We hypothesized that functionally distinct forest types can be mapped at moderate spatial resolutions, using a combination of canopy foliar traits and canopy structure information. Our analysis of LiDAR and imaging spectroscopy data at spatial resolutions ranging from 4 to 200 m (16 m2–40,000 m2), with an emphasis on the 30 m (900 m2) spaceborne hyperspectral spatial resolution, reveals that few remotely sensed canopy properties are needed to successfully identify ecologically distinct forest types at two diverse tropical forest sites in Malaysian Borneo. In testing our second hypothesis that mapped forest types exhibit distinct ecosystem function, we found that forest types identified using remotely sensed leaf P, LMA, Max H, and canopy cover at 20 m height (Cover20) closely align with forest types defined from field-based floristic surveys29,30,31,32,33 and inventory plot-based measurements of growth and mortality rates (Fig. 4b). Our approach, however, enables mapping of their entire spatial extent (Fig. 1) and reveals important structural and functional variation within areas characterized as a single forest type in previous studies (Fig. 3). Current and forthcoming satellite hyperspectral platforms, including PRISMA (30 m), CHIME (20–30 m), and SBG (30 m), have or will have comparable spectral resolution, higher temporal revisits, and much greater geographic coverage. The ability to conduct this type of analysis using remote sensing measurements at 30 m resolution suggests that our method can be applied to these emerging spaceborne imaging spectroscopy data to reveal important differences in structure and function across the world’s tropical forests.
Nested functional forest types revealed
To test our first hypothesis, rather than making an a priori decision about the number of k-means clusters (k), we explored the capacity of remotely sensed data to reveal ecologically relevant variation in forest types. Baldeck and Asner took a similar unsupervised approach to estimating beta diversity in South Africa34. Because the choice of k directly influences analysis outcomes, careful selection of k is required. Different approaches for identifying the number of clusters, using the Gapk and Wk elbow metrics35, yielded varying optimal numbers of clusters for the Sepilok and Danum landscapes (Fig. 1, Supplementary Figs. 4 and 5). However, at both sites, a comparison of results based on different values of k revealed ecologically meaningful structural and functional differences and graduated transitions between forest types (Fig. 2, Supplementary Figs. 7 and 8), indicating that the exploration of traits that aggregate or separate forest types as k changes is a valuable exercise. Overlap between the remotely sensed forest type boundaries and inventory plots within distinct forest types indicate that the series of clustered forests align closely with forest types defined based on in situ data on species composition and ecosystem structure. In part, this type of analysis requires careful selection of the number of clusters. Additionally, however, we gained valuable insights via the exploration of varying numbers of clusters as it relates to biologically meaningful categorization of forest types. Extending this method to other parts of the tropics will require similar decision-making, which will either require user input, or the development of robust automated algorithms for selecting k.
Forest types capture differences in ecosystem dynamics
We further evaluated the canopy traits and structural attributes that were most critical for mapping distinct forest types, hypothesizing that mapped forest types exhibit distinct ecosystem function. Forest types revealed by the cluster analyses were distributed along the leaf economic spectrum, where the leaf economic spectrum characterizes a tradeoff in plant growth strategies36. LMA, which can covary strongly with leaf N and P, is a key indicator of plant growth strategies along the spectrum37. At the slow-return end of the leaf economics spectrum, plants in nutrient-poor conditions with low leaf nutrient concentrations invest in leaf structure and defense, expressed as high LMA, strategizing longer-lived, tougher leaves with slower decomposition rates. This strategy comes at the cost of slower growth. At the quick-return end of the spectrum, plants in nutrient-rich environments with higher leaf nutrient concentrations invest less in structure and defense, enabling faster growth and more rapid leaf turnover, i.e., shorter leaf lifespans. This quick-return growth strategy supports higher photosynthetic rates and more rapid carbon gain36.
