Continental-scale hyperspectral tree species classification in the United States National Ecological Observatory Network

Abstract

Advances in remote sensing imagery and computer vision applications unlock the potential for developing algorithms to classify individual trees from remote sensing at unprecedented scales. However, most approaches to date focus on site-specific applications and a small number of taxonomic groups. This limitation makes it hard to evaluate whether these approaches generalize well across broader geographic areas and ecosystems. Leveraging field surveys and hyperspectral remote sensing data from the National Ecological Observatory Network (NEON), we developed a continental extent model for tree species classification that can be applied to the entire network including a wide range of US terrestrial ecosystems. We compared the performance of the generalized approach to models trained at each individual site, evaluating advantages and challenges posed by training species classifiers at the US scale. We evaluated the effect of geography, environmental, and ecological conditions on the accuracy and precision of species predictions. On average, the general model resulted in good overall classification accuracy (micro-F1 score), with better accuracy than site-specific classifiers (average individual tree level accuracy of 0.77 for the general model and 0.72 for site-specific models). Aggregating species to the genus-level increased accuracy to 0.83. Regions with more species exhibited lower classification accuracy. Trees were more likely to be confused with congeneric and co-occurring species and confusion was highest for trees with structural damage and in complex closed-canopy forests. The model produced accurate estimates of uncertainty, correctly identifying trees where confusion was likely. Using only data from NEON this single integrated classifier can make predictions for 20% of all tree species found in forest ecosystems across the US, suggesting the potential for broad scale general models for species classification from hyperspectral imaging.

Capturing long-tailed individual tree diversity using an airborne multi-temporal hierarchical model

“…These 14 species represent almost 90% of all stems present at the site and include rare species representing less than 1% of the individual trees on the landscape. Compared to the single-site OSBS model created from NEON woody vegetation structure data alone (Marconi et al 2022, evaluation accuracy in Table S2), we doubled the species number predicted by incorporating auxiliary, non-NEON data. As a result, 25% of crowns predicted at OSBS in this study were of species not included in previous efforts.…”

“…An important methodological change from previous tree species classification workflows is that we generated crown predictions for evaluation using an RGB crown detection model independent from the field labeled trees.This approach reflects the process during prediction in which field stem points are not available. The majority of previous papers use fixed size boxes around field stem points or hand-drawn crown polygons on the imagery to generate both training and evaluation data (Maschler et al 2018, Scholl et al 2020, Onishi et al 2022, Marconi et al 2022). This crown delineation approach biases results towards higher accuracy as field points are most often collected on the largest and easiest to differentiate individuals.…”

“…The DeepForest model is a RGB tree detection tool that has been validated using NEON data (Weinstein et al 2020a(Weinstein et al , 2020b(Weinstein et al , 2021. While previous work used a fixed number of pixels closest to the field stem location (Scholl et al 2020, Marconi et al 2022, model-based bounding boxes linked training and evaluation processes more directly to predictions where no stem locations were available. In the event that multiple predicted tree crown boxes overlapped the field point, we chose the box centroid nearest the field stem.…”

“…Following the trend away from hand-crafted features towards deep learning neural networks, numerous publications have applied computer vision approaches to airborne tree species prediction (Mäyrä et al 2021, Abbas et al 2021, Chen et al 2022, Onishi et al 2022, Veras et al 2022). These applications demonstrated that models often achieve between 70% to 90% accuracy for fewer than 10 co-occurring, well-sampled, classes (Marconi et al 2022). A remaining species classification challenge is broadening classifications beyond the most common and differentiable species and applying these models to the scale of entire forested landscapes.…”

Evaluating individual tree species classification performance across diverse environments

“…Remote sensing data are increasingly being used to circumvent the limitations of traditional field-based forest inventories. While species classifications have successfully used light detection and ranging (LiDAR) and multispectral images, complex cases often rely on airborne imaging spectroscopy to achieve sufficient accuracies [8][9][10][11][12][13][14]. Despite the potential of these data to operationally map plant species across large geographic regions, classification studies using imaging spectroscopy are often limited to small geographic regions [10,12,13], and thus the effects of intra-specific spectral variability, driven by variation in canopy traits, on classification accuracies remains underexplored [10].…”