Bayesian Aggregation Improves Traditional Single-Image Crop Classification Approaches.

Area of Science:

• Agricultural Science
• Computer Science
• Data Science

Background:

• Accurate crop information aids in regulating agricultural stocks and financial market strategies.
• Machine learning, particularly with satellite imagery, is key for crop type recognition.
• Challenges include class imbalance and aggregating pixel-level data for field-level predictions.

Purpose of the Study:

• To develop and evaluate a Bayesian methodology for aggregating crop classification results.
• To compare different class balancing techniques for crop classification.
• To assess classical machine learning and U-Net performance against a Bayesian approach for single-satellite-image crop classification.

Main Methods:

• Comparison of various class balancing techniques.
• Evaluation of classical machine learning algorithms and the U-Net convolutional neural network for crop classification using single satellite images.
• Implementation and assessment of a Bayesian aggregation methodology for field-wise classification.

Main Results:

• The best single-satellite-image crop classification achieved 77.4% overall accuracy and a 0.66 Macro F1-score.
• Bayesian aggregation improved field-wise classification results by 1.5% compared to majority voting.
• The Bayesian aggregation approach demonstrated superior overall accuracy over majority voting and averaging strategies.

Conclusions:

• The proposed Bayesian aggregation methodology enhances the accuracy of field-wise crop classification.
• This approach offers a significant improvement over traditional aggregation methods like majority voting.
• Accurate crop classification using satellite imagery and advanced aggregation techniques is crucial for agricultural management and financial markets.