Using Neural Network Classifier Approach for Statistically Forecasting Extreme Corn Yield Losses in Eastern United States. Issue 10 (27th October 2018)

• Record Type: Journal Article
• Authors: Mathieu, J. A.; Aires, F.
• Is Part Of: Earth and space science. Volume 5:Issue 10(2018)
• Page Start: 622
• Page End: 639
• Publication Date: 2018-10-27
• Subjects: statistical -- forecasting -- extreme events -- agriculture; Space sciences -- Periodicals; Geophysics -- Periodicals; 500.5
• DOI: 10.1029/2017EA000343
• Languages: English
• ISSNs: 2333-5084

Abstract

Abstract : This paper presents a statistical method for forecasting extreme corn yield losses caused by weather extremes. A neural network classifier approach is tested over the Eastern United States (time series of 35 years) to detect extreme yield losses for corn from weather‐related information. We first developed a methodology to rank a series of climate‐based predictors according to the accuracy with which they classify extreme from nonextreme yield losses. The classification methodology is adapted in order to be trained with a limited number of extreme cases. Using four weather predictors—the average temperature in July and August, and the SPEI (Standardized Precipitation‐Evapotranspiration Index) in June and July—71% of the extreme cases are well classified by this statistical model. Furthermore, the neural network output represents a good yield severity index and can provide an early quantitative warning for extreme yield anomalies. Plain Language Summary: The impact of weather on agriculture is very important and requires building models that describe the links and effects between weather and crop growth. These models are often efficient but they have great difficulty predicting very low crop yields. These situations refer to extreme yield losses and require other methodologies. Indeed, existing models are not efficient. Besides, very few data are available for these extreme events, which makes models difficult to calibrate. Moreover, the choice of good weather predictors is critical for the quality of the forecast. Our methodology identifies the weather indicators that best discriminate between extreme and nonextreme yield losses. We compare the ability of a variety of weather indicators to represent extreme yields: these indicators encompass single climate variables, water supply indices, but also agroclimatic indices calculated for a month, a year, or a growing season. The goal of the statistical modeling is not to estimate a yield but instead to define a yield severity index that would indicate extreme yield losses. It is extremely difficult to predict precisely in July or August the extreme losses. That is why we are searching for a probability of extreme cases rather than a strict prevision of the yield. Key Points: A neural network classifier forecasts extreme corn yield loss probability using four simple and monthly weather predictors Among the Eastern United States regions, 71% of yield loss extremes are well detected, with less than 15% of nonextremes misclassified as extremes The neural network provides a yield loss severity index that can help anticipating, in July, strong production losses