Optimizing Multi-Class Change Segmentation in High-Resolution Satellite Imagery With a Siamese Network for Low-Resource Environments

This work focuses on multi-class change segmentation to quantify alterations between temporal images. Unlike traditional binary change detection, which requires post-processing to identify specific transformations, we introduce a Siamese architecture called SiamSegCD. This model is designed for direct segmentation of land-use and land-cover changes, particularly suited for low-resource environments. SiamSegCD utilizes deep learning models like UNet and DeepLabv3+ with a lightweight ResNet-50 backbone and stand-alone self-attention to yield a fully self-attentional model, implemented in a Siamese twin framework. The model’s performance is assessed using the HRSCD dataset, revealing better results with the DeepLabv3+ encoderdecoder compared to the UNet-based version, which demonstrated a 5.98% improvement in mean Intersection over Union (mIoU).