Semantic Segmentation of Forest Regions from Satellite Imagery using Transformers

Abstract

Forest conservation is essential to combat global warming, climate change, and threats to global biodiversity. This necessitates innovative technological solutions for effective forest monitoring and management. This study focuses on forest segmentation of 15 districts of the Khyber Pakhtunkhwa (KPK) province in Pakistan using multi-spectral Landsat-8. The dataset comprises images with 11 bands, including RGB. This work presents an advanced approach to forest segmentation utilizing transformer-based architectures. Transformers, known for their powerful feature extraction and attention mechanisms, are leveraged to improve segmentation accuracy. They excel in handling long-range dependencies and operate in parallel mode, which enhances global context modeling by processing entire sequences simultaneously. In this study we investigated two transformer based models, SegFormer and SegForest. SegForest model is inspired by the SegFormer architecture, featuring an encoder-decoder structure. The encoder is akin to that of SegFormer, enabling multiscale feature extraction without requiring positional encoding. The decoder employs an advanced methodology, integrating multi-feature fusion and multi-scale multi-decoder modules for enhanced performance. The key innovation of this work is the introduction of reduction ratios in the encoder’s CNN layers, which optimize performance by selectively downsampling feature maps to balance detail retention and capture finer details. This novel approach enhances the model’s capability to process multiscale features more effectively. We trained and tested our models on subsets of this dataset: RGB bands and 11 selected bands to evaluate performance across different spectral inputs.Our transformer based models with pretrained weights demonstrated superior segmentation accuracy compared to existing methods, achieving a notable increase in accuracy and F1-score and contributed to a 4 percent improvement in accuracy on our dataset, showcasing the effectiveness of our approach in handling diverse spectral inputs.