Artificial Intelligence for the Classification of Earth Observation Satellite Images

Kurzfassung

This work presents a deep learning framework for pixel-level classification of multispectral and hyperspectral satellite imagery for Earth observation. Leveraging data from the Sentinel-2 and EnMAP missions, the study explores convolutional neural network (CNN) architectures (1D, 2D, and 3D) to improve classification accuracy under variable atmospheric and surface conditions. Automatically generated labels from the Python-based Atmospheric Correction (PACO) system were used to train and evaluate the models across diverse global biomes. Results demonstrate that physics-based pixel selection and domain-generalized CNNs enhance the detection of key surface classes such as clouds, water, vegetation, and snow. The proposed approach is sensor-agnostic and scalable to future Earth observation missions.