Reducing the Impact of Cloud Cover on Operational Crop Inventories

Large area cropland mapping using remote sensing has matured due to recent advances in algorithm development and the wide availability of moderate resolution ( 30 m) optical data. Annual crop type mapping requires 2-3 "reasonably cloud-free" scenes per season during key growth stages, although persistent & pervasive cloud cover in many major agricultural areas/times has prevented sufficient information from being reliably collected. Synthetic aperture radar (SAR), such as that freely & openly available from ESA Sentinel-1, provides a way of obtaining key crop information regardless of cloud coverage, although in most cases, single frequency SAR must be used complementarily with optical data to meet standard classification accuracy thresholds (e.g. 85% overall accuracy). Further, power limitations of radar satellites often mean that systematic coverage of large areas is not possible, so guidelines on when, where, and how frequently to acquire SAR for agricultural monitoring are needed. Building on efforts to quantify seasonal cloud cover probability and to define Earth observation requirements for the GEO Global Agricultural Monitoring (GEOGLAM) initiative, this work used a synthetic cloud approach in combining radar and optical satellite data to determine optimal timing of SAR acquisitions, to define "reasonably cloud-free" thresholds for optical data, and to help guide the production of multi-date composites for operational use, over different agricultural landscapes. Preliminary results show that without the use of radar, three optical images with cloud coverages of less than 15% are required at early, middle, and late growth stages to meet the accuracy threshold, but that cloud coverage as high as 57% is acceptable when appropriately timed SAR data are used within the classification. The use of SAR also increases the robustness of classification, with the chances of obtaining acceptable classification accuracy regardless of cloud cover increasing greatly when one to three SAR images are used.