Machine learning-based cloud recognition from all-sky camera data for WFST observation optimization

To support cloud-aware observations with the Wide Field Survey Telescope (WFST), we built a region-level dataset from 400 nighttime all-sky images acquired at the Lenghu site between October 2023 and September 2024. The dataset covered 123 nonconsecutive observing nights and comprised 12,800 region-level samples. The valid sky area in each image was divided into 32 regions, each assigned one of three labels: "cloud-free," "partially obscured," or "fully obscured." Frozen Vision Transformer (ViT) encoders from the DINO (self-distillation with no labels) family were shared across regions for feature extraction, while a separate classifier was trained for each of the four radial levels. Four handcrafted feature sets served as baselines, and 95 representation-classifier configurations were compared using five-fold cross-validation grouped by observation night. The combination of DINOv2 ViT-S/14 class-token features and a support vector machine with a radial basis function kernel (RBF SVM) performed best, with five-fold mean values of 91.69%, 91.48%, and 0.8757 for balanced accuracy, macro-averaged F1 score, and the Matthews correlation coefficient (MCC), respectively. Its balanced accuracy was 7.12 percentage points higher than that of the best handcrafted feature configuration and exceeded 90% in all three lunar-illumination groups (bright, gray, and dark nights). Compared with handcrafted features, this configuration achieved higher overall classification performance and smaller performance differences across the four radial levels. The regional predictions can be mapped back to their original spatial positions to produce a spatial cloud-cover map that can serve as region-level input for WFST dynamic scheduling and cloud-aware observations.