Large-scale crop monitoring and yield estimation are important for both scientific research and practical applications. Satellite remote sensing provides an effective means for regional and global cropland monitoring, particularly in data-sparse regions that lack reliable ground observations and reporting. The conventional approach of using visible and near-infrared based vegetation index (VI) observations has prevailed for decades since the onset of the global satellite era. However, other satellite data encompass diverse spectral ranges that may contain complementary information on crop growth and yield, but have been largely understudied and underused. Here we conducted one of the first attempts at synergizing multiple satellite data spanning a diverse spectral range, including visible, near-infrared, thermal and microwave, into one framework to estimate crop yield for the U.S. Corn Belt, one of the world's most important food baskets. Overall, using satellite data from various spectral bands significantly improves regional crop yield predictions. The additional use of ancillary climate data (e.g. precipitation and temperature) further improves model skill, in part because the crop reproductive stage related to harvest index is highly sensitive to environmental stresses but they are not fully captured by the satellite data used in our study. We conclude that using satellite data across various spectral ranges can improve monitoring of large-scale crop growth and yield beyond what can be achieved from individual sensors. These results also inform the synergistic use and development of current and next generation satellite missions, including NASA ECOSTRESS, SMAP, and OCO-2, for agricultural applications.

Satellite-derived land cover maps play an important role in many applications, including monitoring of smallholder-dominated agricultural landscapes. New cloud-based computing platforms and satellite sensors offer opportunities for generating land cover maps designed to meet the spatial and temporal requirements of specific applications. Such maps can be a significant improvement compared to existing products, which tend to be coarser than 300 m, are often not representative of areas with fast-paced land use change, and have a fixed set of cover classes. Here, we present two approaches for land cover classification using the Landsat archive within Google Earth Engine. Random forest classification was performed with (1) season-based composites, where median values of individual bands and vegetation indices were generated from four years for each of four seasons, and (2) metric-based composites, where different quantiles were computed for the entire four-year period. These approaches were tested for six land cover types spanning over 18,000 locations in Zambia, with ground “truth” determined by visual inspection of high-resolution imagery from Google Earth. The methods were trained on 30% of these points and tested on the remaining 70%, and results were also compared with existing land cover products. Overall accuracies of about 89% were achieved for the season- and metric-based approaches for individual classes, with 93%and 94% accuracy for distinguishing cropland from non-cropland. For the latter task, the existing Globeland30 dataset based on Landsat had much lower accuracies (around 77% on average), as did existing cover maps at coarser resolutions. Overall, the results support the use of either season or metric-based classification approaches. Both produce better results than those obtained from previous classifiers, which supports a general paradigm shift away from dependence on standard static products and towards custom generation of on-demand cover maps designed to fulfill the needs of each specific application.

One of the greatest challenges in monitoring food security is to provide reliable crop yield information that is temporally consistent and spatially scalable. An ideal yield dataset would not only extend globally and across multiple years, but would also have enough spatial granularity to characterize productivity at the field and subfield level. Rapid increases in satellite data acquisition and platforms such as Google Earth Engine that can efficiently access and process vast archives of new and historical data offer an opportunity to map yields globally, but require efficient and robust algorithms to combine various data streams into yield estimates. We recently introduced a Scalable satellite-based Crop Yield Mapper (SCYM) that combines crop models simulations with imagery and weather data to generate 30 m resolution yield estimates without the need for ground calibration. In this study, we tested new large-scale implementations of SCYM, focusing on three regions with varying crops, field sizes and landscape heterogeneity: maize in the U.S. corn belt (390,000 km²), maize in Southern Zambia (86,000 km²), and wheat in northern India (450,000 km²). As a benchmark, we also tested a simpler empirical approach (PEAKVI) that relates yield to the peak value of a time series of spatially aggregated vegetation indices, similar to methods used in current operational monitoring. Both SCYM and PEAKVI were applied to data from all Landsat's sensors and MODIS for more than a decade in each region, and evaluated against ground-based estimates at the finest available administrative level (e.g., counties in the U.S.). We found consistently high correlations (R² ≥ 0.5) between the spatial pattern of ground- and satellite-based estimates in both U.S. maize and India wheat, with small differences between methods and source of satellite data. In the U.S., SCYM outperformed PEAKVI in tracking temporal yield variations, likely owing to its explicit consideration of weather. In India, both methods failed to track temporal yield changes, with various possible explanations discussed. In Zambia, the PEAKVI approach applied to MODIS tracked yield variations much better (R² > 0.5) than any other yield estimate, likely because the frequent cloud cover in this region confounds the other approaches. Overall, this study demonstrates successful approaches to yield estimation in each region, and illustrates the importance of distinguishing between accuracy for spatial and temporal variation. The 30 m resolution of Landsat-based SCYM does not appear to offer large benefits for tracking aggregate yields, but enables finer scale analyses than possible with the other approaches.