Clasificación de la cobertura del suelo en la Provincia del Chimborazo mediante imágenes satelitales y aprendizaje automático

Abstract

In the province of Chimborazo, the lack of information and geographic studies carried out is one of the causes of the lack of information regarding the types of land cover. In the present work, the CART, RF, SVM classification algorithms based on supervised learning of ML and S2 satellite images have been developed. TIC support has played an important role as a provider of data and cloud processing through the GEE platform. 7 classes of land cover have been defined: water, urban, forest, cultivation, bare soil, bushy and snow which is classified using 7 of 13 discriminant bands of S2, in addition to the NDVI, BSI and RGB indices for dataset composition. With 73% of training data (1581) and 27% for testing (598) labeled in the GEE platform, the input datasets have been built. To obtain the classifiers, the phases of the CRISP-DM methodology are applied, which has allowed the development of the three validated ML models, executing three test cases for each one, where it is determined that the RF and SVM algorithms achieve better metrics above ≈80%. Once the ML models have been validated by applying statistical inference, the correlation analysis and hypothesis test have been carried out, with a kappa index different from zero for the three ML models, the existence of concordance is established; therefore, it is claimed that satellite images provide sufficient discriminatory information. Finally, with the analysis and discussion of the results, it is determined that the RF algorithm has a better performance based on the results of general precision ≈88.80% and kappa index of 0.867. The work ends by giving guidelines for future studies such as: comparison of coverage changes and implementation of ML models in decision support systems, taking the present study as a reference.