Intra-annual reflectance composites from Sentinel-2 and Landsat for national-scale crop and land cover mapping

Abstract

Many applications that target dynamic land surface processes require a temporal observation frequency that is not easily satisfied using data from a single optical sensor. Sentinel-2 and Landsat provide observations of similar nature and offer the opportunity to combine both data sources to increase time-series temporal frequency at high spatial resolution. Multi-sensor image compositing is one way for performing pixel-level data integration and has many advantages for processing frameworks, especially if analyses over larger areas are targeted. Our compositing approach is optimized for narrow temporal-intervals and allows the derivation of time-series of consistent reflectance composites that capture field level phenologies. We processed more than a year's worth of imagery acquired by Sentinel-2A MSI and Landsat-8 OLI as available from the NASA Harmonized Landsat-Sentinel dataset. We used all data acquired over Germany and integrated observations into composites for three defined temporal intervals (10-day, monthly and seasonal). Our processing approach includes generation of proxy values for OLI in the MSI red edge bands and temporal gap filling on the 10-day time-series. We then derive a national scale crop type and land cover map and compare our results to spatially explicit agricultural reference data available for three federal states and to the results of a recent agricultural census for the entire country. The resulting map successfully captures the crop type distribution across Germany at 30 m resolution and achieves 81% overall accuracy for 12 classes in three states for which reference data was available. The mapping performance for most classes was highest for the 10-day composites and many classes are discriminated with class specific accuracies >80%. For several crops, such as cereals, maize and rapeseed our mapped acreages compare very well with the official census data with average differences between mapped and census area of 11%, 2% and 3%, respectively. Other classes (grapevine and forest classes) perform slightly less well, likely, because the available reference data does not fully capture the variability of these classes across Germany. The inclusion of the red edge bands slightly improved overall accuracies in all cases and improved class specific accuracies for most crop classes. Similarly, our gap filling procedure led to improved mapping accuracies when compared to nongap-filled 10-day features. Overall, our results demonstrate the valuable potential of approaches that utilize data from Sentinel-2 and Landsat which allows for detailed assessments of agricultural and other land-uses over large areas.