Noise-Tolerant Deep Neighborhood Embedding for Remotely Sensed Images With Label Noise

Recently, many deep learning-based methods have been developed for solving remote sensing (RS) scene classification or retrieval tasks. Most of the adopted loss functions training these models require accurate annotations. However, presence noise in such annotations (also known as label noise) cannot be avoided large-scale RS benchmark archives, resulting from geo-location/registration errors, land-cover changes, and diverse knowledge background annotators. To overcome influence noisy labels on learning process models, we propose a new function called noise-tolerant neighborhood embedding which can accurately encode semantic relationships among scenes. Specifically, target at maximizing leave-one-out K-NN score uncovering inherent structure images feature space. Moreover, down-weight contribution potential by their localized pruning with low scores. Based our newly proposed function, classwise features more robustly discriminated. Our experiments, conducted two datasets, validate effectiveness approach three different interpretation tasks, including classification, clustering, retrieval. The codes this article will publicly available https://github.com/jiankang1991.