Weakly Supervised Deep Soft Clustering for Flood Identification in SAR Images
Fei Ma, Deliang Xiang, Kun Yang, Qiang Yin, Fan Zhang
Abstract
As flood occurs unpredictably, there is not enough time to label the data in practice. The use of clustering inside flood detection deep networks can reduce their demand for labeled data. However, existing clustering algorithms aim at assigning a unique cluster for each pixel. This leads to the fact that clustering process is non-differentiable to the inputs, hindering their incorporation into deep networks. In this study, we introduce a new assignment strategy for single-polarization SAR images to make the clustering differentiable, named “soft association.” Here, each pixel is assigned to various clusters with different probabilities. The greater the probability value, the more likely the pixel will be finally assigned to the cluster. Based on this, an end-to-end trainable semi-supervised clustering network for SAR flood detection is established. Compared with the existing state-of-the-art semi-supervised methods, it can achieve similar performance with fewer labeled samples.