Exploiting multi-resolution earth observation data and intra-parcel heterogeneity information for improving crop type mapping in smallholder cropping systems in Mali

Details

Supervisors

Defourny, Pierre ; Bontemps, Sophie

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

enSmallholder cropping systems present a challenging scenario for crop type mapping with remote sensing approaches. Parcels are small and exhibit complex within-field heterogeneity, containing mixtures of crops and non-crop features such as trees, weeds and bare-soil. Leveraging a unique dataset of parcel-level heterogeneity ratings for various crop parcels in Mali's cotton belt, this study explored the potential of integration with multi-resolution satellite imagery from Sentinel-2, PlanetScope and Pléiades. Pixel-based classification was performed to compare the performance of the three sensors with varying spatial and spectral resolutions and temporal data depth. The study specifically investigated how incorporating intra-parcel heterogeneity information and excluding trees from training data influences classification accuracy. A Pléiades derived cropland mask was further created to refine validation by accounting for non-crop elements within agricultural parcels. A comprehensive classification framework was implemented, incorporating monocrop agricultural classes - focusing on cotton, sorghum, millet, maize, sesame and peanut and various non-crop landcover types, creating a challenging multi-class problem representative of the smallholder landscape mosaic. Incorporating heterogeneity showed to improve classification performance: overall accuracy increased by ~4% for Sentinel-2 (from 71.2% to 76.3%) and PlanetScope (70.2% to 74.4%). Masking out tree-covered pixels within cropland had negligible effects (<0.5% change) across all sensors. Refined validation using a VHR cropland mask produced modest, consistent improvements, revealing that including non-crop elements within parcels in validation data underestimates crop classification performance. Leveraging intra-parcel heterogeneity can enhance crop type mapping, but effective integration requires more detailed characterization of within-field variability. Relating field-assessed heterogeneity to parcel-level spectral metrics proved inconclusive, with no clear patterns across crop types. Future work should aim to decompose heterogeneity into its components (e.g., weed presence, crop density, bare soil) and create denser time series to capture full crop phenologies to improve discrimination between crop types and enable the incorporation of intercropping into classification frameworks.