Evaluating Multi-Sensor Remote Sensing for Parcel-Level Mapping of Milpa Mixed-Cropping and Maize Monocrop Fields in Oaxaca, Mexico

Details

Supervisors

Defourny, Pierre ; Gerald Blasch

Faculty

Faculté des bioingénieurs

Degree label

Master [120] en sciences agronomiques et industries du vivant, à finalité spécialisée: Geo-Information Science and Earth Observation for Environmental Modelling and Management

Abstract

This thesis evaluates whether complementary information from SkySat and PlanetScope improves parcel-level discrimination of the Milpa mixed-cropping system from maize monocrop beyond a seasonal Sentinel-2 baseline in Oaxaca, Mexico. A 2024 field-reference dataset was screened to a final analytical cohort of 149 parcels across four areas of interest; 71 parcels from AOI3 and AOI4 were used for development and 78 parcels from AOI1 and AOI2 formed a secondary, non-confirmatory benchmark. Sentinel-2 time series were harmonised to 36 ten-day dates and represented using spectral, temporal, phenological, heterogeneity and observed-texture features. The primary parcel-level Random Forest achieved binary benchmark accuracy of 0.78, balanced accuracy and macro-F1 of 0.62, while Milpa recall remained 0.36. SkySat and PlanetScope provided some favourable Milpa-sensitive benchmark results, with both native representations reaching Milpa recall of 0.50, although gains were geographically inconsistent and accompanied by additional maize false positives. Predictor concatenation did not provide a consistent advantage. SkySat-guided CNN super-resolution improved image-reconstruction error relative to bicubic interpolation, but this did not translate consistently into better downstream crop-system classification. Among fixed classification-level combinations, Sentinel-2 + SkySat produced the strongest pooled benchmark fusion, while no fusion improved development Milpa recognition. Overall, finer spatial information was complementary but did not remove the difficulty of geographically transferable Milpa discrimination, highlighting the need for stronger reference data and multi-season validation.