Enhancing Predictive Accuracy in Agricultural Land Suitability with Machine Learning Using Feature Selection and Data Balancing

Abstract

Agricultural land assessment and soil quality evaluation are critical for sustainable agricultural practices. However, traditional methods are often inefficient and inaccurate in handling large and complex data sets. This study aims to prove that machine learning can enhance the accuracy of datasets by implementing advanced techniques like data preprocessing and feature selection. Three machine learning models which is Random Forest, Decision Tree, and K-Nearest Neighbors (KNN) were applied to a Kaggle Agricultural Land Suitability and Soil Quality dataset to predict land suitability more effectively. The evaluation results show a significant improvement in performance metrics after preprocessing. For instance, the Random Forest model's AUC increased from 0.491 to 0.838, and its MCC improved from -0.013 to 0.476. Similarly, the Decision Tree and KNN models achieved substantial performance improvement, showing the value of machine learning in enhancing prediction accuracy. This study provides a clear framework for using machine learning to make data-driven decision-making in agriculture, which promote efficient land use and sustainable farming practices.