The automatic observation and recognition of ground-based clouds have guiding significance for analyzing atmospheric motion trends and weather forecasting. To solve the problem of low accuracy of ground-based cloud image classification algorithms and difficulties in deploying them on embedded terminals, a ground-based cloud image classification network model GBcNet based on residual network structure and a hardware implementation architecture based on ZYNQ are proposed. The PS end is used to load the weight parameters and cloud data of the model, and the PL end implements DDR3 read-write control and GBcNet hardware acceleration. For the GBcNet network, accelerated IP cores corresponding to each module were designed, including sliding window, convolutional layer, pooling layer, batch normalization layer, and fully connected layer. Experiments were conducted on the CCSN dataset, and the results showed that the GBcNet model achieved an accuracy of 96.02% on a PC platform. After hardware acceleration, the accuracy remained at 94.5%. Compared with the recognition rate of the PC model, the accuracy loss for each cloud class did not exceed 3%, and the overall accuracy loss was less than 1.5%. The maximum resource usage on the FPGA did not exceed 48%, and the inference time for a single ground-based cloud image was 0.13 seconds. Compared to existing ground-based cloud recognition methods, this approach demonstrates higher accuracy and shorter inference time. The proposed recognition model and acceleration method provide a reference solution for the development of multi-node, portable ground-based cloud observation equipment.