Feature Relevance Evaluation for Predicting Corn Crop Yield in Colombia’s Neotropical Region

Abstract

Creating precise predictive models for estimating corn crop yields accurately is crucial for informed decision-making in sustainable agriculture. While various approaches, including Fuzzy Logic, Association Rules, and Machine Learning, aim to achieve this, some encounter limitations due to the complexity and non-linearity of factor interactions. Although Machine Learning techniques can achieve high precision, incorporating multiple attributes may diminish it. This research focuses on identifying key regional factors affecting corn crop yields in Colombia, situated in the Neotropical zone. Utilizing climatological time series and yield records, a CRISP-DM-based methodology is employed, involving related work review, data cleaning, transformation, relevance evaluation using RReliefF algorithm, and performance verification of influential factors through prediction algorithms. Results reveal that solar radiation, precipitation, vapor pressure, and maximum and minimum temperatures significantly influence yield prediction, with relevance factors of 0.033, 0.032, 0.026, 0.022, and 0.021, respectively. Validating the performance of selected factors, two predictive models are implemented. The first, employing Artificial Neural Networks, yields RMSE of 0.1216 and 0.1403 with subset and all variables, respectively. The second, using Linear Regression, results in RMSE of 0.1417 and 0.1424 with subset and all variables, respectively. These findings underscore the importance of selected features as influential climatic factors in constructing highly accurate predictive models in the Neotropical zone.