Characterizing major agricultural land change trends in the Western Corn Belt

Abstract

In this study we developed annual corn/soybean maps for the Western Corn Belt within the United States using multi-temporal MODIS NDVI products from 2001 to 2015 to support long-term cropland change analysis. Based on the availability of training data (cropland data layer from the USDA-NASS), we designed a cross-validation scheme for 2006–2015 MODIS data to examine the spatial generalization capability of a neural network classifier. Training data points were derived from a three-state sub-region consisting of North Dakota, Nebraska, and Iowa. Trained neural networks were applied to the testing sub-region (South Dakota, Kansas, Minnesota, and Missouri) to generate corn/soybean maps. Using a default threshold value (neural network output signal⩾0.5), the neural networks performed well for South Dakota and Minnesota. Overall accuracy was higher than 80% (kappa>0.55) for all testing years from 2006 to 2015. However, we observed high variation of classification performance for Kansas (overall accuracy: 0.71–0.82) and Missouri (overall accuracy: 0.65–0.77) for various testing years. We developed a threshold-moving method that decreases/increases threshold values of neural network output signals to match MODIS-derived corn/soybean acreage with the NASS acreage statistics. Over 70% of testing states and years showed improved classification performance compared to the use of a default 0.5 threshold. The largest improvement of kappa value was about 0.08. This threshold-moving method was used to generate MODIS-based annual corn/soybean map products for 2001–2015. A non-parametric Mann-Kendall test was then used to identify areas that showed significant (p<0.05) upward/downward trends. Areas showing fast increase of corn/soybean intensities were mainly located in North Dakota, South Dakota, and the west portion of Minnesota. The highest annual increase rate for a 5-km moving window was about 6.8%.