How does pre-training a Convolutional Neural Network decrease the amount of in-situ data needed to perform agricultural field delineation?

Details

Supervisors

Defourny, Pierre ; Deffense, Quentin

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

enAccurate agricultural field boundary delineation is essential for crop monitoring, yield estimation, and sustainable land management, yet in-situ reference data are often scarce, expensive to obtain or unavailable. CNN models offer a powerful approach for deriving such information from RS data, however, their performance typically depends on large amounts of labelled training data. In data scarce regions, there is a need for strategies to reduce reliance on in-situ data while maintaining reliable segmentation accuracy. Pre-training a model has been shown to decrease training time and improve model performance in some applications, particularly under limited in-situ data conditions. EL approaches have also demonstrated strong potential for improving segmentation quality and can be used to generate unsupervised field boundary information suitable for pre-training. However, the effect of pre-training a CNN using EL for the application of agricultural field boundary delineation remains largely unexplored in literature. This study contributes to addressing this gap. In this study, two modelling approaches are compared, a Scratch model trained solely on labelled data and a Pretrained model pretrained on EL derived segmentation products. Both are applied under progressively reduced in-situ training datasets. Pre-training reduces the number of training epochs by an average of 23% compared to the Scratch models. Pre-training also improves pixel-based performance, with MCC increasing from 0.63 (best Scratch model, 100% dataset) to 0.69 (Pretrained, 25% dataset) and from 0.41 (1% Scratch) to 0.55 (2% Pretrained). In contrast, improvements in object-based metrics are less pronounced, with IoU ≥ 0.5 and median IoU increasing slightly from 0.16 and 0.11 (Scratch) to 0.23 and 0.18 (Pretrained), respectively. Overall, the benefits of pre-training are most evident under conditions of limited in-situ data. Overall, EL based pre-training enhances CNN performance to limited amounts of in-situ data and improves field and non-field discrimination, but further work is required to refine boundary precision and reduce oversegmentation. The approach shows potential for application in data-scarce regions where high-quality field boundary datasets are unavailable.