In this study, the principal components and clustering results yielded forest types that are indicative of community level differences associated with leaf economic spectrum differences. The nutrient rich sites (Danum1 and Danum2, Supplementary Fig. 8) show high canopy N and P and low LMA compared to the nutrient poor and acidic sites (Sandstone and Kerangas), which contributes to lower leaf photosynthetic capacity (Vcmax) and growth (Fig. 4b). Foliar N:P also increased with site fertility, confirming that tropical forests are primarily limited by phosphorus, and not nitrogen38,39, with large implications for carbon sequestration in these forests. Orthogonal differences in canopy structure and architecture between Danum forest types and Sepilok Sandstone and Alluvial forests could be indicative of ecosystem scale differences in the sensitivity of these forests to endogenous disturbance processes40.
The significant differences in aboveground carbon stocks and growth and mortality rates between forest types further suggests strong differences in ecosystem dynamics. In general, growth rates varied inversely to aboveground carbon, and higher aboveground carbon corresponded to lower mortality rates. As an example, the Sepilok sandstone forests, which are largely comprised of slow-growing dipterocarp species29,33, had the highest median aboveground carbon (236 Mg C ha−1), with higher canopy P and N, and lower LMA. The taller canopy and low canopy leaf nutrient concentrations are consistent with the low growth and mortality rates found in the sandstone forest, indicating a slow-growth strategy yielding larger trees and higher aboveground carbon stocks. In contrast, alluvial forests exhibit high turnover with mortality and growth rates higher relative to Sandstone forests corresponding to lower aboveground carbon on average. Kerangas forests exhibited low aboveground carbon despite an intermediate plot-level growth rate, and mortality rates that were significantly lower than the Danum or alluvial forest types. Kerangas forests, which were characterized by the highest LMA, lowest foliar P and N (Fig. 2a), and the lowest plot-level aboveground carbon density (186 Mg C ha−1; Fig. 4a), are known to have higher stem densities, lower canopy heights, and long-lived leaves5,32,41, suggesting well-developed strategies for nutrient retention42. Interestingly, despite significantly different aboveground carbon and demography, the kerangas and sandstone forests did not differ in LAI or canopy architecture (P:H); although maximum height, Cover20, and Hpeak LAI were significantly higher in the sandstone forest, highlighting the need to account for differences beyond LAI when scaling processes from leaves to ecosystems.
In addition, when three forest types were distinguished at Sepilok, the alluvial inventory plot had significantly higher aboveground carbon than the remote sensing-derived alluvial forest extent (Fig. 4a, p < 0.001). It was only when the mudstone and alluvial forests were differentiated when k = 4 that the inventory plot and clustered alluvial forest areas exhibited similar aboveground carbon distributions, with significantly lower carbon in the mudstone forest. Although Sepilok mudstone and alluvial forests are often characterized as a single forest type5,43, independent research first identified mudstone hills as unique based on differences in soil cation exchange capacity, pH, and nutrient concentrations that translated into intermediate plant growth rates in mudstone forests44. More recently, higher clay fractions and higher exchangeable Mg, Ca, and K were found at varying soil depths in Sepilok mudstone forests compared to alluvial forests, although alluvial forests exhibited higher foliar N, P, K, and Mg concentrations compared to mudstone forests30. Our remote sensing findings independently support the uniqueness of mudstone forests based on both leaf traits and structural attributes (Figs. 2 and 3). The lower aboveground carbon in the mudstone forest may be due to lower leaf nutrient concentrations and higher soil acidity, as well as differences in hydrology associated with seedling and sapling responses to flooding that influence the species assembly45,46. Because the mudstone forests in Sepilok are also closer to anthropogenic forest edges than alluvial forests, edge effects—which have been shown to significantly influence large tree mortality and lower aboveground carbon—may also be a factor47,48,49.
At Danum, our results indicate that the region is comprised of one to three forest types that differ in LMA, foliar N and P, canopy height, and vertical structure (Figs. 1 and 2; Supplementary Figs. 7–8). Our finding that two of these forest types (Danum 1 and 2) were found within the 50 ha Smithsonian ForestGEO inventory plot interestingly aligns with recent, independent research. Differences in species composition and soil characteristics have been identified between the northeast corner and the remainder of the 50 ha plot30. A recent study also identified the northeast corner (Danum 1) has having lower species richness, diversity, stem density, and basal area compared to the rest of the plot (Danum 2), linked to less acidic soils with a higher cation exchange capacity and higher Ca, Mg, and Ni content31.
