Semantic Segmentation Using Y-Net Architecture for Change Detection on Hyperspectral Imagery

The multiclass change detection problem entails delineating and categorizing land changes using bi-temporal remote-sensing imagery. Technology has recently attracted a lot of interest as a powerful tool for deciphering remote-sensing data and holds enormous potential for precisely identifying changes taking place on the Earth's surface. However, the majority of available algorithms are primarily built for binary change detection and are unable to offer detailed details on the changes' nature. This study offers Y-Net, an end-to-end densely connected network with a Siamese architecture specifically designed for multiclass change detection, as a response. This cutting-edge network uses a dual-stream Dense Net during the encoding stage to extract features related to bi-temporal shifts. It uses an attention fusion strategy to increase the focus on change-related elements during the decoding phase. Furthermore, dice loss is used in multiclass change detection to lessen the effect of sample imbalance. To facilitate change graph reconstruction across multiple levels, the proposed network employs a deep supervision approach for each decoding layer. The performance of Y-Net is evaluated using the challenging multiclass change detection dataset SECOND, and an ablation study demonstrates the potency of both the attention fusion module and dice loss in the multiclass change detection setting.