/qPCR primer sequences for Vd and Xf). DS was scored using a 0–4 rating scale according to the percentage of the tree crown showing disease symptoms.In Apulia, the Xf-olive database comprised a total of 15 olive groves surveyed during the June 2016 and July 2017 campaigns. Visual assessments for infection were conducted on 7296 trees (3324 in 2016 and 3972 in 2017). In 2016, 1886 symptomatic (and 1438 asymptomatic) trees were surveyed (762 trees labelled as DS = 1; 802 DS = 2; 250 DS = 3 and 72 DS = 4). In 2017, 1365 were reported as symptomatic (and 2607 asymptomatic) (686 DS = 1; 542 DS = 2; 122 DS = 3 and 15 DS = 4). qPCR assays were carried out to diagnose Xf infection in 77 olive trees, whereby 39 trees tested negative (qPCR = 0) and 38 tested positive (qPCR = 1).On the island of Majorca and at the Alicante province, the field-based Xf-almond database comprised a total of 19 almond groves surveyed in 2018 and 2019, respectively. In Alicante, the field surveys covered 83 ha with 9 almond groves consisting of 943 almond trees. During the field campaigns, almond trees were visually assessed to evaluate Xf-induced DI and DS indices. From this analysis, we identified 593 symptomatic trees and 350 asymptomatic trees. Out of all symptomatic trees, 163 were rated as DS = 1, 214 DS = 2, 157 DS = 3, and 59 DS = 4. Furthermore, qPCR analysis was carried out on 226 almond trees to diagnose Xf infection, resulting in 48 non-infected (qPCR = 0) almond trees and 178 infected trees (qPCR = 1). In Majorca, field surveys in July 2019 covered a total of 2803 ha and comprised 10 almond groves. During the field campaigns, visual observations were carried out on over 4048 almond trees to assess DI and DS, yielding 1387 symptomatic and 2661 asymptomatic trees. From symptomatic trees, 537 were rated as DS = 1449 DS = 2, 359 DS = 3 and 42 DS = 4. We conducted AmplifyRP XRT + assays on 265 almond trees for diagnosing Xf infection the same day they were sampled and identified 141 negative trees (qPCR = 0) and 124 positive trees (qPCR = 1).We carried out physiological measurements of leaf chlorophyll, anthocyanins, flavonoids and nitrogen contents with a Dualex Scientific + (Force-A, Orsay, France) instrument as well as leaf reflectance (400–1000 nm spectral range) and steady-state chlorophyll fluorescence (Ft) using the PolyPen RP400 and FluorPen FP100 instruments (Photon Systems Instruments, Drasov, Czech Republic) during the field evaluations of almond and olive groves conducted in Majorca, Alicante and Apulia regions. In the Xf-olive study site in Apulia, we generated 1023 leaf measurements with Dualex, 1543 single leaf reflectance spectra, as well as 1402 Ft readings over 67 olive trees. In the Xf-almond study sites in Majorca, we measured 1242 leaves with Dualex, 1094 leaves with the PolyPen and 1218 with the Fluorpen instruments from 87 almond trees across a wide range of disease severity levels. For the Xf-almond study sites located at Alicante, we conducted 1649 leaf measurements with Dualex, 632 leaf measurements with PolyPen and 563 leaf measurements with FluorPen FP100 over 43 almond trees.We assessed Vd-infected olive trees from 11 olive groves by surveying an area of over 3000 ha in Castro del Rio and Ecija, southern Spain, in 2011 and 2013, respectively. In Castro del Rio, we conducted visual assessments in an infected area of 96 ha comprising 1878 olive trees, thus identifying 1569 asymptomatic and 283 symptomatic olive trees. Out of the 283 symptomatic trees, 218 were rated as DS = 1; 45 DS = 2; 12 DS = 3 and 8 DS = 4. We measured leaf Fs and Fm’ fluorescence parameters from 25 leaves per tree using a PAM-2100 Pulse-Amplitude Modulated Fluorometer (Heinz Walz GMBH, Effeltrich, Germany). In addition, leaf PRI570 was measured from 25 leaves per tree using a custom-made PlantPen device (Photon System Instrument, Drasov, Czech Republic). Finally, we measured leaf conductance (Gs) on five leaves per tree using a leaf porometer (model SC-1, Decagon Devices, Washington, DC, USA). In the Écija region, the surveyed area covered 3424 ha, and 5223 olive trees were evaluated. We performed visual assessment to determine DI and DS indices of Vd-infected trees, identifying 5040 asymptomatic olive trees. Of the remaining 183 olive trees that were symptomatic, 112 were trees rated as DS = 1; 41 DS = 2; 22 DS = 3 and 8 DS = 4.Trees were evaluated for disease severity and incidence by visual assessment in each outbreak region. PCR assays were carried out on a subset of these trees within each orchard to (i) validate that the pathogen (Xf or Vd) was actually present and the biotic source of symptoms; and (ii) validate that asymptomatic (DS = 0) but infected (PCR = 1) trees were detected using the hyperspectral plant traits estimated through the methodology described in this paper. In general, PCR assays are (i) time consuming and costly, and (ii) difficult to make in large infected trees due to the non-uniform distribution of the infection within each tree crown. These PCR data for each tree along with the field evaluations of DS, DI and non-destructive physiological measurements derived for each tree within every orchard were matched with the high-resolution hyperspectral images to build the biotic databases used in this study. We carried out the field work at each orchard guiding the evaluations and measurements using a high-resolution image to map