Automatic field extraction using deep learning in the context of smallholder agriculture

Details

Supervisors

Defourny, Pierre

Faculty

Faculté des bioingénieurs

Degree label

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

Abstract

Sub-Saharan Africa is the center of the world’s population growth but is particularly vul- nerable to food shortages. In the framework of collaboration between the Institut National de la Statistique du Mali (INSTAT) and UCLouvain, the goal is to achieve a better understanding of agricultural dynamics and improve agricultural statistics. Automatic field extraction from satellite imagery is an essential tool to achieve agricultural landscape mapping, which acts as a decision-support tool to help improve food security. A newly acquired set of reference agricultural field delimitations of 16.000km2 has been prepared by the Malian partner of the Institut d’Economie Rurale (IER) in order to make the first step in that direction. This thesis aimed to evaluate the performance of an automatic field extraction methodology from High Resolution satellite images, in the context of smallholder agriculture, which is characteristic of the sub-Saharan region. The methodology relies on a Deep Learning model trained with satellite images from 2021 captured by the Planet satellite and the reference dataset. The first step consisted of realizing a calibration analysis to set the best hyperparameters for the model training. The baseline method was then evaluated with carefully selected metrics showing the performance of our model. The last step includes an assessment of the model’s portability within Africa, as well as an exploration of possible improvements by adjusting the quality of the input data and the imagery’s spatial resolution. Overall, the results were mixed. The model is mostly capable of finding the fields’ extent but struggles to capture field boundaries correctly. The portability of the model is promising, but it suffers from land cover and topography differences. A major drawback is the lack of reliability of the reference dataset. Finally, recommendations and improvement opportunities were discussed for a potential continuation of experiments.