A Resource-Efficient FPGA Architecture for Real-Time Classification of Optical Weather Images

Real-time analysis of optical weather images is important for resource-constrained embedded vision systems. However, the computational cost of deep learning models limits their deployment on resource-constrained platforms. This paper presents a resource-efficient FPGA architecture for six-class weather image classification. A lightweight convolutional neural network, WeatherNet, employs depthwise separable convolution to reduce model complexity while maintaining recognition performance. A dataflow-aware accelerator improves memory efficiency through on-chip feature reuse, optimized data movement, and pipelined computation. The proposed design is implemented on a Xilinx ZYNQ-7020 FPGA using hardware–software co-design. The evaluation uses ground-level RGB images labeled cloudy, dew, lightning, rainy, sandstorm, and sunrise. The floating-point WeatherNet achieves 94.08% accuracy, while the deployed fixed-point FPGA implementation achieves 93.75% on the same test set. WeatherNet contains 12,647 parameters and requires 10.426 million MACs per image for the convolutional and fully connected layers. The FPGA implementation achieves 200 FPS at 100 MHz. Based on an estimated on-chip power consumption of 2.201 W, the corresponding energy efficiency is 90.87 FPS/W. These results demonstrate the feasibility of the proposed architecture for weather image classification on the evaluated FPGA platform.