Performance Improvement of Encoder/Decoder-Based CNN Architectures for Change Detection from Very High-Resolution Satellite Imagery

Abstract

Over the past few years, Convolutional Neural Networks (CNN) have grown in popularity with the remote sensing community due to their relatively easy training process, excellent generalization capacity and state-of-the-art performance. CNN models following an encoder-decoder architecture to perform end-to-end semantic segmentation have been applied to Change Detection (CD) applications with remarkable results. In this paper, we aim to further improve the performance of such models. First, we experiment with the introduction of additional boundary information into an encoder-decoder architecture that performs semantic segmentation for CD. We use the Dense Extreme Inception Network (DexiNeD) to produce the semantically informed edges. Second, we propose a training process that implicitly teaches the model to become more robust to misregistration errors. We evaluate our proposed approaches on a CD dataset, which consists of very high resolution RGB satellite image pairs, using two encoder-decoder models, UNet and UNet++, as our backbone architecture. The evaluation results suggest that both enhancements improve the performance of the CD network, with the average improvement on precision, recall, F1score and IoU ranging between 1% and 2% when incorporating boundary features into our architectures, and up to 2.5% when modeling the misregistration errors in the training process.