Automating field boundary delineation with multi-temporal Sentinel-2 imagery

Abstract

Listen

Translate article

Knowledge of the extent and location of cultivated fields are critical for agricultural monitoring, food security planning and commodity trading. The increased observational capacity of remotely sensed data, such as Sentinel-2 imagery, constitutes a major advantage over in situ methods, particularly for agricultural applications that require frequent updates. However, the development of fully automated cultivated field delineation methods remains a challenge. In this study an existing object-based image analysis (OBIA) methodology, called Canny edge-detection in conjunction with watershed segmentation (CEWS), was modified to automatically identify and delineate agricultural fields from multi-temporal Sentinel-2 imagery in five diverse agricultural landscapes. The robustness of the technique was evaluated by comparing its outputs to those of standard per-pixel, supervised classifications. Reference field boundaries were manually digitised and used for quantitative accuracy assessments. Area and edge metrics were used to evaluate the accuracy of the extracted boundaries. Results show that, on average, CEWS produced significantly higher field boundary accuracies than the supervised per-pixel approach in terms of both mean absolute error (MAE) (~31 m lower) and overall accuracy (~13.8% higher). The closed field boundaries produced by CEWS, as well as its capability to operate without any a priori knowledge, was found to be its main strengths for integration into operational workflows. However, in complex landscapes (e.g. Swartland and Mooketsi) it was found that some fields had weak boundaries due to the homogeneity between adjacent fields, making edge detection problematic. The incorporation of higher resolution imagery into the edge detection process is proposed, particularly in agricultural areas with very small and irregularly shaped fields. Despite these challenges, it was concluded that the multi-temporal CEWS approach will perform well in most agricultural areas and is suitable for operational implementation.