Validación y refinamiento de alertas de deforestación para la amazonia colombiana mediante el método MapBiomas alerta

Abstract

This technical report documents the outcomes of an internship carried out at the Gaia Amazonas Foundation within the framework of the MapBiomas Alerta project, aimed at validating and refining deforestation alerts in the Colombian Amazon to optimize the precision of open geospatial data using SCCON, Google Earth Engine, and Alerts Workspace platforms. Through a multi-temporal human-validation methodology supported by a structured catalog of technical criteria, 682 initial alerts (4,224.69 hectares) were processed, identifying and filtering out 33.58% of false positives mainly caused by spatial redundancy among sensors. Following advanced stages of manual geometric polygon refinement and a rigorous internal quality audit that yielded a 97.28% reliability rate, the official publication of 378 final alerts was consolidated (3,429.32 hectares), heavily concentrated within the departments of Caquetá, Guaviare, and Meta. The spatial analysis characterized extensive livestock farming as the primary land-use driver (45.24%), while successfully identifying 25.66% of the alerts as indigenous traditional cultural land use (chagras), thereby preventing the criminalization of ancestral subsistence practices. Lastly, visual quality glitches during the re-exporting process of Planet imagery were identified as the main technical bottleneck (accounting for 78.26% of rejections during the refinement phase), leading to the recommendation of integrating partial automated filtering tools within the platform to optimize operational times without replacing essential human interpretation.