Harnessing temporal and spectral dimensionality to map and identify species of individual trees in diverse tropical forests

Abstract To understand how tropical rainforests will adapt to climate change and the extent to which their diversity imparts resilience, precise, taxonomically informed monitoring of individual trees is required. However, the density, diversity and complexity of tropical rainforests present considerable challenges to remote mapping and traditional field-based approaches are limited in scale. This study introduces a new approach for mapping tree species linking a multi-temporal implementation of the convolutional neural network method, detectree2 , to segment tree-crowns from aerial photographs to machine learning classifiers to identify species from hyperspectral data (416 - 2500 nm). We build upon previous work in two ways. Firstly, we aimed to improve the accuracy of crown delineations by surveying the same patch of forest with UAV-RGB ten times over six months and fusing multi-date information on the location and shape of individual trees. Secondly, we extended the scope of species identification to include far more species than has been previously attempted (169 compared to 20 previously). We trained and tested our algorithms on subsets of a database of 3500 ground truth, labelled tree crown polygons representing 239 species in French Guiana that we had delineated by hand and field verified. We assessed how well our segmentation approach could locate and delineate individual tree crowns and how well our classification approach predicted the species of those crowns. We extracted information on waveband importance for distinguishing species from our classification model. Based on an existing phylogeny of the trees in our dataset, we tested for phylogenetic signal across the hyperspectral bands and probed how species were being classified by comparing the phylogenetic signal to the importance of bands for separating species. The accuracy of delineations increased gradually as additional dates of tree crown maps were stacked and combined. Stacking increased the F1-score from 0.69 (a single date) to 0.78 (all dates). The overall (weighted) F1-score for species classification was 0.75. A total of 65 species were predicted from the hyperspectral data with F1-score > 0.7. The performance for classifying a species increased with the number of crowns in the database available for that species: 8 training crowns were needed to achieve an expected F1-score = 0.7 for crown level classification. With this new approach, we assessed that 70% of tree crown area at landscape-scale was accurately mapped. The most important wavebands for discriminating species were narrowly clumped on the NIR side of the red edge region (748 - 775 nm). While most wavebands showed some phylogenetic signal, waveband importance for species classification was negatively correlated with phylogenetic signal. Our integrated approach makes a significant contribution to the ongoing development of efficient and accurate methodologies for mapping canopy tree species in tropical forests, providing a framework for mapping trees in diverse tropical forests that is far more comprehensive than its predecessors.