Model for Predicting Maize Crop Yield on Small Farms Using Clusterwise Linear Regression and GRASP

Abstract

Planting a crop involves several key steps: resource assessment, crop selection, crop rotation, planting schedules, soil preparation, planting, care, and harvesting of crops. In this context, estimating the productivity of a crop based on available information, such as expected climatic conditions and agricultural practices, helps farmers reduce the uncertainty of their investment. In Colombia, maize is the fourth most important crop in the country. Significant efforts are required to improve productivity in traditional and technified production systems. In this sense, this research proposes and evaluates an approach called Clusterwise Linear Regression (CLR) to predict the crop maize yield in small farms, considering data on climate, soil, fertilization, and management practices, among others. To develop the CLR model, we conducted the following steps: data collection and preparation, clustering using k-means, cluster optimization with Greedy Random Adaptive Search Procedure (GRASP), and performance evaluation. The cluster optimization process allows the identification of clusters with similar characteristics and generates multiple linear regression models with mixed variables that explain the yield of the farms on each cluster. The Simulated Multiple Start Annealing (MSSA) metaheuristics were also evaluated, but the results of GRASP were the best. The results indicate that the proposed CLR approach is more effective than the linear and nonlinear algorithms mentioned in the literature, such as multiple lasso linear regression, random forests, XGBoost, and support vector machines. These algorithms achieved an accuracy of 70%. However, with the new CLR model, a significantly improved accuracy of 87% was achieved with test data. The clusters’ studies revealed key factors affecting crop yield, such as fertilization, drainage, and soil type. This transparency is a benefit over black-box models, which can be harder to interpret. This advancement can allow farmers to make better decisions about the management of their crops.

Keywords

• clusterwise
• linear regression
• metaheuristics
• clustering
• site-specific agriculture