CloudMatch: Weak-to-Strong Consistency Learning for Semi-Supervised Cloud Detection

TL;DR

CloudMatch advances semi-supervised cloud detection by combining view-consistency learning with a dual scene-mixing augmentation that includes inter-scene patch blending and intra-scene within-image transformations. It enforces class-level consistency between weak and strong augmented views and uses pseudolabel supervision for unlabeled data, all within a CD-Mamba backbone that captures long-range dependencies. Empirical results on Biome, SPARCS, and RICE show state-of-the-art performance under low-label regimes and strong cross-dataset generalization, highlighting robustness to diverse cloud morphologies and ground conditions. The approach reduces annotation costs while maintaining high segmentation fidelity, enabling practical deployment for remote sensing applications.

Abstract

Due to the high cost of annotating accurate pixel-level labels, semi-supervised learning has emerged as a promising approach for cloud detection. In this paper, we propose CloudMatch, a semi-supervised framework that effectively leverages unlabeled remote sensing imagery through view-consistency learning combined with scene-mixing augmentations. An observation behind CloudMatch is that cloud patterns exhibit structural diversity and contextual variability across different scenes and within the same scene category. Our key insight is that enforcing prediction consistency across diversely augmented views, incorporating both inter-scene and intra-scene mixing, enables the model to capture the structural diversity and contextual richness of cloud patterns. Specifically, CloudMatch generates one weakly augmented view along with two complementary strongly augmented views for each unlabeled image: one integrates inter-scene patches to simulate contextual variety, while the other employs intra-scene mixing to preserve semantic coherence. This approach guides pseudolabel generation and enhances generalization. Extensive experiments show that CloudMatch achieves good performance, demonstrating its capability to utilize unlabeled data efficiently and advance semi-supervised cloud detection.

Figures (11)

• Figure 1: CD-Mamba network architecture. The model is based on a U-shaped structure, integrating convolutional modules with Cloud-SMB (Cloud Spatial Mamba Block) modules and incorporating dual-attention blocks (DA Blocks) in the skip connections to enhance cloud boundary detection accuracy.
• Figure 2: CloudMatch architecture and scene-mixing augmentation framework. The prediction represents a probabilistic prediction map, where each pixel value ranges from $[0,1]$, while the pseudolabel corresponds to a binary prediction map with values of $\{0,1\}$.
• Figure 3: Comparative experimental results of different backbones on the Biome dataset.
• Figure 4: Comparative experimental results on the RICE dataset.
• Figure 5: Comparative experimental results on the SPARCS dataset.