Cloud image classification aims to judge whether there is cloud in remote sensing images, which plays an important role in onboard cloud detection tasks. Cloud image classification methods mainly include threshold-based, machine learning-based, and deep learning-based approaches. While deep learning-based cloud image classification methods can provide better performance for accurate cloud detection under complex scenarios, they suffer from long processing time and high onboard resource consumption. Therefore, scholars usually select threshold-based methods combined with machine learning methods based on multiple texture features to realize high-efficiency on-orbit cloud discrimination. However, in order to deal with the interference of high-brightness ground objects such as ice and snow, there are problems such as complex calculation and large amount of intermediate result data when constructing features in cloud image classification method. These problems seriously affect processing efficiency. Therefore, a cloud image classification method based on tree analysis process using spectral-improved texture features is proposed in this article, and it is accelerated by field-programmable gate array (FPGA). The proposed method quickly detects the dark ground in the image through tree branch setting constructed by two levels of adaptive threshold, which simplifies the method calculation. Furthermore, a novel diagonal feature of gray level co-occurrence matrix (GLCM) is designed to reduce the calculation complexity while maintaining the ability to distinguish between cloud and high-brightness ground. In addition, the time-consuming pixel statistics process is accelerated by FPGA. Experimental results demonstrate that the proposed method has a reduction of more than 99% in the number of operations and more than 95% in the volume of intermediate data than MobileNetV3 with high classification accuracy, while achieving a similar cloud false alarm rate. In addition, the acceleration architecture of the proposed method can achieve processing for an image with a resolution of 7 300×6 908 pixels in 1.20 s, which greatly improves the efficiency.