Clasificación de imágenes de radar de apertura sintética aplicando Corine Land Cover adaptada para Colombia mediante redes neuronales convolucionales

Abstract

In recent years Deep Learning, a field of the Artificial Intelligence has come gaining in strength with the advances in technology, inside their different algorithms there the Convolutional Neural Networks (CNN) which are widely used in image classification and object detection. In this paper we use this network with Synthetic Aperture Radar Satellital Images that have the capacity of obtain land photographs independent of atmospheric conditions. We trained this network with classified images applying "CORINE Land Cover" (CLC) adapted to Colombia in some specific country zones using Sentinel-1 Images. Due Limitations with the first designed network, the data was scaled between 0 - 255 values and we used the "RestNet50" architecture, achieving accuracy of 87.9% and 82.9% for CLC level 1 and 2 respectively. In total were used 9 of 15 categories of level 2 CLC, given to the lack of training examples of certain categories, this represent the limitations of the amount of data required by the CNN. Notwithstanding of this limitations, it can be assumed that with more amount of data for training, metric and quantity of covers can increase, this shows great potential for the future application of these techniques in classification and maps creation of the land coverage.