WaveST-Yield: a novel spatio-temporal deep learning framework with frequency-domain refinement for UAV-based maize...

Area of Science:

• Agricultural Science
• Remote Sensing
• Deep Learning

Background:

• Accurate maize yield prediction is vital for precision agriculture and food security.
• Existing models struggle with spatio-temporal dynamics and feature extraction from UAV multispectral data.
• UAV-based remote sensing offers rich data but requires advanced models for effective analysis.

Purpose of the Study:

• To develop a novel hybrid deep learning framework, WaveST-Yield, for accurate plot-scale maize yield prediction.
• To improve the mining of spatio-temporal dynamics and yield-sensitive features from multi-temporal multispectral data.
• To enhance the separation of spatial details from background noise in remote sensing data.

Main Methods:

• Proposed WaveST-Yield framework integrating Spatio-Temporal Phenology Encoder (SPE), Multiscale Frequency-Spatial Refiner (MFSR), and Adaptive Yield-Sensitive Re-calibrator (AYSR).
• SPE uses ConvLSTM for temporal dynamics; MFSR employs Haar Wavelet Downsampling (HWD) for detail preservation and noise reduction.
• AYSR utilizes 3D-CBAM attention for yield-trait enhancement and background suppression.

Main Results:

• WaveST-Yield achieved high prediction accuracy (R² of 0.883 in Field 1, 0.775 in Field 2) outperforming existing models.
• The model demonstrated superior error control and robustness in cross-validation and external validation.
• Ablation studies confirmed the effectiveness of integrated spatio-temporal encoding, frequency refinement, and attention mechanisms.

Conclusions:

• WaveST-Yield provides a highly accurate and robust methodological framework for maize yield monitoring.
• The hybrid deep learning approach significantly improves generalization ability across different fields.
• This framework advances high-throughput crop yield monitoring for precision agriculture applications.