Explainable light-weight deep learning pipeline for improved drought stress identification.

Area of Science:

• Agricultural Science
• Computer Science
• Plant Physiology

Background:

• Early detection of crop drought stress is crucial for mitigating yield loss.
• Non-invasive imaging and sensor data are valuable for machine learning in agriculture.
• Existing methods need optimization for real-world field conditions.

Purpose of the Study:

• To develop a novel deep learning framework for classifying drought stress in potato crops.
• To enable real-time, accurate drought stress identification in natural agricultural settings using UAV imagery.
• To enhance the interpretability of deep learning models in agricultural applications.

Main Methods:

• A novel deep learning framework combining a pre-trained network (DenseNet121) with custom layers for feature extraction and dimensionality reduction.
• Classification of drought stress in potato crops using imagery captured by unmanned aerial vehicles (UAVs).
• Integration of Gradient-Class Activation Mapping (Grad-CAM) for model interpretability and visualization of decision-making processes.

Main Results:

• The proposed framework achieved a precision of 97% for the stressed class and 91% overall accuracy.
• Demonstrated superior performance compared to state-of-the-art object detection algorithms.
• The explainable nature of the model provides actionable insights into drought stress identification.

Conclusions:

• The developed explainable deep learning framework accurately identifies drought stress in potato crops under natural conditions.
• This approach offers a powerful tool for precision agriculture, enabling timely interventions.
• Enhanced model interpretability fosters trust and facilitates practical adoption in field applications.