PREDICTIVE ABILITY OF NEURAL NETWORKS FOR SITE-SPECIFIC YIELD ESTIMATION^*

Scott T. Drummond, Computer Specialist
Kenneth A. Sudduth, Agricultural Engineer
USDA Agricultural Research Service
Cropping Systems and Water Quality Research Unit
Columbia, Missouri, USA

Anupam Joshi, Assistant Professor
Department of Computer Science & Electrical Engineering
University of Maryland, Baltimore County
Baltimore, Maryland, USA

ABSTRACT

Precision farming attempts to improve cropping efficiency by variable application of crop treatments such as fertilizers, pesticides, etc., on a point-by-point basis within fields. Therefore, a more complete understanding of the relationships between yield and soil and site properties is of critical importance to precision farming. A necessary first step in this process is the search for techniques able to identify functional relationships between measured soil and site characteristics and crop yield. Ten site-years of crop yields and corresponding site and soil characteristics were studied. A variety of supervised feed-forward neural networks were investigated in an attempt to identify methods able to relate soil properties and crop yields on a point-by-point basis, within individual site-years. To avoid overfitting, evaluations were based on generalization ability, using a 5-fold cross-validation technique. The results for neural methods were compared with results for linear and nonlinear, non-parametric methods. In terms of generalization ability, the neural techniques consistently outperformed all other methods, providing minimal errors in each site-year. However, in site-years with relatively fewer observations, and in site-years where water stress was not apparent, the improvements achieved by neural networks over linear methods were minimal. A second phase of the experiment involved estimation of crop yield across multiple site-years by including climatological data. The ten site-years of data were appended with climatological variables, and generalization errors were computed. The results showed that significant overfitting had occurred, and indicated that a much larger number of climatologically unique site-years would be required for this type of analysis to be successful.

Research Publications

Home | People | Projects | Publications | Upcoming Events | Links | Partners