Authors:
Joel Parente de Oliveira
1
;
Marly Guimarães Fernandes Costa
2
and
Cícero Ferreira Fernandes Costa Filho
2
Affiliations:
1
Operations and Management Center of the Amazon Protection System (CENSIPAM), Manaus, Brazil
;
2
Department of Electrical and Computer Engineering, Federal University of Amazonas, Manaus, Brazil
Keyword(s):
Remote Sensing, Image Segmentation, Land Use, Image Database.
Abstract:
This study presents an image database and a convolutional neural network for the segmentation of land use in agriculture, forest and pasture classes. LANDSAT-8/OLI images from an area of the Brazilian Amazon region were used. The reference data were extracted from the results of the TerraClass project in 2014. The image database was generated in two versions: the first with six bands and the second with three bands. Each version of the data set has 4,000 images and size 400x400 pixels. Each image was generated using the mosaic technique. Each mosaic image is created from small agricultural, forest and grassland patches that are extracted from satellite images. The mosaic image is created with almost the same amount of agriculture, forest and pasture patches. The convolutional neural network architecture was evaluated together with three optimization methods: SGDM, ADAM and RMSProp and the dropout and L2 regularization for generalization improvement. The best model, CNN + optimization
method + technique for generalization improvement, evaluated on the validation set, was used to segment some regions of the Amazon. The best results were obtained using the ADAM optimization method and L2 regularization. The accuracy values obtained for the evaluated images were above 94%.
(More)