Detecting abandoned farmland using harmonic analysis and machine learning

It is critical to inventory abandoned farmland soon after it is generated, to better manage agricultural resources and to prevent negative consequences that would otherwise follow. This study aims to distinguish abandoned farmlands from active croplands-rice paddy and agricultural fields-by discerning the phenological trajectories over a short-term period of three years (Jan. 2016 to Dec. 2018) in Gwanyang City in South Korea. For Support Vector Machine (SVM) classification, we fully utilized parameters derived from harmonic analyses of the three vegetation indices (VIs: NDVI, NDWI, and SAVI) extracted from Sentinel-2A imagery. The harmonic analyses proved that higher-order sinusoid components produced better fitting to explain the trajectory of the VIs-the maximum adjusted R-2 was 95.23%-and the multiple VIs diversified the attributes for the classifications. Consequently, the higher-order harmonic components and the additional VIs increased the accuracy when used in SVM classification. The best performing classification was achieved with a composite of harmonic terms derived from the three VIs, yielding overall accuracy of 90.72%, Kappa index of 0.858, and user's accuracy for abandoned farmland of 93.40%. The proposed method here would greatly improve the process of detecting abandoned farmland, despite a relatively short observation period, and enable a rapid response to the occurrence of abandonment.