Abstract:Semi-supervised classification of remote sensing (RS) hyperspectral image (HSI) aims at exploiting both labeled and unlabeled samples for accurate land cover recognition. However, imbalanced data distribution and different classification difficulties negatively affect classification performance. Focused on this, a novel dual-stream class-adaptive network (DSCA-Net) is proposed for semi-supervised HSI classification, in this article. First, a superpixel-guided label propagation (SGLP) module is introduced to alleviate the negative effect of imbalanced data distribution. Specifically, approximate estimation of labels for unlabeled samples is achieved via superpixel-wise similarity measure and label propagation, so that equal sampling is applied to each class. Then, a consistency regularization-based dual-stream network is constructed, which shares the same encoder for feature representation of either labeled or unlabeled samples. Based on this, two distinct classifiers are designed to force similar predictions that can be achieved for various perturbed versions of the same unlabeled sample, thereby allowing unlabeled samples to train the model in a supervised manner. Finally, since different classes always have various degrees of learning difficulty, equal treatment may lead to overfitting of "easy" classes and biased prediction of "hard" classes. Unlike the traditional selection of unlabeled samples with a fixed threshold, dynamic class-adaptive thresholds are calculated according to the learning status of the model. In this manner, a higher threshold is assigned to "easy" classes to reduce sample redundancy, and a lower threshold is set for "hard" classes to select more samples. Experiment results demonstrate the effectiveness and superiority of the proposed method. Codes are available at https://github.com/luting-hnu/DSCA-Net.

IDN: Inner-Class Dense Neighbours for Semi-Supervised Learning-Based Remote Sensing Scene Classification

A lightweight and stochastic depth residual attention network for remote sensing scene classification

Superpixel-Based Long-Range Dependent Network for High-Resolution Remote-Sensing Image Classification

Remote Sensing Scene Classification by Gated Bidirectional Network

Semi-Supervised Subcategory Centroid Alignment-Based Scene Classification for High-Resolution Remote Sensing Images

Semisupervised Scene Classification for Remote Sensing Images: A Method Based on Convolutional Neural Networks and Ensemble Learning

Semi-Supervised Scene Classification for Remote Sensing Images Based on CNN and Ensemble Learning.

Remote Sensing Image Scene Classification with Self-Supervised Paradigm under Limited Labeled Samples

Index Your Position: A Novel Self-Supervised Learning Method for Remote Sensing Images Semantic Segmentation

Scene Classification of High Resolution Remote Sensing Images Via Self-Paced Deep Learning.

Semi-supervised remote sensing image scene classification with prototype-based consistency

Deep Updated Subspace Networks for Few-Shot Remote Sensing Scene Classification

Hierarchical Self-Attention Embedded Neural Network With Dense Connection for Remote-Sensing Image Semantic Segmentation

Duplex-Hierarchy Representation Learning for Remote Sensing Image Classification

Collaborative Learning-based Network for Weakly Supervised Remote Sensing Object Detection

A Cross-Layer Nonlocal Network for Remote Sensing Scene Classification

Class-level Prototype Guided Multi-Scale Feature Learning for Remote Sensing Scene Classification with Limited Labels

Multigranularity Decoupling Network With Pseudolabel Selection for Remote Sensing Image Scene Classification

Remote sensing scene classification based on contextual attention mechanism of lie group space

DOCNet: Dual-Domain Optimized Class-Aware Network for Remote Sensing Image Segmentation

Dual-Stream Class-Adaptive Network for Semi-Supervised Hyperspectral Image Classification