$ show sentinel-hub/sentinel2-cloud-detector
The sentinel2-cloud-detector repository provides a Python package (s2cloudless) for detecting clouds in Sentinel-2 satellite imagery. It implements a pixel-based cloud detector using a pre-trained LightGBM model that takes Sentinel-2 spectral bands as input and outputs cloud probability maps or binary masks with optional post-processing (convolution and dilation). The package is designed as a reusable library for integration into remote sensing workflows. The architecture consists of three core modules: cloud_detector.py (containing the main S2PixelCloudDetector class), pixel_classifier.py (a generic wrapper for applying pixel-wise classifiers over image data), and utils.py (helper functions for data loading and image processing operations). The PixelClassifier class is particularly notable for its reusability, as it decouples the classification logic from the specific model and can work with any classifier adhering to the scikit-learn API (e.g., scikit-learn models, LightGBM). The package includes comprehensive unit tests, CI/CD pipelines for multiple Python versions, and follows modern packaging standards for distribution via PyPI and Conda. The package actively maintains compatibility with current dependencies (lightgbm, numpy, opencv, sentinelhub) and avoids deprecated APIs. It demonstrates strong software engineering practices through clear separation of concerns, type hints, and comprehensive documentation. While the S2PixelCloudDetector class is tailored to the specific cloud detection use case, the underlying PixelClassifier utility represents a broadly applicable pattern for pixel-wise image classification tasks.
The LLM identified 1 files worth highlighting for reuse:
The PixelClassifier class provides a generic wrapper for applying pixel-wise classifiers (e.g., scikit-learn models, LightGBM) over multi-band image data. It efficiently reshapes 4D image data (samples, height, width, bands) into 2D pixel arrays, applies the classifier's predict/predict_proba methods, and reshapes results back to image dimensions. This avoids explicit loops and is optimized for performance. The class is decoupled from specific classifier types via Protocol typing, making it reusable in any image classification task requiring per-pixel inference (e.g., land cover segmentation, object detection) with minimal adaptation.