DNN-Based PolSAR Image Classification on Noisy Labels

Deep neural networks (DNNs) appear to be a solution for the classification of polarimetric synthetic aperture radar (PolSAR) data in that they outperform classical supervised classifiers under condition sufficient training samples. The design classifier is challenging because DNNs can easily overfit due limited remote sensing samples and unavoidable noisy labels. In this article, softmax loss strategy with antinoise capability, namely, probability-aware sample grading (PASGS), developed overcome limitation. Combined proposed strategy, two DNN-based are implemented perform PolSAR image demonstrate its effectiveness. framework, difference distribution implicitly reflects probability clean, clean labels distinguished from according method statistics. Then, employed reweight corresponding each during process locate prevent participation calculation network. As number iterations increases, statistics will constantly adjusted without supervision, eventually identified train Experiments on three datasets methods also superior state-of-the-art methods.