Evaluation of Three Feature Dimension Reduction Techniques for Machine Learning-Based Crop Yield Prediction Models.

Area of Science:

• Agricultural Science
• Data Science
• Machine Learning

Background:

• Machine learning (ML) is crucial for crop yield forecasting, but identifying critical features remains a challenge.
• Existing methods like feature selection (FS) and feature extraction (FX) are used independently.
• No prior research has compared FS, FX, and their combined approach (FSX) for crop yield prediction.

Purpose of the Study:

• To propose and evaluate a framework comparing FS, FX, and FSX for crop yield forecasting.
• To investigate the performance benefits of combining feature selection and feature extraction.
• To establish a baseline using all features (All-F) for performance comparison.

Main Methods:

• Developed 21 rice yield forecasting models using ML algorithms (Linear, SVM, Tree, ANN, Ensemble).
• Applied feature selection (FS), feature extraction (FX), and a combined FSX approach.
• Utilized Vegetation Condition Index (VCI) and Temperature Condition Index (TCI) data for eight sub-regions in Vietnam.

Main Results:

• FSX-based models demonstrated superior performance, achieving the best results in 18 out of 21 models.
• FSX, FS, and FX models improved upon the All-F baseline, reducing RMSE by an average of 21% (up to 60%).
• Ensemble (13 models) and Tree (6 models) were the most effective ML algorithms for yield prediction.

Conclusions:

• The combination of feature selection and feature extraction (FSX) significantly enhances ML-based crop yield forecasting accuracy.
• FSX leverages the strengths of both FS and FX, offering substantial improvements over individual methods and using all features.
• Ensemble and Tree algorithms, when combined with FSX, show great promise for accurate and reliable crop yield prediction.