Forest Change Detection Using Dynamic Squeeze and Excitation Attention with Pyramid Scene Parsing Network from Satellite Imagery

The accurate detection of forest changes is significant for monitoring forest sources and ensuring accuracy of land data. This research utilizes Deep Learning (DL) methods to perform forest change detection through analyzing Sentinel-2 dataset from mountain regions. The Dynamic Squeeze and Excitation attention with Pyramid Scene Parsing Network (DSE-PSPNet) is proposed in this research for forest change detection from satellite images. The DSC-PSPNet enhances both feature representation and segmentation accuracy. The model uses DSE to control spatial and channel-wise attention dynamically therefore it prioritizes important forest alterations with reduced background effects. PSPN et implements a multi-scale feature fusion method which effectively monitors both smaller scale and broader scale variations in forest regions. Preprocessing techniques of augmentation and normalization improve forest change detection operations by raising accuracy together with model robustness. The application of flipping and rotation alongside brightness which helps the model avoid overfitting and improve its ability to handle previously unseen data points. The normalization technique speeds up the training process since it maintains equal distributions of gradients. The DSE-PSPNet attained 99.73% precision, 99.65% fl-score and 99.69% sensitivity for Sentinel-2 dataset compared to U-Net++ model.