Abstract

Accurate and update delimitation of agricultural areas is essential for supporting official statistics in Mexico and designing public policies that support food security and sustainable development. This article presents an automated methodology for identifying agricultural areas, using a combination of remote sensing techniques and artificial intelligence. This work focuses on a supervised classification method at the pixel level. The model is based on a multivariate dataset that integrates various geospatial sources to maximize predictive capacity. Sentinel-2 satellite imagery serves as the basis, using 12 spectral bands derived from geomedian composites to achieve a spatial resolution of 10 meters, while reducing temporal variability and cloud interference. To improve class separability, we include spectral indices such as the Enhanced Vegetation Index (EVI) and the Normalized Difference Built-up Area Index (NDBI). Additionally, key topographic variables—altitude, slope, and aspect—from the Mexican Elevation Continuum (CEM 3.0) are incorporated to account for terrain characteristics that influence agricultural land use. Training is based on a set of continuously refined, labeled polygons representing the various land cover types across the country. From these labels, we implement a machine learning process using an ExtraTrees classifier combined with stacking techniques to optimize performance. Tests conducted on sample areas of 150 × 150 km reflect results with classification accuracy greater than 98% on the validation datasets. In conclusion, with the tests conducted in the selected areas, the feasibility of integrating advanced artificial intelligence and remote sensing techniques for the identification of agricultural areas is demonstrated.