Enhancing Crop Yield Prediction Utilizing Machine Learning on Satellite-Based Vegetation Health Indices.

Area of Science:

• Agricultural Science
• Remote Sensing
• Machine Learning

Background:

• Accurate crop yield forecasting is vital for food industry decision-making.
• Existing models often use a one-fits-all approach, ignoring spatial variability in vegetation and thermal conditions.
• Redundant data in nonlinear machine learning models can negatively impact prediction accuracy.

Purpose of the Study:

• To propose a novel framework for enhancing crop yield prediction accuracy.
• To address the challenges of spatial variability and data redundancy in crop yield forecasting.
• To improve the reliability of vegetation and thermal condition index-based prediction models.

Main Methods:

• Utilized higher-order spatial independent component analysis (sICA) to consolidate spatial variability.
• Implemented a Principal Component Analysis (PCA) and Machine Learning (ML) combination (PCA-ML) to handle data challenges.
• Applied the framework to subregional rice yield forecasting in Vietnam.

Main Results:

• Subregional models improved rice yield forecasting accuracy by 20%–60% compared to the one-fits-all approach.
• The PCA-ML combination outperformed ML-only models by 18.5%–45%.
• The framework achieved reliable predictions 1–2 months ahead of harvest with an average 5% error.

Conclusions:

• The proposed framework effectively enhances crop yield prediction accuracy by considering subregional spatial variability.
• The combination of PCA and ML offers a significant improvement over traditional ML-only approaches.
• This method provides a reliable and accurate tool for early crop yield forecasting.