Land Cover Maps Production with High Resolution Satellite Image Time Series and Convolutional Neural Networks: Adaptations and Limits for Operational Systems

The Sentinel-2 satellite mission offers high resolution multispectral time-series image data, enabling the production of detailed land cover maps globally. When mapping large territories, the trade-off between processing time and result quality is a central design decision. Currently, this machine learning task is usually performed using pixel-wise classification methods. However, the radical shift of the computer vision field away from hand-engineered image features and towards more automation by representation learning comes with many promises, including higher quality results and less engineering effort. In particular, convolutional neural networks learn features which take into account the context of the pixels and, therefore, a better representation of the data can be obtained. In this paper, we assess fully convolutional neural network architectures as replacements for a Random Forest classifier in an operational context for the production of high resolution land cover maps with Sentinel-2 time-series at the country scale. Our contributions include a framework for working with Sentinel-2 L2A time-series image data, an adaptation of the U-Net model (a fully convolutional neural network) for dealing with sparse annotation data while maintaining high resolution output, and an analysis of those results in the context of operational production of land cover maps. We conclude that fully convolutional neural networks can yield improved results with respect to pixel-wise Random Forest classifiers for classes where texture and context are pertinent. However, this new approach shows higher variability in quality across different landscapes and comes with a computational cost which could be to high for operational systems.