Abstract:Hyperspectral remote sensing technology is swiftly evolving, prioritizing affordability, enhanced portability, seamless integration, sophisticated intelligence, and immediate processing capabilities. The leading model for classifying hyperspectral images, which relies on convolutional neural networks (CNNs), has proven to be highly effective when run on advanced computing platforms. Nonetheless, the high degree of parameterization inherent in CNN models necessitates considerable computational and storage resources, posing challenges to their deployment in processors with limited capacity like drones and satellites. This paper focuses on advancing lightweight models for hyperspectral image classification and introduces EBCNN, a novel binary convolutional neural network. EBCNN is designed to effectively regulate backpropagation gradients and minimize gradient discrepancies to optimize BNN performance. EBCNN incorporates an adaptive gradient scaling module that utilizes a multi-scale pyramid squeeze attention (PSA) mechanism during the training phase, which can adjust training gradients flexibly and efficiently. Additionally, to address suboptimal training issues, EBCNN employs a dynamic curriculum learning strategy underpinned by a confidence-aware loss function, Superloss, enabling progressive binarization and enhancing its classification effectiveness. Extensive experimental evaluations conducted on five esteemed public datasets confirm the effectiveness of EBCNN. These analyses highlight a significant enhancement in the classification accuracy of hyperspectral images, achieved without incurring additional memory or computational overheads during the inference process.

Enhancing Binary Convolutional Neural Networks for Hyperspectral Image Classification