Sentinel-2 Masking CNNs Trained on Physics-Supervised Labels

This paper presents a method to improve pixel-level classification of Sentinel-2 imagery by integrating spectral index-based masking with deep learning approaches using 1D, 2D, and 3D Convolutional Neural Networks (CNN1D, CNN2D, and CNN3D). Rather than relying on manually labeled data, the proposed method selects high-quality training samples from PACO (Python-based Atmospheric COrrection software), using pixel selection strategies to remove ambiguous or inconsistent labels. Three selection strategies are explored: full inclusion, uniqueness-based filtering, and physics-based rules. Unlike traditional masking algorithms based only on spectral indices, the CNN models leverage spatial correlations among neighboring pixels across all spectral bands, plus auxiliary features like elevation and illumination, enabling the extraction of more informative representations and improved classification accuracy, particularly in complex scenes. The model is trained using a large global training dataset from PACO, while a separate validation dataset from the same source is used to monitor performance during learning and prevent overfitting. Final evaluation is performed using two independent manually labeled testing datasets (TD1 and TD2) that span diverse land cover types and atmospheric conditions. Compared to PACO's baseline classification, our CNN approaches achieve consistent improvements for normalized Matthews Correlation Coefficient (nMCC), with maximum gains of +3.3 percentage points (pp) on TD1 (from 0.855 to 0.888) and +18.3pp on TD2 (from 0.665 to 0.848). The largest class- wise gains are observed for shadows and clear land-related classes, with up to +22.7pp improvement. These results confirm the effectiveness of the proposed training strategy and its potential for improving label quality in large-scale Earth observation pipelines.