Improving Rice Seed Quality Assessment with Multi-Spectral Imaging and Environmental Data Integration

Accurate assessment of rice seed quality is essential for improving crop yield, enhancing resource efficiency, and ensuring long-term agricultural sustainability. Conventional seed quality evaluation methods, such as germination tests and vigor measurements, are time-consuming, labor-intensive, and unsuitable for large-scale or continuous field monitoring. This study proposes an advanced, non-destructive assessment framework that integrates multi-temporal multi-spectral imagery with synchronized environmental data to improve rice seed quality prediction. The Multitemporal Multispectral Imagery for Rice Yield and Phenology Prediction dataset was utilized, comprising spectral bands (Blue, Green, Red, Red Edge, and NearInfrared) and environmental parameters (temperature, humidity, soil moisture, rainfall, and solar radiation) collected across multiple phenological stages. The framework employs an intermediate fusion strategy, combining spectral-spatial features extracted via convolutional neural networks with temporal patterns from environmental variables processed through long short-term memory networks. Experimental results show that the integrated model significantly outperforms unimodal and traditional machine learning approaches, achieving $95.4 \%$ classification accuracy and an $\mathbf{R}^{\mathbf{2}}$ value of 0.92 for regression tasks. Feature analysis reveals that red-edge and NIR bands, along with soil moisture and temperature, are the most influential predictors of seed quality. This research demonstrates that combining multi-spectral imaging with environmental data provides a robust and scalable solution for precise, real-time rice seed quality monitoring, offering valuable insights for precision agriculture applications.