Semisupervised Change Detection With Feature-Prediction Alignment

Abstract

Change detection (CD) has received raising attention for its broad application value. However, traditional fully-supervised CD methods have a huge demand for the pixel-level annotations, which are laborious and even impossible in some few-shot scenarios. Recently, several semi-supervised CD (SSCD) methods have been proposed to utilize numerous unlabeled RSI pairs, which can largely reduce the annotation dependence. These methods are mainly based on: 1) the adversarial learning, whose optimization direction is difficult to control as a black-box method; or 2) the feature-consistency learning, which has no explicit physical meaning. To deal with these difficulties, we propose a novel progressive SSCD framework in this paper, termed feature-prediction alignment (FPA). FPA can efficiently utilize unlabeled RSI pairs for training by two alignment strategies. First, a class-aware feature alignment (FA) strategy is designed to align the area-level change/no-change feature extracted from different unlabeled RSI pairs (i.e., across regions) with the awareness of their locations, in order to reduce the feature difference within the same classes. Secondly, a pixel-wise prediction alignment (PA) is devised to align the pixel-level change prediction of strongly-augmented unlabeled RSI pairs to the pseudo-labels calculated from the corresponding weakly-augmented counterparts, in order to reduce the prediction uncertainty of various RSI transformations with physical meaning. Experiments are carried out on four widely-used CD benchmarks, including LEVIR-CD, WHU-CD, CDD, and GZ-CD, and our FPA achieves the state-of-the-art performance. The experimental results demonstrate the superiority of our method in both effectiveness and generalization. Our code is available at https://github.com/zxt9/FPA-SSCD.