Estudio de viabilidad en la delimitación de coberturas con imágenes satelitales mediante el modelo Segment Anything: Comparación con la metodología CORINE Land Cover en el municipio de Guatavita

Abstract

This research evaluates the viability of using the open-source tool SAM to delimit land cover at the municipal level in Guatavita, Cundinamarca, by assessing its performance against cartography generated using the CLC methodology for Colombia in 2018. Sentinel-2 satellite images with an L2A processing level, captured in 2018, were used as inputs. A set of tiles was constructed for each satellite image, along with its respective CLC reference. The methodology consisted of selecting and cropping the image according to the study area, executing SAM in automatic mode using the SamGeo package, and using the IoU, Dice, TPR, FPR, and pixel accuracy validation metrics, as well as guided instructions (points, boxes, and text). Results showed better performance on water surfaces (IoU: 0.72; Dice: 0.80), followed by agricultural areas (IoU: 0.41; Dice: 0.50), and forests/semi-natural areas (IoU: 0.37; Dice: 0.42). Performance was lowest on artificial areas (IoU: 0.25; Dice: 0.33). In this study, spatial instructions outperformed text and automatic mode. SAM is not viable under the automatic method. However, it showed great potential with georeferenced instructions, such as points and boxes. Further fine-tuning and evaluation of its use as support in supervised segmentation tasks is recommended.