Scalable Parcel-Based Crop Identification Scheme Using Sentinel-2 Data Time-Series for the Monitoring of the Common Agricultural Policy

• Published in: Remote sensing (Basel, Switzerland) Vol. 10; no. 6; p. 911
• Main Authors: Sitokonstantinou, Vasileios, Papoutsis, Ioannis, Kontoes, Charalampos, Arnal, Alberto, Andrés, Ana Pilar, Zurbano, José Angel
• Format: Journal Article
• Language: English
• Published: MDPI AG 01-06-2018
• Subjects: common agricultural policy; crop identification; feature importance; Landsat-8 OLI; multispectral image time-series; random forest; Sentinel-2 MSI; support vector machines
• DOI: 10.3390/rs10060911
• EISSN: 2072-4292

This work investigates a Sentinel-2 based crop identification methodology for the monitoring of the Common Agricultural Policy's (CAP) Cross Compliance (CC) and Greening obligations. In this regard, we implemented and evaluated a parcel-based supervised classification scheme to produce accurate crop type mapping in a smallholder agricultural zone in Navarra, Spain. The scheme makes use of supervised classifiers Support Vector Machines (SVMs) and Random Forest (RF) to discriminate among the various crop types, based on a large variable space of Sentinel-2 imagery and Vegetation Index (VI) time-series. The classifiers are separately applied at three different levels of crop nomenclature hierarchy, comparing their performance with respect to accuracy and execution time. SVM provides optimal performance and proves significantly superior to RF for the lowest level of the nomenclature, resulting in 0.87 Cohen's kappa coefficient. Experiments were carried out to assess the importance of input variables, where top contributors are the Near Infrared (NIR), vegetation red-edge, and Short-Wave Infrared (SWIR) multispectral bands, and the Normalized Difference Vegetation (NDVI) and Plant Senescence Reflectance (PSRI) indices, sensed during advanced crop phenology stages. The scheme is finally applied to a Lansat-8 OLI based equivalent variable space, offering 0.70 Cohen's kappa coefficient for the SVM classification, highlighting the superior performance of Sentinel-2 for this type of application. This is credited to Sentinel-2's spatial, spectral, and temporal characteristics.