IRCNN: An Irregular-Time-Distanced Recurrent Convolutional Neural Network for Change Detection in Satellite Time Series

Deep learning (DL)-based methods incorporating convolutional neural networks (CNNs) and recurrent neural networks (RNNs) have been successfully applied to change detection in satellite time series. However, traditional RNNs assume identical time interval between image sequences, which hardly meets the real case in satellite time series because of clouds and shadows. In this letter, a novel irregular-time-distanced recurrent CNN (IRCNN) is proposed. IRCNN consists of three sub-networks: a multi-branch Siamese CNN, irregular-time-distanced long short-term memory (ILSTM), and fully connected (FC) layers. Superior to the existing methods, IRCNN can account for temporal dependency among time series with irregular time distances. It is end-to-end trainable with samples generated using an automatic annotation generation method, which is proposed based on the prior knowledge from the continuous change detection and classification (CCDC) approach. IRCNN was tested over five study areas using Landsat time series collected between 2013 and 2020. Experiments demonstrate the effectiveness and stability of the proposed network with better performance, compared to the state-of-the-art approaches in terms of both qualitative and quantitative aspects. Our IRCNN Pytorch code and data are available at https://github.com/thebinyang/IRCNN.

Automated summary

This letter proposes a novel method called IRCNN for change detection in satellite time series. IRCNN is based on deep learning and incorporates convolutional neural networks and irregular-time-distanced recurrent neural networks. Experimental results show that IRCNN outperforms other methods in both qualitative and quantitative aspects.