Sentinel-2 satellites for crop-type mapping: use case in Kolda, Senegal, using spectral, temporal augmented and non-augmented data

Details

Supervisors

Defourny, Pierre ; Bontemps, Sophie

Faculty

Faculté des bioingénieurs

Degree label

Master [120] : bioingénieur en sciences et technologies de l'environnement, à finalité spécialisée

Abstract

Smallholder agriculture is crucial for food production in Africa, particularly in Senegal, but faces challenges that create an urgent need for accurate crop monitoring and mapping systems. Remote sensing data and advanced machine learning techniques offer promising solutions to map smallholder agricultural landscapes, serving as decision-making tools and providing comprehensive land-use change assessments. While crop monitoring and mapping research has expanded to include sub-Saharan Africa, results remain variable, and accurate classification of certain crop types continues to pose challenges. This variability in outcomes underscores the need for further investigation and method refinement. Therefore, the objective of this master’s thesis is to design and compare two classification approaches using Sentinel-2 imagery: one utilizing time series data and another using selected cloud-free images. To meet these objectives, an extreme gradient boosting classifier was employed to test nine different scenarios for the time series classification. These scenarios explored various SMOTE algorithms to resample the in-situ data. The same classifier was also used for the classification on selected cloud-free images. Both approaches utilized spectral bands and computed metrics specific to crop temporal patterns, aiming to optimize performance in the challenging context of Senegalese smallholder agriculture. The time series approach using both spectral bands and computed metrics showed varying performance across different classes, with non-crop classes consistently achieving higher accuracy than crop classes. The best overall performance was achieved using the SMOTEENN algorithm, combining spectral bands and metrics. The cloud-free image approach demonstrated a slight improvement in overall accuracy compared to the time series method, but its applicability was limited to a single Sentinel-2 tile. Both approaches faced challenges in accurately classifying specific crop types, particularly those with similar spectral signatures or grown in mixed or low-density fields. The study concludes by discussing potential improvements to the classification approaches and proposing ideas for further development.