SpectroChangeNet Model for Change Detection in Synthetic Aperture Radar

Change detection using Synthetic Aperture Radar (SAR) has gained increased attention from the research community.The majority of existing change detection models have focused primarily on detection performance, with limited attention given to computational time and memory usage.To address this concern and achieve better change detection performance, a new deep learning model named SpectroChangeNet is introduced in this paper.The SpectroChangeNet model utilizes feature vectors from both the frequency and spatial domains to improve change detection results.Specifically, in the frequency domain, discrete coefficients are extracted from the SAR images using the Discrete Cosine Transform (DCT).In the spatial domain, the Multi-Region Convolution (MRC) module extracts deep feature vectors from the input images.The combined discrete coefficients and deep feature vectors are then used to classify uncertain pixels as changed or unchanged using the softmax function.In addition, a hybrid loss function is integrated with the MRC module to reduce the model's computational time and memory usage.The focal loss efficiently down-weights well-classified samples to concentrate more on hard or mis-classified uncertain samples.Further, the Mean Absolute Error (MAE) ensures that the gradient updates remain balanced and smooth.This model achieved 86.98%, 96.20%, and 94.28% of Kappa Coefficient (KC), and 96.36%, 98.92%, and 98.52% of Percentage of Correct Classification (PCC) on the Yellow River, Sulzberger, and Ottawa datasets.The achieved change detection results are superior compared to the existing models.