Mapeamento de cobertura e uso de solo a partir de dados de sensoriamento remoto utilizando redes neurais

Abstract

Land cover and land use are two key elements that describe the terrestrial environment in relation to human and natural activities. Land cover is characterized by the biophysical features of the terrestrial environment, and land use is the way that those features are used by humans. The map generated by these two elements is usually man-made, thus taking a lot of time for it's development. This work aims to generate a land use and land cover map on remote sensing data using convolutional neural networks and autoencoders. The Python language was used for the development along with a Machile Learning API named Keras. The databases used for training were the RSI-CB128, UCMerced Land-Cover and DLRSD, of which all of the databases contains RBG images of 128x128 pixels. Lastly a land use and land cover map was generated for the city of Pato Branco using satellite images provided online by Google.