Implications for modeling tropical forest biomes
We did not find significant variation in total ecosystem LAI across forest types in this study. In contrast, vertical variation in structure was more strongly linked to differences between functionally distinct forest types. LAI is considered one of the most important ecophysiological attributes of vegetation, and is widely used in terrestrial ecosystem and biosphere models to upscale estimates of leaf-level processes to ecosystems and to model land atmosphere interactions26,27. LAI varies significantly among the world’s biomes50. Within tropical forests, previous studies have shown that variation is correlated with maximum water deficit, minimum temperature, and forest protection status51. However, as shown here, LAI variation does not vary significantly across the different lowland forest types (alluvial, sandstone, mudstone, and kerangas) found at Sepilok and Danum. Average canopy LAI (estimated from the airborne LiDAR data using the method described in ref. 52) was similar across forest types, ranging from 5 to 6. A recent study argued that total LAI may not be directly relevant for many processes in ecosystems (e.g., productivity) beyond a value of three53.
However, our finding of significant variation in vertical LAI profiles, despite no substantial variation in average ecosystem LAI, provides important evidence that vertical foliar distributions may be more important than the absolute amount of leaf area for characterizing differences across ecosystems, underscoring the importance of evaluating additional LiDAR-derived metrics and leaf traits. In particular, the vertical distribution of leaf area is important for many canopy processes since the total amount of leaf surface area and its vertical organization can vary independently54,55. Although terrestrial biosphere models differ in their representation of vertical forest strata, there has been a growing effort to incorporate vertical variation more directly in many models56. Several recent model developments are at the cutting edge of representing vertical variation, which implement vertical gradients of irradiance, water content, and leaf temperature in ways that better enable models to capture differences in function within and between forest ecosystems57,58,59.
In conclusion, we explore the ability to map forest types at the spatial resolution of forthcoming hyperspectral satellite sensors and evaluate the ability of those forest types to capture differences in three aspects of ecosystem dynamics: aboveground biomass stocks, growth rates, and mortality rates. It was beyond the scope of this study to conduct ground validation of the entire extent of the mapped forest types. However, an important next step for further research entails a more detailed analysis of differences in forest composition, structure, and ecosystem dynamics at the larger landscape scale, beyond the inventory plots evaluated in this study. Establishing 1 ha plots and repeatedly censusing them across the extent of each forest type mapped would yield more robust accuracy assessments of the forest type maps reported here and will be critical for disentangling the mechanisms and processes underpinning differences in structure and function. Despite this need for further in situ analyses, this study provides the first step toward reliably characterizing differences in forest types over large areas, where forest inventory plots are not readily available. This type of mapping alone offers invaluable insights into differences across vast areas of tropical forest that are otherwise often characterized as a single biome in ecosystem models. The ability to capture this variation within the global critical tropical forest biome sets a new bar for terrestrial biosphere modeling. Results from this type of analysis can be used to parameterize and benchmark earth system models, further constraining high uncertainty in the future dynamics of these ecosystems.
Since the main axes of variation in canopy properties correspond to quantities measurable from spaceborne LiDAR and imaging spectroscopy, our approach offers a framework for large-scale mapping of functionally distinct forest types that can be employed across highly diverse tropical forest ecosystems at regional and global scales. Canopy leaf phosphorus (P) and leaf mass per unit area (LMA) were critical for distinguishing between forest types and will thus be essential to accurately map from spaceborne sensors for ecological applications. Maximum canopy height, and the fraction of canopy cover taller than 20 m, were important for distinguishing forest types, although variation in structure alone was insufficient to capture differences. These types of analyses at pan-tropical scales will be invaluable for improving understanding of ecosystem variation across a biome that is both incredibly threatened by and critical to mitigating global climate change. Our findings also underscore synergies between ground-based and remote-sensing ecological analyses, whereby landscape-scale remote surveys can efficiently pinpoint locations that can be targeted as high priority for discovery-oriented fieldwork and ground-based measurements.
Source: Ecology - nature.com
