Moisture Prediction from Simple Micrometeorological Data

Abstract

Four linear regression methods and a generalized regression neural network (GRNN) were evaluated for estimation of moisture occurrence and duration at the flag leaf level of wheat. Moisture on a flat-plate resistance sensor was predicted by time, temperature, relative humidity, wind speed, solar radiation, and precipitation provided by an automated weather station. Dew onset was estimated by a classification regression tree model. The models were developed using micrometeorological data measured from 1993 to 1995 and tested on data from 1996 and 1997. The GRNN outperformed the linear regression methods in predicting moisture occurrence with and without dew estimation as well as in predicting duration of moisture periods. Average absolute error for prediction of moisture occurrence by GRNN was at least 31% smaller than that obtained by the linear regression methods. Moreover, the GRNN correctly predicted 92.7% of the moisture duration periods critical to disease development in the test data, while the best linear method correctly predicted only 86.6% for the same data. Temporal error distribution in prediction of moisture periods was more highly concentrated around the correct value for the GRNN than linear regression methods. Neural network technology is a promising tool for reasonably precise and accurate moisture monitoring in plant disease management.