the location of each tree within the orchard. Due to the planting grids typical of almond and olive species, which were not contiguous or in row-structured patterns, the identification of each individual tree in the images was straightforward.Collection of abiotic stress field dataWe monitored over 3600 ha of olive and almond groves located outside any infected area in Cordoba and Seville, Southern Spain. We performed a multitemporal analysis to study the spectral plant-trait alterations induced by abiotic stress relative to healthy olive and almond trees with data we collected over a 468 ha area comprising two olive and one almond groves throughout July 2016 and August 2017 growing seasons. We analysed 2975 olive and 1964 almond trees in 2016, and 2865 olive and 2063 and almond trees in 2017. At both study sites, we monitored the midday stem water potential (SWP) using a pressure chamber (Soil Moisture Equipment Corp. model 3000, Santa Barbara, CA, USA) on 16 trees per grove. SWP values showed differences between two existing irrigation levels (well-watered and mild water stress), averaging –1.7 and –1.9 MPa across the season in the case of almonds. In olive, SWP in one of the groves reached –3.8 and –3.5 MPa. In 2017, water potential levels averaged –2.9 and –2.0 MPa. In the second grove, irrigation levels were higher, reaching an average SWP of –1.5 MPa. We used an additional study site located in Casariche (Seville province), southern Spain, to validate the results obtained from the multitemporal analysis. This study site covered 3371 ha containing 55 olive groves grown under irrigated and rainfed conditions, with 21,071 olive trees used for statistical analysis.The multitemporal dataset was used to evaluate the water-induced abiotic stress by quantifying the evolution of the importance of the most sensitive spectral traits by clustering non-stressed trees (C0) against groups of trees exposed to increasing levels of water stress (C1 to C4). The multitemporal component of this assessment enabled the evaluation of every single tree across time, therefore selecting the trees for each cluster based on a sustained water stress level, avoiding the selection of trees under short-term stress dynamics. Thus, the clusters were determined based on their CWSI levels, and only the trees with stable water stress levels across two consecutive years (2016 and 2017) were selected for the analysis. For this purpose, we did not include trees that deviated beyond 95% of the CWSI differences calculated between the first and second year in the analysis. After this trimming step, we retained 5484 olive trees (from 5566 trees) and 3652 almond trees (from 3882 almond trees). Trees were then grouped through CWSI clustering analysis using a modified three-sigma rule58. This rule describes the density of a distribution within standard deviation bands on both sides of the mean point into the 68th, 95th and 99.7th percentiles58, representing µ ± σ, µ ± 2σ and µ ± 3σ, respectively. The first interval defined by the classic three-sigma rule (µ ± σ) represented most trees, while the third interval (µ ± 3σ) consisted of very few trees, raising issues for the determination of statistical significance analysis. Based on this observation, we adjusted the breakpoints between groups as follows: we classified those trees that were in the lowest 10th percentile as C0. Trees between the 10th and 68th percentiles (µ + σ) were classified as C1, trees between the 68th and 85th percentile were classified as C2, trees between the 85th and 95th percentile were classified as C3 and finally the trees over the 95th (µ + 2σ) percentile were classified as C4. We thus selected 488 C0, 3066 C1, 1090 C2, 618 C3 and 222 C4 olives trees. Likewise, we grouped almond trees into 390 C0, 1776 C1, 1248 C2, 214 C3 and 24 C4 clusters. Moreover, the analysis of the contribution of a given trait was performed using ML modelling strategies to classify unstressed trees against the clusters defined above that were exposed to increasing levels of water stress. Furthermore, we assessed the consistency of the obtained indicators by performing the classification between stressed and non-stressed trees at an independent olive study site. For this purpose, we evaluated our predictors and compared their contribution over an additional site (Casariche).Model inversion methods for plant-trait estimationWe quantified chlorophyll content (Ca+b), carotenoid content (Cx+c), anthocyanin content (Anth.), mesophyll structure (N), leaf area index (LAI) and average leaf angle (leaf inclination distribution function or LIDF) by radiative transfer model inversion of PROSPECT-D59 and 4SAIL60, as in Zarco-Tejada et al.24. We inverted PROSPECT-D + 4SAIL using a look-up-table (LUT) generated with randomised input parameters. The LUT was generated with 100,000 simulations within fixed ranges (Supplementary Table 3). We implemented a wavelet analysis61 into six wavelets by a Gaussian kernel, estimating the parameters in the top 1% entries ranking the lowest root mean square error (RMSE) values. We then retrieved each plant trait independently by training supported vector machine (SVM) algorithms using the simulated reflectance data as input. We built SVMs in Matlab (MATLAB; Statistics and Machine Learning toolbox and Deep Learning toolbox; Mathworks Inc., Matick, MA, USA) using a Gaussian kernel (radial basis function) with