An Efficient Change Detection for Large SAR Images Based on Modified U-Net Framework

Abstract

Large SAR images usually contain a variety of land-cover types and accordingly complicated change types, which cause great difficulty for accurate change detection. The U-Net is a special fully convolutional neural network that not only can capture multiple features in the image context but also enables precise pixel-by-pixel image classification. Therefore, we explore the U-Net to describe accurately the differences between bi-temporal SAR images for high-precision change detection. However, large scene SAR images often have significantly different statistical distributions for various change types, which prevents the U-Net from working properly. We modified the U-Net by introducing the batch normalization (BN) operation at the input of every neuron to regularize the statistical distributions of its input data for avoiding the risk of gradient disappearance or dispersion during the network training. In addition, the ELU (Exponential Linear Unit) activation function replaces the ReLU (Rectified Linear Unit) function to improve further the gradients backpropagation. Then we selected bi-temporal Sentinel-1SAR data covering Jiangsu Province, China, to discuss quantitatively and qualitatively the detection performance and model complexity of the modified network with different numbers of convolutional kernels.