Satellite

Quantum mechanics

Engineering

Aerospace engineering

Computer science

Spectroradiometer

Image resolution

Physics

Moderate-resolution imaging spectroradiometer

Optics

Geography

Data set

Environmental science

Artificial intelligence

Remote sensing

Temporal resolution

Reflectivity

Science of Remote Sensing

Generating crop type maps using satellite remote sensing requires robust data acquisition at both high spatial and temporal resolutions to resolve rapid phenological transition at the field scale. The increasing availability of freely-available, moderate-resolution satellite data such as the Landsat and Sentinel series of satellites offers an unprecedent opportunity for large-area crop type mapping. In this study, we evaluated the utility of Landsat (7&8), Sentinel-2 (A&B), Sentinel-1 (A&B) and the Moderate Resolution Imaging Spectroradiometer (MODIS) for mapping corn and soybean in the United States. We designed a series of classification experiments using these satellite data over a nationally distributed sample as input and the United States Department of Agriculture (USDA) Cropland Data Layer (CDL) as reference for training and accuracy assessment. A set of tests were performed with data from each satellite senor as input to derive the potential accuracy achievable by the satellite/sensor. In comparison, another set of tests were conducted with all data from all sensors as input to derive the combined accuracy as well as to evaluate the utility of each sensor, spectral band and acquisition date. Results showed that data from one satellite/sensor, either Landsat, Sentinel-2 or Sentinel-1, could achieve 94.8–96.8% accuracy, whereas the coarse-resolution MODIS produced about 92% accuracy, for both corn and soybean. Combing data from all sensors marginally improved the accuracy to 97.0% for both crops. Based on the criterion of deviation reduction in decision tree models, Landsat was identified as the most useful satellite/sensor for soybean classification, especially the two short-wave infrared bands, whereas Sentinel-2 was recognized as the most valuable satellite/sensor for corn classification, especially the red edge, near infrared and short-wave infrared bands. Optical data were always chosen over Synthetic Aperture Radar (SAR) data by the pixel-based supervised classification algorithm except in some persistently cloudy regions, although using SAR data alone can also achieve very high accuracy. The virtual constellation of Landsat and Sentinel-2 increased data revisit frequency to 4–7 days in the U.S. during June to September 2017. However, cloud and shadow reduced clear-view observations by half. Satellite data acquisitions in July were most critical for mapping corn whereas data in July and August were most important for mapping soybean. Our analysis suggested that, without the practical limitation of training data, current freely-available, moderate-resolution satellite data including Landsat, Sentinel-2, Sentinel-1 and MODIS, can achieve a potential accuracy of over 95% for national-scale crop type mapping over large industrial agricultural regions such as the United States. Expanding the spatial coverage and maintaining consistent acquisitions of Sentinel-1 data is a high priority to enable operational crop mapping and monitoring over large areas.

An evaluation of Landsat, Sentinel-2, Sentinel-1 and MODIS data for crop type mapping