A Multi-Angle Encoding Spiking Convolutional Neural Network for Remote Sensing Classification

Spiking Convolutional Neural Networks (SCNNs), known as the third generation of neural networks, are favored for their low energy consumption and biological plausibility, making them ideal for energy-limited applications like satellite remote sensing image classification. Traditional Convolutional Neural Networks (CNNs) consume significant energy, prompting a shift towards more efficient architectures like binary and adder neural networks. However, SCNNs have been overlooked due to their binary information transmission, which typically results in lower accuracy. This paper introduces the Multi-Angle Encoding Spiking Convolutional Neural Network (MASCNN), featuring a Multi-Angle Encoding Layer and a Deep Feature Extraction Module to enhance input information and improve classification accuracy. A new Multi-Angle Loss Function is also proposed to enrich learning. Testing on various datasets shows that MASCNN outperforms other low-energy networks in accuracy while maintaining minimal energy use.