To what extent are field data dispensable for training the ResUNet-a neural network in automatic field boundaries delineation?

Details

Supervisors

Defourny, Pierre

Faculty

Faculté des bioingénieurs

Degree label

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

Abstract

Accurate field delineation is critical to the implementation of up-to-date and accurate agricultural monitoring systems. Faced with challenges such as limited access to field data and modern technology, developing countries need innovative solutions to ensure effective monitoring of their agricultural systems. The traditional method of manual parcel delimitation, as used in Europe, is proving too costly, time-consuming and error-prone to be applicable in developing countries. In this context, the use of deep learning could meet the needs of these countries, but one drawback is that it requires a very significant volume of field data to be effective. The main objective of this thesis is therefore to train a deep learning model using a classical unsupervised image segmentation algorithm in order to minimise the need for large amounts of field data through the use of transfer learning. This research is carried out in Wallonia, Belgium, taking advantage of the large volume of field data provided by its LPIS. The ResUNet-a neural network is used and trained several times independently, with different types of training and amounts of field data, resulting in a number of different versions of the model. Comparison of the performance of each model shows that pre-training a model with algorithmically generated data can substantially improve performance, but only for small training datasets. The study also demonstrates the resilience of the ResUNet-a model through its ability to perform despite the drastic reduction in the number of labelled fields used for training. The approach studied could therefore be beneficial to all countries that do not have an annual parcel system, enabling significant savings in both manpower and image acquisition costs by reducing the need for manual digitisation based on visual interpretation by up to 90%, with only a slight decrease in model performance. Deep learning proved to be a robust solution, offering automatic feature learning and end-to-end learning capabilities, overcoming the drawbacks of traditional segmentation approaches that rely on manual feature refinement and multiple stages processing. Nonetheless, despite promising results, the approach studied is not necessarily the most effective. In-depth study of deep learning, and in particular transfer learning from a region with abundant field data to another where data is scarce, could help to create more appropriate and effective solutions for providing countries, that do not have an annual parcel system, with accessible decision-making tools in the field of agriculture.