Image-based Semantic Segmentation Network for the Brazilian Cerrado based on Public Databases

We have developed fully convolutional networks (FCN) for semantic segmentation of satellite imagery based on five Land Use and Land Cover (LULC) categories: native vegetation, agriculture, pasture, urban region, and waterbody. To this end, we gathered and preprocessed public Brazilian data into an annotated dataset with 26,000 segmented 224 × 224 image patches. We obtained images from the Sino-Brazilian CBERS-04A satellite program and segmentation masks from the Terra Class project (INPE/Embrapa). We performed transfer learning on four backbone models: DeepLabV3+, MobileNet, ResNet-50, and VGG-16. We evaluated their performance with IoU, with the respective scores of 45.96%, 34.40%, 45.58%, and 62.78%. However, our dataset is unbalanced, and a balanced IoU yields scores lower than 20% for all models, indicating specialization on majority classes. Despite these shortcomings, our models’ masks have a 16-times higher pixel-density than previously available masks, and taking only images as input, without external data or expert curation.