USING FUZZY LOGIC TO PREDICT AND EVALUATE THE MAGNITUDE AND DISTRIBUTION OF PRECIPITATION ON RAINFED MAIZE AND SOYBEAN YIELDS IN NEBRASKA

作者：Rudnick D R; Sharma V; Meyer G E^*; Irmak S

来源：Transactions of the ASABE, 2015, 58(5): 1215-1229.

DOI：10.13031/trans.58.10831

• ANFIS
• GIS
• Maize
• Precipitation
• Rainfed
• Regional
• Soybean
• Yield prediction

Adaptive neuro-fuzzy inference system (ANFIS) models were developed to predict and evaluate the impact of temporal and spatial precipitation amounts on rainfed maize and soybean yields in Nebraska for individual years. Precipitation data from 1996 to 2012 were divided into three datasets: training, checking, and testing. The testing dataset comprised all counties with reported grain yields for the years 2000, 2005, and 2010. The statewide average seasonal (May 1 to Sept. 30) precipitation for 2000, 2005, and 2010 was 270, 358, and 482 mm, respectively, and the long-term (1996 to 2012) average was 361 mm. The ANFIS models were tested by adding individual cumulative monthly precipitation amounts as model inputs. ANFIS was unable to accurately predict maize and soybean yields with only May precipitation, with R-2 and root mean squared difference (RMSD) values, respectively, of 0.01 and 2.48 Mg ha(-1) for maize and 0.00 and 0.75 Mg ha(-1) for soybean. Model performance improved as monthly precipitation amounts were added to the maize and soybean ANFIS models, with the ANFIS model Y = f(May, June, July, Aug., Sept.) performing best for the pooled testing years (2000, 2005, and 2010), with R-2 and RMSD values, respectively, of 0.57 and 1.64 Mg ha(-1) for maize and 0.47 and 0.56 Mg ha(-1) for soybean. Longitude was added as a model input to implicitly account for differences in climatic variables (e.g., solar radiation) that impact grain yield across Nebraska. The inclusion of longitude into the time-step ANFIS models improved the performance for both individual and pooled testing years, with greater improvement for maize than soybean. The potential effects of dormant (Oct. 1 to April 30), seasonal, and annual (Oct. 1 to Sept. 30) precipitation amounts on grain yields were also investigated The best performing model of the pooled testing data for both maize and soybean was Y = f(Lng, Annual), with R-2 and RMSD values, respectively, of 0.75 and 1.22 Mg ha(-1) for maize and 0.67 and 0.44 Mg ha(-1) for soybean. Furthermore, prior to the growing season, the Y = f(Lng, Dormant) model explained 65% and 61% of the variability in maize and soybean grain yield, respectively, for the pooled testing data. This study demonstrated that ANFIS can be used to predict maize and soybean grain yields in Nebraska with reasonable accuracy from precipitation data.

• 出版日期2015-10