A Swin Transformer, YOLO, and Weighted Boxes Fusion-Based Approach for Tree Detection in Satellite Images

Forests are terrestrial ecosystems that provide ecological balance and environmental benefits. Due to uncontrolled logging, increasing population, and commercial use, forest reserves are rapidly decreasing worldwide. This situation carries risks that will lead to climate crisis and biodiversity loss. For this reason, trees need to be protected and kept under control. Since human-powered control efforts are insufficient, the use of satellite images and artificial intelligence-based systems has come to the fore to prevent forest destruction. Nowadays, there are studies on tree detection applications from satellite images using different deep learning models. This article aims to combine Swin transformers and YOLO models with a post-processing technique known as weighted boxes fusion to enhance object detection capabilities. The object detection approaches of the two models are different. When the wide variation detection ability of the Swin transformers method is combined with the low variation but sharp detection ability of the YOLO model, the results combine the advantages of both models in tree detection. In this way, the proposed method detected trees with %3.7 higher precision than Swin transformers and %13.1 higher precision than YOLO.