Accuracy of Rapid Crop Cover Maps of Conterminous United States for 2008 - 2016
Dates
Publication Date
2018-04-12
Start Date
2008
End Date
2016
Citation
Dahal, D., Wylie, B.K., and Howard, D., 2018, Accuracy of rapid crop cover maps of conterminous United States for 2008 - 2016: U.S. Geological Survey data release, https://doi.org/10.5066/F7B27TG8.
Summary
Spatially accurate annual crop cover maps are an important component to various planning and research applications; however, the importance of these maps varies significantly with the timing of their availability. Utilizing a previously developed crop classification model (CCM), which was used to generate historical annual crop cover maps (classifying nine major crops: corn, cotton, sorghum, soybeans, spring wheat, winter wheat, alfalfa, other hay/non alfalfa, fallow/idle cropland, and ‘other’ as one class for remaining crops), we hypothesized that such crop cover maps could be generated in near real time (NRT). The CCM was trained on 14 temporal and 15 static geospatial datasets, known as predictor variables, and the National Agricultural [...]
Summary
Spatially accurate annual crop cover maps are an important component to various planning and research applications; however, the importance of these maps varies significantly with the timing of their availability. Utilizing a previously developed crop classification model (CCM), which was used to generate historical annual crop cover maps (classifying nine major crops: corn, cotton, sorghum, soybeans, spring wheat, winter wheat, alfalfa, other hay/non alfalfa, fallow/idle cropland, and ‘other’ as one class for remaining crops), we hypothesized that such crop cover maps could be generated in near real time (NRT). The CCM was trained on 14 temporal and 15 static geospatial datasets, known as predictor variables, and the National Agricultural Statistics Service (NASS) Cropland Data Layers (CDL) was used as the dependent variable. We were able to generate a NRT crop cover map by the first day of September through a process of incrementally removing weekly and monthly data from the CCM and comparing the subsequent map results with the original maps and NASS CDLs. Initially, our NRT results revealed training error of 1.4% and test error of 8.3%, as compared to 1.0% and 7.6%, respectively for the original CCM. Through the implementation of a new ‘two-mapping model’ approach, we were able to substantially improve the results of the NRT crop cover model. We divided the NRT model into one ‘crop type model’ to handle the classification of the nine specific crops and a second, binary model to classify crops as presence or absence of the ‘other’ crop. Under the two-mapping model approach, the training errors were 0.8% and 1.5% for the crop type and binary model, respectively, while test errors were 5.5% and 6.4% for crop type and binary model, respectively. With overall mapping accuracies for annual maps reaching in 60s to 70s percent, this approach shows strong potential for generating crop type maps of current year in September.
