Metodología para la generación (detección y clasificación) de cambios de cobertura de la tierra mediante el análisis de imágenes multilaterales basada en algoritmos de deep learning

Abstract

This research presents the results of the application of a methodological framework for the detection, classification and quantification of land cover changes due to mining and deforestation through a multitemporal analysis of multispectral satellite images based on Deep Learning algorithms. The area of interest for the estimation of areas of change due to mining corresponds to a region of the Cerrejón open pit mine (Department of La Guajira), in the period October 2017 to November 2019; while the area of interest for the detection of deforestation corresponds to a region of the Tinigua National Natural Park (PNN) and an area of the Cordilleras de los Picachos PNN (Department of Meta), in the period December 2015 to December 2019. For the formulation of this framework, a comparison of two change detection and classification schemes was carried out through the application of Deep Learning (DL) algorithms. The first scheme was based on the post-classification comparison of thematic maps of mining areas and deforested areas generated by two convolutional neural network (CNN) architectures: U-Net and FPN (Feature Pyramidal Network) over two time series (2017 to 2019 - Mining and 2015 to 2019 - Deforestation) of Planet multispectral satellite images. The second scheme is based on direct change detection between pairs of Planet multispectral satellite images using a modified UNet network. The results indicate that change detection by mining was more accurate using a direct detection scheme (kappa greater than 0.9 in each period of the time series) versus the post-classification comparison scheme (kappa between 0.7 and 0.9). While the detection of deforested areas is more accurate using a post-classification change detection scheme (kappa between 0.7 and 0.9). (average kappa 0.97) versus an average kappa of 0.94 for the direct detection scheme. The results obtained and the different processes used allowed us to propose a methodological framework for the detection, delimitation and classification of changes on this type of land cover under DL algorithms in a timely manner, analyzing large volumes of data and obtaining a high thematic accuracy in the different schemes evaluated.