A Multi-Method Diagnostic Framework for Detecting Feature–Label Independence: A Case Study on Crop Health Classification

Abstract

We worked on a machine learning project to predict crop health from agronomic data. We used a real-world dataset of 212,019 field records with 32 attributes like temperature, humidity, rainfall and soil properties. Our goal was to see how well four machine learning models. Logistic Regression, Random Forest, Histogram-based Gradient Boosting and XGBoost. Could predict crop health. We tested these models using a method called stratified 5fold crossvalidation and then fine-tuned the best model. The dataset only had a healthy or unhealthy label for each crop so we created a new label with five levels of severity. Surprisingly all four models performed about well as guessing the most common answer, which was not very good. We did checks to see if the data or models were the problem and it seems the data just doesn't have enough information to predict crop health. We think it's essential to share results, like this so other don't waste time on similar projects.