SkySense-CCS: a ground-based cloud segmentation method using connected domain algorithm

Due to its particularity, the ground-based cloud image segmentation algorithm is often applied in ordinary processor environments such as single-chip microcomputers with weak computing power, and requires high interpretability to adapt to the complexity and variability of cloud image scenarios. Factors such as irregular cloud shapes, blurred edges, significant illumination variations, and low contrast between thin clouds and the sky have led to obvious limitations in traditional threshold methods, manual feature-dependent machine learning approaches, and data-driven deep learning methods regarding segmentation accuracy, adaptability, and edge preservation. To address this, we propose SkySense-CCS, a lightweight intelligent segmentation algorithm for complex cloud scenes. Based on connected component segmentation theory, the algorithm establishes a three-stage processing workflow: thresholding, labeling, and filtering. First, it employs physical priors from blue/red channel ratios for adaptive thresholding to achieve illumination-robust cloud-sky initial separation. Second, it efficiently extracts morphological and positional attributes of cloud blocks through connected component labeling. Finally, a shape-position joint filter is designed to effectively eliminate noise interference from birds, power lines, and solar flares using multi-dimensional criteria like area, eccentricity, and zenith distance. Experimental results demonstrate that SkySense-CCS achieves millisecond-level real-time inference on cloud images without GPU dependency, with segmentation confidence exceeding 0.9 in thick cloud scenarios. This significantly outperforms traditional methods like Otsu and K-means, providing an interpretable and no-training-required efficient cloud segmentation solution for embedded meteorological observation devices.