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

Accurate and generalizable plot-scale maize yield prediction is critical for precision agriculture and food security. While UAV-based multispectral remote sensing provides rich phenotyping data, existing yield prediction models often struggle with insufficient mining of complex spatio-temporal dynamics, ineffective separation of spatial details from background noise, and inadequate focus on yield-sensitive features throughout the crop growth cycle. To address these limitations, this study proposes WaveST-Yield, a novel hybrid deep learning framework tailored for multi-temporal multispectral data. The proposed model integrates three core modules: a Spatio-Temporal Phenology Encoder (SPE) based on ConvLSTM to capture the temporal dynamic patterns and spatio-temporal correlations across the entire growth period; a Multiscale Frequency-Spatial Refiner (MFSR) utilizing Haar Wavelet Downsampling (HWD) to preserve image details and decouple noise without early loss of key physiological features; and an Adaptive Yield-Sensitive Re-calibrator (AYSR) leveraging a 3D-CBAM attention mechanism to enhance the extraction of critical yield-related traits while suppressing background interference. The model was rigorously evaluated on two independent maize experimental fields using 5-fold cross-validation and cross-plot external validation. Results demonstrate that WaveST-Yield consistently outperforms traditional machine learning algorithms and single-structure deep learning models, achieving the highest prediction accuracy (Overall R² of 0.883 and 0.775 in Field 1 and Field 2, respectively) with superior error control. Extensive ablation and multi-model comparison experiments confirm that the synergistic integration of spatio-temporal encoding, frequency-domain refinement, and 3D attention mechanisms significantly improves model robustness and cross-regional generalization ability. This study provides a highly accurate, robust, and generalizable methodological framework for high-throughput crop yield monitoring.