MMPC-Net: Multigranularity and Multiscale Progressive Contrastive Learning Neural Network for Remote Sensing Image Scene Classification
Shaofan Li, Mingjun Dai, Bingchun Li
Abstract
With the development of convolutional neural network (CNN), significant progress has been achieved in remote sensing image scene classification (RSISC). However, wide spatial range changes, complex scenes, as well as the high similarity between various classes and the significant difference in the same class, make it difficult to classify remote sensing images scenes. In this work, targetted at using finite remote sensing images to learn sufficient distinguishing features in a contrastive manner, we propose a novel multi-granularity and multi-scale progressive contrastive learning neural network (MMPC-Net). More specifically, we construct an end-to-end CNN model to mine discriminative features from multi-scale and multi-granularity representations. Afterwards, the discriminative knowledge between different features is summarized by introducing a progressive contrastive learning module, which can learn meaningful feature linking subtle changes in positive and negative pairs from massive samples. Experimental results over three widely-used benchmark datasets demonstrate that our methods can achieve comparative performance.