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

摘要

Over the past few years, Convolutional Neural Networks (CNN) have grown in popularitywith the remote sensing community due to their relatively easy training process, excellentgeneralization capacity and state-of-the-art performance. CNN models following an encoderdecoderarchitecture to perform end-to-end semantic segmentation have been applied toChange Detection (CD) applications with remarkable results. In this paper, we aim to furtherimprove the performance of such models. First, we experiment with the introduction ofadditional boundary information into an encoder-decoder architecture that performs semanticsegmentation for CD. We use the Dense Extreme Inception Network (DexiNeD) to producethe semantically informed edges. Second, we propose a training process that implicitlyteaches the model to become more robust to misregistration errors. We evaluate our proposedapproaches on a CD dataset, which consists of very high resolution RGB satelliteimage pairs, using two encoder-decoder models, UNet and UNetþþ++, as our backbone architecture.The evaluation results suggest that both enhancements improve the performanceof the CD network, with the average improvement on precision, recall, F1score and IoUranging 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.