SatRed: New classification land use/land cover model based on multi-spectral satellite images and neural networks applied to a semiarid valley of Patagonia

Abstract

In this article we describe a new model, SatRed, which classifies land use and land cover (LULC) from Sentinel-2 imagery and data acquired in the field. SatRed performs pixel-level classification and is based on a densely-connected neural network. The study site is the lower Chubut river valley which has an extension of 225 km2 and is located in estern semiarid Patagonia. SatRed showed a 0.909 ± 0.009% (mean ± sd, n = 7) overall accuracy and outperformed the seven most traditional Machine Learning methods, including Random Forest. Our model accurately predicted buildings, shrublands, pastures and water and yielded the best results with classes harder to classify by all methods considered (Fruit crops and Horticulture). Further improvements involving textural information and multi-temporal images are proposed. Our model proved to be easy to run and use, fast to execute and flexible. We highlight the capacity of SatRed to classify LULC in small study areas as compared to large data sets usually needed for state-of-the-art Deep Learning models suggested in literature.