hyperparameters optimised for each model. The training processes were carried out in parallel using the Matlab parallel computing toolbox. With these trained models, we then used the spectral reflectance extracted from the delineated crowns (as show in Fig. 1) to predict plant traits for each individual tree at each study site. The model inversions were carried out for each tree using the crown reflectance. The latter was calculated as an average across all the pixels belonging to the tree crown, delineated using segmentation. This method52 avoids the problem of pixels from within-crown shadows, from tree edges or from sunlit or shaded soil background affecting the spectra, as it retrieves the plant traits from pure sunlit vegetation components of the trees. We also calculated narrow-band spectral indices from reflectance spectra (Supplementary Table 1), which are sensitive to leaf traits and potentially related to disease-induced symptoms. Tree-crown radiance and temperature were used to calculate sun-induced chlorophyll fluorescence at 760 nm (SIF@760) and the crop water stress index (CWSI)37. SIF@760 was quantified using the O2-A in-filling Fraunhofer Line Depth (FLD) method63 and CWSI was calculated by incorporating the tree temperature and the weather data obtained at each study site37.Statistical analysisWe implemented random forest (RF)64 algorithms to classify healthy vs. infected (biotically stressed) trees, and non-stressed vs. water (i.e. abiotically) stressed trees for both tree species. RF algorithms have been widely used in remote sensing studies since they have shown excellent classification accuracies and high processing speeds with high-dimensional data62 and have shown to be accurate in detection of several diseases29,65,66,67. The spectral plant traits estimated by radiative transfer model inversion (Ca+b, Cx+c, Anth., LAI and LIDF), CWSI and SIF@760 were used as inputs for the models. In addition, using a recursive feature elimination approach68 the narrow-band indices that improved the classification in terms of overall accuracy (OA) and kappa coefficient (κ) were added to the models. The pool of narrow-band indices was reduced based on a variance inflation factor (VIF) analysis69 to avoid collinearity among the input features.The RF algorithms were built in Matlab and the hyperparameters were optimised using Bayesian optimisation. The importance of a feature using the RF algorithm was assessed based on the permutation of out-of-bag (OOB) predictor methodology70. To compare the relative differences of the spectral traits in classification of the biotic and abiotic stress, the importance was normalised by dividing the importance of each trait by the highest contribution obtained for each pathogen/species. For the RF models, 500 iterations were run by randomly partitioning each dataset into training (80% of samples) and testing sets (20% of samples). For the training subset, a balanced number of samples from each class was randomly selected at each iteration. The importance obtained by the OOB permutation algorithms was used to build a feature-weighted random forest algorithm (based on Liu and Zhao45), accounting for the importance of each variable on the classification process, evaluating the model against PCR data and visual observations for each biotic stress dataset in terms of OA and κ levels.Probabilities of the predictions were obtained for each sample71 and the uncertain trees were assessed. To extract the uncertainty for each individual tree on the classification, we evaluated the probability distribution for each class from each dataset independently. Then, those trees with a classification probability below the 68th percentile (µ [mean] + σ [standard deviation]) were considered as uncertain and incorporated into a second-stage classification process. The second stage consisted of an unsupervised graph theory–based spectral clustering algorithm72 and included traits selected by focusing on the divergent biotic–abiotic stress obtained from the biotic and the abiotic stress databases. Spectral clustering was performed in R using the kernlab package73.To determine the spectral traits that differed between Xf- and Vd-infected plants and those from the abiotic pathway, we first normalised the importance of the specific traits independently. Then, we compared the common traits between abiotic and biotic stress sets, selecting only biotic stress-related traits that differed in ratio by >0.5 over their homologous abiotic stress trait values. Traits that were only expressed under biotic stress conditions and that showed a normalised importance over 0.5 were included for the second-stage classification process only including those divergent-specific biotic and abiotic stress-related spectral traits as inputs. Specifically, NPQI, Anth. and SIF@760 were considered for the classification of Xf-infected olive trees. Ca+b, SIF@760 and PRIn were used for classifying Xf-infected almond trees. Furthermore, NPQI, Anth. and B spectral traits were selected for classifying uncertain Vd-infected olive trees. Finally, we validated our feature-weighted methodology coupled with the second-stage spectral clustering method against qPCR assays and visual assessment of symptom severity.Reporting summaryFurther information on research design is available in the Nature Research Reporting Summary linked to this article. More
