In this study, we automate tree species classification and mapping using field-based training data, high spatial resolution airborne hyperspectral imagery, and a convolutional neural network classifier (CNN). We tested our methods by identifying seven dominant trees species as well as dead standing trees in a mixed-conifer forest in the Southern Sierra Nevada Mountains, CA (USA) using training, validation, and testing datasets composed of spatially-explicit transects and plots sampled across a single strip of imaging spectroscopy. We also used a three-band ‘Red-Green-Blue’ pseudo true-color subset of the hyperspectral imagery strip to test the classification accuracy of a CNN model without the additional non-visible spectral data provided in the hyperspectral imagery. Our classifier is pixel-based rather than object based, although we use three-dimensional structural information from airborne Light Detection and Ranging (LiDAR) to identify trees (points > 5 m above the ground) and the classifier was applied to image pixels that were thus identified as tree crowns. By training a CNN classifier using field data and hyperspectral imagery, we were able to accurately identify tree species and predict their distribution, as well as the distribution of tree mortality, across the landscape. Using a window size of 15 pixels and eight hidden convolutional layers, a CNN model classified the correct species of 713 individual trees from hyperspectral imagery with an average F-score of 0.87 and F-scores ranging from 0.67–0.95 depending on species. The CNN classification model performance increased from a combined F-score of 0.64 for the Red-Green-Blue model to a combined F-score of 0.87 for the hyperspectral model. The hyperspectral CNN model captures the species composition changes across ~700 meters (1935 to 2630 m) of elevation from a lower-elevation mixed oak conifer forest to a higher-elevation fir-dominated coniferous forest. High resolution tree species maps can support forest ecosystem monitoring and management, and identifying dead trees aids landscape assessment of forest mortality resulting from drought, insects and pathogens. We publicly provide our code to apply deep learning classifiers to tree species identification from geospatial imagery and field training data.

Reducing uncertainty of terrestrial carbon cycle depends strongly on the accurate estimation of changes of global forest carbon stock. However, this is a challenging problem from either ground surveys or remote sensing techniques in tropical forests. Here, we examine the feasibility of estimating changes of tropical forest biomass from two airborne lidar measurements of forest height acquired about 10 yr apart over Barro Colorado Island (BCI), Panama. We used the forest inventory data from the 50 ha Center for Tropical Forest Science (CTFS) plot collected every 5 yr during the study period to calibrate the estimation. We compared two approaches for detecting changes in forest aboveground biomass (AGB): (1) relating changes in lidar height metrics from two sensors directly to changes in ground-estimated biomass; and (2) estimating biomass from each lidar sensor and then computing changes in biomass from the difference of two biomass estimates, using two models, namely one model based on five relative height metrics and the other based only on mean canopy height (MCH). We performed the analysis at different spatial scales from 0.04 ha to 10 ha. Method (1) had large uncertainty in directly detecting biomass changes at scales smaller than 10 ha, but provided detailed information about changes of forest structure. The magnitude of error associated with both the mean biomass stock and mean biomass change declined with increasing spatial scales. Method (2) was accurate at the 1 ha scale to estimate AGB stocks (R² = 0.7 and RMSE_mean = 27.6 Mg ha⁻¹). However, to predict biomass changes, errors became comparable to ground estimates only at a spatial scale of about 10 ha or more. Biomass changes were in the same direction at the spatial scale of 1 ha in 60 to 64% of the subplots, corresponding to p values of respectively 0.1 and 0.033. Large errors in estimating biomass changes from lidar data resulted from the uncertainty in detecting changes at 1 ha from ground census data, differences of approximately one year between the ground census and lidar measurements, and differences in sensor characteristics. Our results indicate that the 50 ha BCI plot lost a significant amount of biomass (−0.8 Mg ha⁻¹ yr⁻¹ ± 2.2(SD)) over the past decade (2000–2010). Over the entire island and during the same period, mean AGB change was 0.2 ± 2.4 Mg ha⁻¹ yr⁻¹ with old growth forests losing −0.7 Mg ha⁻¹ yr⁻¹ ± 2.2 (SD), and secondary forests gaining +1.8 Mg ha yr⁻¹ ± 3.4 (SD) biomass. Our analysis suggests that repeated lidar surveys, despite taking measurement with different sensors, can estimate biomass changes in old-growth tropical forests at landscape scales (>10 ha).

There is an increasing interest in identifying theories, empirical data sets, and remote‐sensing metrics that can quantify tropical forest alpha diversity at a landscape scale. Quantifying patterns of tree species richness in the field is time consuming, especially in regions with over 100 tree species/ha. We examine species richness in a 50‐ha plot in Barro Colorado Island in Panama and test if biophysical measurements of canopy reflectance from high‐resolution satellite imagery and detailed vertical forest structure and topography from light detection and ranging (lidar) are associated with species richness across four tree size classes (>1, 1–10,>10, and>20 cm dbh) and three spatial scales (1, 0.25, and 0.04 ha). We use the 2010 tree inventory, including 204 757 individuals belonging to 301 species of freestanding woody plants or 166 ± 1.5 species/ha (mean ± SE), to compare with remote‐sensing data. All remote‐sensing metrics became less correlated with species richness as spatial resolution decreased from 1.0 ha to 0.04 ha and tree size increased from 1 cm to 20 cm dbh. When all stems with dbh>1 cm in 1‐ha plots were compared to remote‐sensing metrics, standard deviation in canopy reflectance explained 13% of the variance in species richness. The standard deviations of canopy height and the topographic wetness index (TWI) derived from lidar were the best metrics to explain the spatial variance in species richness (15% and 24%, respectively). Using multiple regression models, we made predictions of species richness across Barro Colorado Island (BCI) at the 1‐ha spatial scale for different tree size classes. We predicted variation in tree species richness among all plants (adjustedr²= 0.35) and trees with dbh>10 cm (adjustedr²= 0.25). However, the best model results were for understory trees and shrubs (dbh 1–10 cm) (adjustedr²= 0.52) that comprise the majority of species richness in tropical forests. Our results indicate that high‐resolution remote sensing can predict a large percentage of variance in species richness and potentially provide a framework to map and predict alpha diversity among trees in diverse tropical forests.

Most plant species are non-randomly distributed across environmental gradients in light, water, and nutrients. In tropical forests, these gradients result from biophysical processes related to the structure of the canopy and terrain, but how does species richness in tropical forests vary over such gradients, and can remote sensing capture this variation? Using airborne lidar, we tested the extent to which variation in tree species richness is statistically explained by lidar-measured structural variation in canopy height and terrain in the extensively studied, stem-mapped 50-ha plot on Barro Colorado Island (BCI), Panama. We detected differences in species richness associated with variation in canopy height and topography across spatial scales ranging from 0.01-ha to 1.0-ha. However, species richness was most strongly associated with structural variation at the 1.0-ha scale. We developed a predictive generalized least squares model of species richness at the 1.0-ha scale (R² = 0.479, RMSE = 8.3 species) using the mean and standard deviation of canopy height, mean elevation, and terrain curvature. The model demonstrates that lidar-derived measures of forest and terrain structure can capture a significant fraction of observed variation in tree species richness in tropical forests on local-scales.