Multi-model Approach for Tree Detection and Classification in Wallonia Region (Belgium)

Abstract. Forests play a pivotal role in global ecosystems by sequestering carbon, preserving biodiversity, and providing valuable resources for both humans and wildlife. Monitoring and management of these forests require accurate, up-to-date information on individual trees and species composition—challenges that can be addressed with advanced remote sensing and deep learning. This paper presents a multi-season, multi-year approach to automatic tree detection and species classification in heterogeneous forests. Using over 5,000 high-resolution (0.25 m) RGB orthophoto tiles from the Wallonia region (spanning 2018–2023), we annotated more than 100,000 individual trees representing 14 classes of deciduous and coniferous species. A Faster R-CNN model trained for tree detection achieved a F1 score of 0.828 and a mAP@50 of 0.827, effectively locating tree crowns under varying illumination and phenological conditions. Meanwhile, a convolutional neural network (CNN) for species classification attained an overall accuracy of 0.937, accurately distinguishing most species and age classes. Despite strong performance, limitations persist, particularly in identifying small saplings and visually similar species (e.g., oak vs. beech). These findings highlight the potential of multi-temporal aerial imagery and deep learning to enhance forest inventories, reduce field survey costs, and inform targeted management.