Evaluación de redes convolucionales para la segmentación de objetos geográficos: un insumo para la cartografía básica a escala 1:2000 basado en el catálogo del IGAC

Abstract

This research work explores the use of convolutional neural networks (CNNs) for the automatic segmentation of geographic objects in the generation of basic cartography at a scale of 1:2000, focusing on municipalities in Colombia. The similarities between geographic objects (roads, green areas, forests, water bodies, buildings) and the physical characteristics of the region are analyzed, aligning with the IGAC object catalog. The CNN architectures UNet, DeepLabV3, and LinkNet were selected and evaluated, implementing Transfer Learning in UNet. Data from the IGAC was collected and selected, creating a training dataset and performing preprocessing. The performance of the architectures was evaluated using metrics such as Test Loss, IoU, F1, precision, accuracy, and recall. The results indicated that UNet with Transfer Learning achieved the best overall performance, excelling in IoU, F1, precision, accuracy, and recall. It is important to consider practical factors such as training time and adaptability to new data when choosing the most suitable architecture.