A model to predict ordinal suitability using sparse and uncertain data

Abstract

We describe the development of the algorithms that comprise the Spatial Decision Support System (SDSS) CaNaSTA (Crop Niche Selection in Tropical Agriculture). The system was designed to assist farmers and agricultural advisors in the tropics to make crop suitability decisions. These decisions are frequently made in highly diverse biophysical and socioeconomic environments and must often rely on sparse datasets. The field trial datasets that provide a knowledge base for SDSS such as this are characterised by ordinal response variables. Our approach has been to apply Bayes’ formula as a prediction model. This paper does not describe the entire CaNaSTA system, but rather concentrates on the algorithm of the central prediction model. The algorithm is tested using a simulated dataset to compare results with ordinal regression, and to test the stability of the model with increasingly sparse calibration data. For all but the richest input datasets it outperforms ordinal regression, as determined using Cohen’s weighted kappa. The model also performs well with sparse datasets. Whilst this is not as conclusive as testing with real world data, the results are encouraging.