StressNet: a spatial-spectral-temporal deformable attention-based framework for water stress classification in maize.

Area of Science:

• Agricultural Science
• Remote Sensing
• Computer Vision

Background:

• High-throughput crop monitoring is crucial for agricultural health.
• Existing methods often rely on visible spectrum imagery, limiting abiotic stress detection.
• Water deficiency is a major abiotic stress affecting crop yield.

Purpose of the Study:

• To develop and evaluate a deep learning model for accurate water stress classification in maize.
• To investigate the efficacy of multispectral, multi-temporal imagery for stress detection.
• To enhance spatial and temporal feature extraction for improved classification accuracy.

Main Methods:

• Utilized multispectral imagery from Unmanned Aerial Vehicles (UAVs) during maize growth stages.
• Developed a Convolutional Neural Network (CNN) with deformable convolutions for spatial-spectral feature extraction.
• Employed an Attention-based Bi-Directional Long Short-Term Memory (Bi-LSTM) network for temporal feature analysis.
• Integrated spatial, spectral, and temporal features for final water stress prediction.

Main Results:

• The proposed model achieved a high accuracy of 91.30% for water stress classification.
• Precision and recall rates were 0.8888 and 0.8857, respectively.
• Multispectral, multi-temporal data significantly outperformed mono-temporal classification.

Conclusions:

• Multispectral, multi-temporal UAV imagery is effective for detecting water stress in crops.
• The combination of deformable convolutions and attention-based Bi-LSTM enhances spatial-spectral-temporal feature extraction.
• The developed model offers a valuable tool for precise and efficient crop water stress monitoring.