Self-organising explainable multi-view representation learning for remote sensing scene classification

Remote sensing scene classification is widely considered to be a challenging task due to high intraclass variability and interclass similarity in remotely sensed imagery. While existing deep neural networks achieve promising performance, they often lack transparency and generalisation capability. To enhance interpretability without sacrificing accuracy, a novel self-organising transparent multi-view representation learning framework based on evolving fuzzy neural encoders for remote sensing scene classification is introduced in this paper. The framework leverages multiple pre-trained convolutional neural network backbones with different architectures to extract image embeddings from multiple views. The multi-view image embeddings are projected into a lower-dimensional feature space using multilayer evolving fuzzy neural networks trained in a supervised or self-supervised fashion as encoders and subsequently fused for scene classification. Extensive experiments on six benchmark datasets (Optimal-31, WHU-RS, UCMerced, AID, RSI-CB256, and PatternNet) demonstrate the framework’s superior performance, achieving average accuracies of 99.81 %, 98.83 %, 97.86 %, 98.37 %, 99.84 %, and 98.83 %, respectively, without fine-tuning to the specific context. Ablation studies confirm the complementary contributions of the multi-view, supervised, and self-supervised components in the proposed framework. The proposed framework provides an effective solution for remote sensing scene classification, achieving high accuracy with enhanced transparency and interpretability.