Realization of Remote Sensing Image Classification Algorithm Based on SOPC

Recently, Convolutional Neural Networks (CNNs) have made significant progress in the field of remote sensing image scene classification. However, commonly used CNN models are mainly deployed on GPU, which is bulky and power-hungry, making them unsuitable for customized requirements of on-board embedded applications. To meet the needs of on-board embedded applications, this study selects System On a Programmable Chip (SoPC), known for their low power consumption and compact size, as the processing platform for accelerating CNNs. Employing the knowledge distillation method, this paper uses the classical CNN scene classification algorithms to create lightweight models. These optimized models are then deployed on resource-constrained ZCU104 SoPC hardware platform. Experimental results demonstrate that the optimized lightweight models achieve image classification tasks with over 95% accuracy on SoPC hardware platform. Furthermore, due to minimal impact from quantization errors, the accuracy drop is less than 1%. Compared to GPU (GTX1650), the throughput on SoPC is increased by 96.43%.