Crop Yield Forecasting from Remotely Sensed Aerial Images with Self-Organizing Maps

Abstract

Crop yield varies spatially and temporally due to several limiting factors, including soil type, nutrient availability, water availability, weather, and crop varieties. It is difficult to quantify the severity of these limiting factors on crop yield as well as the spatial variation of crop yield before harvest. Remotely sensed aerial images can be used to improve estimates of mid-season crop yields. However, the success of such approaches is typically dependent on the image classification technique used. In this article, we develop an automated self-organizing map (SOM) neural image classification technique and evaluate it with respect to the widely used ISODATA clustering method using visual assessment of preservation of image texture. We also develop hybrid SOM-back-propagation neural network (BPNN) models to predict crop yield from the Oakes Irrigation Test Area (OITA) research sites at Oakes, North Dakota. We used the SOM image classification technique to classify the composite visible and near-infrared (NIR) band aerial images from several parts of the OITA research site. SOM cluster centroids from individual grid images were used as inputs to the BPNN yield prediction models created for the years 1997, 1998, 1999, 2000, and 2001. The models using individual red, green, and NIR band digital data provided average corn yield prediction accuracies of more than 80% for every year, except for the NIR model in 1998 (45%). Individual year models were validated with individual band training and testing datasets from the other four years. Models using the green band dataset provided better validation results. Average corn yield prediction accuracies ranged from 55% to a maximum of 96%. Average corn yield prediction accuracies of validated models using the red band and NIR band datasets ranged from 33% to a maximum of 94% and from 36% to a maximum of 95%, respectively. This article establishes the use of hybrid SOM-BPNN modeling techniques to predict corn yield from aerial images at mid-crop season.