Assessing fire-exposed vegetation in the Orinoco Basin using Sentinel-2 and deep learning.

The Orinoco Basin in Colombia is undergoing significant land-use change, where fire is commonly employed for pasture renewal, crop management and deforestation. This study aims to enhance the understanding of vegetation types affected by fire in the Orinoco Basin by applying a deep learning approach to burned area detection using Sentinel-2 imagery. The model used local spatial context from 3×3 multispectral image chips and learned spectral representations through convolutional neural networks (CNNs) to identify burned area fragments. The Gini index was used to assess the relative importance of MSI spectral bands and guide the selection of explanatory variables for the CNN classifier. Using PlanetScope imagery as reference data we found a Dice Coefficient (DC) of 0.72 compared to 0.37 for MCD64A1. The higher spatial resolution of Sentinel-2 particularly reveals the high frequency of small burned patches and highlights fire impacts on riparian vegetation and tropical forests. By enabling more accurate burned area mapping, this approach contributes to improved assessments of land degradation and deforestation in critical ecosystems, this work directly supports the objectives of SDG 15 (Life on Land). Furthermore, since fires are closely linked to the increase in greenhouse gas emissions, improving the accuracy of burned area estimates is fundamental for climate accounting and advancing the goals of SDG 13 (Climate Action).