Image-driven in situ grading of compost maturity using deep feature clustering and supervised prediction

Area of Science:

• Agricultural Science
• Computer Science

Background:

• Accurate compost maturity assessment is vital for safe agricultural application and efficient composting.
• Current methods often lack precision or are not cost-effective for real-time monitoring.

Purpose of the Study:

• To develop an image-driven, in situ compost maturity grading framework.
• To enable multi-level maturity differentiation beyond simple binary classification.

Main Methods:

• Utilized deep learning models (P-ResNet-18) for image classification and feature extraction.
• Employed latent Dirichlet allocation (LDA) for refined maturity grading and multi-level prediction.
• Analyzed feature evolution using Grad-CAM, identifying color and texture as key indicators.

Main Results:

• P-ResNet-18 achieved excellent classification performance (metrics > 0.94) with strong generalization.
• LDA demonstrated superior grading performance (80% coverage, 90% purity) and correlation with composting duration.
• Supervised multi-level prediction models exceeded 90% accuracy.
• The framework showed robustness to noise (>93% consistency).

Conclusions:

• The developed framework offers accurate, cost-effective, and multi-level compost maturity grading.
• It supports intelligent composting management by improving end-point determination and reducing over-composting.
• Image analysis, particularly color and texture, provides valuable insights into compost development stages.