Tree Segmentation from Low-Resolution Digital Orthophotos using a Hybrid Deep Learning Model

This study presents a cost-effective tree crown segmentation framework using a hybrid deep learning model that combines a ResNet-34 encoder with a U-Net decoder.Our approach operates on low-resolution RGB Digital Orthophotos (DOPs) collected from urban and peri-urban areas in Bochum, Germany, simulating real-world data constraints.We processed 450 orthophoto-mask pairs through a comprehensive preprocessing pipeline including resizing (from 2000020000 to 256256), augmentation, and noise simulation.The model was trained using 10-fold cross-validation, achieving a Dice coefficient of 0.8678, Intersection over Union (IoU) of 0.7754, precision of 0.8410, and recall of 0.9103.These results demonstrate that even with downsampled imagery, reliable segmentation of tree crowns is feasible, making our approach suitable for low-cost forest inventory and precision agroforestry applications.Unlike previous studies relying on high-resolution LiDAR, this work is among the first to show robust tree crown segmentation using low-resolution orthophotos, making it accessible for widespread use in resource-constrained settings.