A Data-Efficient Rice Yield Classification Framework Combining Hybrid PCA-RFE and Regularized Random Forest

Abstrak

Accurate rice yield prediction by remote sensing data and machine learning is still a big challenge in limited resources of field survey where the sample size is often very small. This study tackles the core challenge of small sample size (n=39) in rice yield classification by proposing a hybrid feature engineering framework that combines Principal Component Analysis (PCA) and Recursive Feature Elimination (RFE) in a regularized Random Forest (RF) classifier. The methodology is based on 30 spectral features from multi-temporal Sentinel-2A imagery (15 spectral bands i.e. Bands 2, 3, 4, 5 and 8 for three acquisition dates and 15 vegetation indices). PCA revealed three principal components that explained approximately 90% of the total variance and RFE selected the five most discriminative spectral bands (SAMPLE_14, SAMPLE_15, SAMPLE_17, SAMPLE_23 and SAMPLE_26). These were fused in a compact eight-dimensional hybrid feature space. Model evaluation was conducted using repeated stratified cross-validation (5-fold x 10 repeats) and an independent test set (holdout 30%). The average cross-validation accuracy of the hybrid model was 83.36%, ROC-AUC 0.9304, independent test accuracy 91.67% and Cohen's Kappa 0.8333. There was no over-fitting of the training-validation gap to 0.10 in the learning curve. The optimal classification threshold was identified as 0.5373 in the Youden Index optimization. The RFE-selected spectral bands contributed the largest share (76%) in the feature importance analysis, and the PCA components provided a complementary 24%, confirming the synergistic value of unsupervised extraction and supervised selection. However, spatial uncertainty mapping revealed limitations to the overall extrapolation, with near-maximum values (0.999) in unsampled areas, emphasizing the need for targeted field verification in high-uncertainty zones. The study provides a replicable methodological blueprint for crop yield classification under very limited data conditions and provides practical guidance for agricultural monitoring in developing countries with logistical and financial constraints.Received: 2025-09-15 Revised: 2026-06-26 Accepted: 2026-07-28 Published: 2026-08-05

Kata kunci

• Geography
• planning
• and development
• PCA
• RFE
• Random Forest
• Rice Yield Classification
• Spatial Uncertainty
• Computer in Earth Science (GIS and Remote Sensing)