A Comparative Study of Machine Learning Models For Crop Disease Prediction Under Data Scarcity

Studi Komparatif Model Machine Learning untuk Prediksi Penyakit Tanaman dalam Kondisi Keterbatasan Data

Abstract

In many agricultural settings, particularly in developing regions, access to large, high-quality labeled datasets remains a persistent constraint. This study examines whether reliable crop-disease classification can be obtained when data are scarce, and whether the choice of learning algorithm—rather than the volume of data—is the dominant factor governing accuracy. Using a controlled simulation of 150 samples described by three environmental parameters (temperature, humidity, and soil pH) across eight leaf-disease classes, we train and evaluate five classifiers: K-Nearest Neighbors (KNN), Decision Tree, Random Forest, Support Vector Machine (SVM), and Logistic Regression. Each class is generated from characteristic, overlapping environmental ranges so that a learnable but non-trivial signal exists. Models are assessed on a stratified 30% hold-out set using accuracy, macro precision, recall, and F1-score, with Principal Component Analysis (PCA) used to visualize decision boundaries. At a fixed and small sample size, accuracy varied widely across models (from 51.11% for the Decision Tree to 68.89% for Logistic Regression), all substantially above the 12.50% random baseline for eight balanced classes. Among the three tree-family and instance-based models that motivated the study, Random Forest was the strongest (64.44%). These results support the central premise that, in low-resource conditions, principled model selection contributes more to predictive performance than simply enlarging the dataset.