Intelligent Tree Canopy Segmentation From UAV Imagery to Support Urban and Rural Garden Design

Accurate tree canopy mapping is fundamental for urban planning, ecological monitoring, and garden design. This project presents an end‑to‑end deep learning pipeline that automates tree canopy segmentation from UAV aerial imagery and makes it accessible through a web application. Two transformer‑enhanced segmentation models were developed and evaluated: AerialFormer (Swin Transformer Tiny encoder with CNN stem and multi‑scale decoder) and UNetFormer (ResNet50 encoder with memory‑efficient window attention in decoder blocks). Both were trained on the high‑resolution restor/tcd dataset using a combined BCE + Dice loss, AdamW optimization, and cosine annealing. For the front‑end interface, the best‑performing model (AerialFormer) is deployed within a Flask‑based web application. Users can upload UAV images and instantly receive a complete segmentation panel including predicted mask, probability heatmap, canopy‑overlay visualization, and coverage percentage. The system delivers actionable spatial insights for tree distribution assessment, land‑use decisions, and sustainable landscape design in both urban and rural environments

Keywords: Tree canopy segmentation; UAV imagery; deep learning; Swin Transformer; AerialFormer; UNetFormer; Flask web application; garden design; remote sensing.