Creación de un dataset sobre coberturas del suelo en el oriente antioqueño – caso de estudio: Cocorná, Carmen de Viboral, Granada y Sonsón

Abstract

The main objective of this research project is to create a dataset on land cover in Eastern Antioquia, with case studies in Cocorná, Carmen de Viboral, Granada, and Sonsón. This study highlights the lack of updated, georeferenced land use information, which poses a challenge for territorial and environmental planning. To address this need, the research project proposes creating the dataset using satellite image processing, segmentation, manual labeling, and deep learning techniques. Using the Google Earth Engine platform, images were collected, processed, augmented, and divided into small 96x96 pixel patches. These were then manually labeled into five land cover categories: Forests and semi-natural areas, agricultural land, artificial land, road network, and water bodies. To theoretically represent the categories, a formal ontology was developed using the Methodology methodology and implemented in Protegé software The MobileNetV2 architecture was used, which is a convolutional neural network (CNN) trained with more than 20,000 images organized by class. The model was validated with 20% of the total set. The results showed that the model achieved an accuracy of 76%. The classes with the best predictions were artificial land and agricultural land. However, there were problems predicting the class corresponding to forests and semi-natural areas, which was the class with the worst prediction. The road network class also had low accuracy, as it was poorly represented in the original dataset, necessitating the application of data augmentation strategies. Finally, the water surface class performed well in terms of model accuracy.