Sentinel 2 Classification Using CNNs Trained on Multi Label Pre-classification Masks and Validation on Independent Datasets

Kurzfassung

This study presents a deep learning method for pixel-level classification of Sentinel-2 Level-1C imagery, trained using multi-label masks that are automatically generated during the early stages of atmospheric correction by PACO (Python-based Atmospheric Correction) software. The method uses 1D, 2D, and 3D convolutional neural networks that combine spectral bands, digital elevation, and illumination data to model spatial and spectral context. The training dataset is filtered using physics-based pixel selection rules defined on PACO's multi-label masks. These rules discard combinations of active class labels that are not physically compatible, such as simultaneous detection of snow and cloud. The model incorporates label uncertainty into the training process by including the full multi-label information in the loss function. Validation uses two independent, manually labeled datasets. The first includes globally distributed scenes labeled through visual inspection. The second focuses on pixels with high classification uncertainty. Compared to the PACO baseline, the trained models show improvements in normalized Matthews correlation coefficient (nMCC) of up to +3.3 percentage points on the first dataset and +18.3 points on the second. The largest gains are observed in challenging classification cases, particularly for the shadow and clear-sky classes. These results highlight the potential of combining convolutional models with automatically generated labels and pixel selection rules to support atmospheric correction for Sentinel-2.