Near Real-Time Corn and Soybean Mapping at Field-Scale By Blending Crop Phenometrics With Growth Magnitude From Multiple Temporal and Spatial Satellite Observations

Timely and accurate crop mapping is essential for predicting crop production, estimating water use, and informing market forecasts. However, operational crop maps are typically accessible more than four months subsequent to harvest, rather than in real-time or near real-time (NRT). Recently, in-season crop mapping has emerged by leveraging rich satellite data sources at various scales in the United States (US) Corn Belt – a prominent food-producing agricultural region dominated by corn and soybeans. However, challenges persist due to inadequate clear-sky satellite observations and the absence of field-scale in-season crop phenometrics. Recognizing that SWIR (shortwave infrared reflectance) is able to reflect the asynchronous temporal variations in plant canopy water contents and that combining phenological shift and growth magnitude can enhance the classification of crop types, this study developed two canopy Greenness and Water (GW) content indices that are GW-I, which is a ratio of the kernel NDVI (normalized difference vegetation index) to SWIR to distinguish phenological shift of different crops, and GW-II, which is a product of kernel NDVI and SWIR to separate growth magnitude of different crops. To reconstruct gap-free field-scale GW-I and GW-II time series, historical and timely available multi-scale satellite observations, including Harmonized Landsat and Sentinel-2 (HLS), Visible Infrared Imaging Radiometer Suite (VIIRS), and Advanced Baseline Imager (ABI), are dynamically fused every week. The potential future GW-I and GW-II values are further predicted using a recently developed algorithm of Spatiotemporal Shape Matching Model (SSMM) and combined with the timely available time series for retrieving NRT phenometrics (greenup onset, mid-date of greenup phase, and maturity onset) every week during the crop greenup phase. Multiple Gaussian mixture models are used to independently estimate the weekly probability of corn and soybean types using three NRT crop phenometrics and the latest (≤3 days' latency) GW-II. Finally, the corn and soybean probabilities (estimated from GW-I phenometrics and GW-II crop growth magnitude together) are integrated to produce NRT corn and soybean mapping every week during the early growing season. The accuracy of NRT corn and soybean mapping is evaluated using the Cropland Data Layer (CDL). The result shows that our method can map corn and soybean in diverse croplands across the US Corn Belt with an overall accuracy of ∼90 % at a relatively early date (late July), although the local heterogeneity of agricultural landscapes potentially impacts the accuracy during the early stages. These findings underscore the feasibility of applying the developed method to produce near real-time corn and soybean mapping not only across the US Corn Belt but also in other countries and diverse agricultural regions.