Resumo

SUAREZ L, Arnol S; JIMENEZ L, Andrés F; CASTRO-FRANCO, Mauricio  e  CRUZ-ROA, Angel. Classification and automatic mapping of land covers in satellite images using Convolutional Neural Networks. Orinoquia [online]. 2017, vol.21, suppl.1, pp.64-75. ISSN 0121-3709.  https://doi.org/10.22579/20112629.432.

Land cover classification is important for studies of climate change and monitoring of ecosystem services. Conventional coverage classification methods are performed by the visual interpretation of satellite imagery, which is expensive and inaccurate. Implementing computational methods could generate procedures to classify coverage in satellite images automatically, quickly, accurately and economically. Particularly, automatic learning methods are promising computational methods for estimating soil cover changes. In this work we present an automatic learning method based on convolutional neural networks of ConvNet type architecture for the automatic classification of soil coverings from Landsat 5 TM images. The ConvNet was trained from the manual annotations by means of visual interpretation on the satellite images with which the experts generated the map of Tuparro national park, of National Natural Park of Colombia. The validation model was performed with data from the Colombian Amazon cover maps made by the Colombian Environmental Information System. The results obtained from the diagonal of the confusion matrix of the average accuracy were 83.27% in training and 91.02% in validation; for the classification in patches between forests, areas with herbaceous and / or shrub vegetation, open areas with or without vegetation and Inland waters.

Palavras-chave : Automatic learning; land cover; natural parks; convolutional neural networks; remote sensing..

        · resumo em Português | Espanhol     · texto em Espanhol     · Espanhol ( pdf )