UTRNet: An Unsupervised Time-Distance-Guided Convolutional Recurrent Network for Change Detection in Irregularly Collected Images

Change detection in time series is among the most critical problems in earth monitoring and attracts extensive attention in the remote sensing community. The task is, however, nontrivial because available images are irregularly collected due to interference by clouds and shadows. Traditional recurrent neural networks neglect such information and thus degrade the possibility of distinguishing pseudochanges (caused by intra-annual and inter-annual dynamics) and real changes. To this end, we proposed an unsupervised time-distance-guided convolutional recurrent network (UTRNet) for change detection in irregularly collected images. UTRNet is distinctive because the influence of pseudochanges can be suppressed by adopting a novel time-distance-guided long short-term memory (TLSTM) unit, in which input and forget gates are modified to adapt to irregular time distances. To the best of our knowledge, this is the first time that the influence of pseudochanges can be suppressed using irregular time distances. Moreover, to make UTRNet more applicable, a weighted prechange detection model is proposed to extract the most reliable training samples automatically. In the training process, unlike existing approaches that only care about changed and unchanged sample imbalance, our UTRNet also pays attention to imbalance between hard and easy samples and proposes a new focal weighted cross-entropy loss, which helps to make the training process focus on hard changed samples. The proposed UTRNet is validated on Landsat 8 time series data collected over nine typical scenes in the 2013-2021 period. Qualitative and quantitative comparisons with several state-of-the-art methods suggest the superior performance of UTRNet. Our dataset and codes are available at https://github.com/thebinyang/UTRNet.