Assessing the capabilities of high-resolution spectral, altimetric, and textural descriptors for mapping the status of citrus parcels

Abstract

Agricultural land abandonment is an increasing phenomenon around the world with relevant environmental and socio-economic implications. In the European Union about 11 % of agricultural land is at high risk of abandonment. The Comunitat Valenciana region (Spain) is the most important citrus producer in Europe suffering from this problem. Identifying the status of citrus crops at the parcel level is essential for policymakers in agriculture. This work assessed the use of WorldView-3 data, Very High-Resolution Airborne Images, and Structure from Motion point clouds to identify the status of citrus parcels using two machine learning algorithms: Random Forest and Support Vector Machines. Different analyses involving combinations of the three data sources were carried out to assess the impact on classification accuracy. The results showed the high potential of airborne imagery (OA ≈ 0.967) and WorldView-3 (OA ≈ 0.936) to detect parcel status using a single image. The SfM data showed a lower potential (OA ≈ 0.825). Adding SfM point cloud to the multispectral information produced small improvements (0.4–2.0 %) in classification accuracy. The class separability analysis showed the importance of WV-3 SWIR bands to detect abandoned parcels as they produce more spectral separability over the productive parcels in the 1570 nm – 2330 nm spectrum. The results also show the importance of GLCM texture features extracted from sub-metric images due to their ability to model spatial planting patterns typical of fruit crops.