High-Dimensional Satellite Image Compositing and Statisticsfor Enhanced Irrigated Crop Mapping

Abstract

Accurate irrigated area maps remain difficult to generate, as smallholder irrigation schemes often escape detection. Efforts to map smallholder irrigation have often relied on complex classification models fitted to temporal image stacks. The use of high-dimensional geometric median composites (geomedians) and high-dimensional statistics of time-series may simplify classification models and enhance accuracy. High-dimensional statistics for temporal variation, such as the spectral median absolute deviation, indicate spectral variability within a period contributing to a geomedian. The Ord River Irrigation Area was used to validate Digital Earth Australia’s annual geomedian and temporal variation products. Geomedian composites and the spectral median absolute deviation were then calculated on Sentinel-2 images for three smallholder irrigation schemes in Matabeleland, Zimbabwe, none of which were classified as areas equipped for irrigation in AQUASTAT’s Global Map of Irrigated Areas. Supervised random forest classification was applied to all sites. For the three Matabeleland sites, the average Kappa coefficient was 0.87 and overall accuracy was 95.9% on validation data. This compared with 0.12 and 77.2%, respectively, for the Food and Agriculture Organisation’s Water Productivity through Open access of Remotely sensed derived data (WaPOR) land use classification map. The spectral median absolute deviation was ranked among the most important variables across all models based on mean decrease in accuracy. Change detection capacity also means the spectral median absolute deviation has some advantages for cropland mapping over indices such as the Normalized Difference Vegetation Index. The method demonstrated shows potential to be deployed across countries and regions where smallholder irrigation schemes account for large proportions of irrigated area.