Metodologia de Fusão de Dados usando aprendizado profundo para segmentação semântica de usos de solo na Amazônia

Abstract

This study proposes a methodology that uses deep learning and a multiresolution segmentation algorithm to perform the semantic segmentation of remote sensing images. The objective of the semantic segmentation is to classify the land use in three regions: forest, pasture and agriculture. Initially, the image is segmented using a convolutional network. Then, an image with homogeneous regions is generated using a multiresolution segmentation algorithm. Finally, a data fusion process is proposed to merge the information from these two segmentation processes. The field of study were areas of the Brazilian Amazon region. The input data used were LANDSAT-8/OLI images. The reference data were extracted from the results of the TerraClass project in 2014. Two sets of data were evaluated: the first with six bands and the second with three bands. Three convolutional network architectures were evaluated along with three optimization methods, SGDM, ADAM, and RMSProp, and two methods for generalization improvement: dropout and L2 regularization. The best model, defined as the association of a network architecture, an optimization method and a generalization method, which had the best performance in the validation set, was submitted to a 5-folder cross validation methodology. The results obtained with the proposed models were compared with pre-trained networks using the knowledge transfer methodology. For this purpose the following pre-tested networks were used: ResNet50, InceptionResnetv2, MobileNetv2 and Xception. Finally, the proposed methodology was evaluated in regions used by other authors. The accuracy values obtained for the images evaluated were higher than 99%, which shows the excellence of the land use classification technique developed in this work for the classification of remote sensing images.