Mapeamento de uso e cobertura da terra utilizando sensoriamento remoto e redes neurais convolucionais

Abstract

The increase in the world population requires an expansion in food demands, consequently increasing agricultural production. Land Use and Land Cover (LULC) detailing plays an essential role in the agricultural sector, enabling efficient monitoring, planning, and management of these areas. In this segment, remote sensing techniques have proved to be a valuable tool for mapping large agricultural areas. Therefore, the general objective of this research was to explore machine learning methods to carry out the LULC mapping through satellite images of three study areas in the state of Paraná. In addition, the generalization of the models was evaluated through cross-site classification. The work was divided into three stages covered in different scientific papers. The first paper proposed a one-dimensional Temporal Convolutional Neural Network (1D-TempCNN) to classify LULC using Satellite Image Time Series (SITS). Two other classifiers, Random Forest (RF) and Support Vector Machine (SVM), were used to compare the results. The Overall Accuracy (OA) was above 98% for all models when the test was performed in the same training area. However, in the cross-site classification, 1D-TempCNN showed better OA values (between 94.34% and 98.67%) and greater generalization. Two Data Augmentation (DA) techniques, sliding window and scaling, contributed to the generalization of the models. This way, the proposed architecture proved viable for cross-site classification and can be used in different crop years (cross-year) or agricultural areas (cross-site). The second paper explored the early classification using the 1D-TempCNN architecture and two classic models, Multilayer Perceptron (MLP) and RF. The models showed similar performance, reaching OA above 95% at the end of December. However, in the cross-site classification, only the 1D-TempCNN model achieved OA above 95% in all test scenarios, reaching this value between the beginning of December and the first half of February. Thus, this model demonstrated generalization capacity and can be used for early classification in different training areas. The third paper addressed the use of semantic segmentation to build LULC maps. Two pre-trained DeepLabv3 architectures (ResNet-50 and ResNet-101) were evaluated along with two different segmentations (true color and false color) and two training image sizes (256 x 256 and 512 x 512 pixels). The reference maps used in training and testing were derived from the results of the first paper. The OA presented results between 74.91% and 77.81%, and those of the Mean Intersection over Union (MIoU) metric between 39.46% and 52.56%. In addition, the combination of false color bands was superior to true color, and the use of smaller images resulted in more detailed and accurate maps. The model with the ResNet-101 base network presented the best results in most of the analyzed metrics. However, distinguishing between soybean and corn classes was the most significant difficulty. Therefore, this model presented generalization capacity, proving to be a viable option for constructing large area LULC maps, which allows the monitoring and planning of agricultural areas.