SARCDNet-an enhanced deep learning network for change detection from bi-temporal SAR images

Abstract

Abstract Change detection analysis is a crucial process in microwave remote sensing enabling the identification of changes on the Earth’s surface due to natural and human activities. Bi-temporal Synthetic Aperture Radar (SAR) imaging is particularly effective for flood detection, environmental monitoring, disaster response and urban planning due to its all-weather, day-and-night capabilities. Conventional optical imaging methods face challenges such as cloud cover and daylight dependency making SAR a preferred alternative. The availability of SAR data from satellite resources such as RADARSAT, Sentinel-1, Envisat has helped in advancing change detection analysis. This paper introduces SARCDNet (SAR Change Detection Network) model which is an enhanced deep learning network for change detection from bi-temporal SAR images. Adaptive fusion block is designed, which incorporates the adaptive global filtering operation that extracts features in frequency domain and channel attention mechanism to enhance the relevance of extracted features. SARCDNet is computationally efficient design that mitigates the effects of speckle noise and improves the prediction accuracy. SARCDNet takes advantage of spatial and frequency domain methods and has demonstrated improved performance across public datasets namely Yellow river and Farmland which has higher incidence of speckle noise. The proposed SARCDNet model has demonstrated considerable performance improvement of 4.26% F1 score, 0.841% PCC, 4.92% $$\kappa$$ and 4.67% MCC in flood change detection in Chao Lake dataset.