CNN Sensitivity Analysis for Land Cover Map Models Using Sparse and Heterogeneous Satellite Data

Land cover maps provide detailed information on the land use of territories, which is useful for public policy making. Constant changes in the landscape limit the usefulness of these maps over time, so they need to be constantly updated. In this context, remote sensing images combined with the use of deep neural networks can be used for this purpose. Although several models are trained on different datasets, we do not know their ability to transfer the learned patterns to new data. In this paper, we evaluate several pre-trained semantic segmentation models on deep convolutional neural networks (CNN) using freely available global RGB data from Sentinel-2. Four CNN models with 32 different architectures were evaluated on data from three continents, on seven different classes. The results show that the best model is the PSPNet with seresnet18, obtaining a test macro F1 score of 0.4950 when the model is trained with data augmentation and fine-tuning.