??AI for Maize: Revolutionizing Crop Yield Prediction in Breeding ??

Improving yield prediction is not easy, as traditional models struggle to generalize across different conditions and make sense of complex, multi-modal data.

One of the latest researches of Purdue University, United States ???? investigates the prediction of maize grain yield using deep learning models that integrate multi-modal remote sensing data, including hyperspectral imagery and LiDAR point clouds, alongside weather and genetic data.

The study focuses on improving prediction accuracy and interpretability by utilising attention mechanisms within the models.

Role of AI

AI plays a critical role in this research through the application of advanced deep learning techniques. Specifically:

1. LSTM Networks: These recurrent neural networks are designed to handle time series data from RS inputs throughout the growing season. They capture the temporal dependencies in crop development and enable accurate predictions.
2. Attention Mechanisms: Attention mechanisms are employed to improve both the accuracy and interpretability of the models. They assign importance to different RS data inputs at various growth stages, identifying critical periods in maize development that significantly influence yield.
3. Multi-modal Deep Learning: The AI model assimilates inputs from heterogeneous data streams—hyperspectral imagery, LiDAR point clouds, and environmental data (e.g., weather)—to generate accurate yield forecasts. The combination of multiple modalities enhances the model's ability to capture the complex relationships between the environment, genetics, and crop growth.

Core Numerical Insights

• The models achieved R2 values ranging from 0.82 to 0.96, indicating high accuracy in predicting maize grain yields.
• The multi-modal architecture performed best, with R2 values of 0.96 and RMSE of 3.21 tons per hectare, demonstrating superior performance over traditional vanilla LSTM models.
• Attention mechanisms improved model interpretability by showing that LiDAR data contributed most during early growth stages, while hyperspectral data had higher importance during mid- and late-season stages.
• Four different prediction scenarios were evaluated, with the inclusion of all RS dates providing the most accurate predictions, though reduced time-step scenarios still yielded robust results (e.g., R2 values over 0.8).

This study highlights the potential of AI-driven, multi-modal RS models to improve yield predictions and support more informed decision-making in plant breeding and crop management.

The research concludes that multi-modal remote sensing, integrated with deep learning models and attention mechanisms, can provide accurate and interpretable predictions of maize grain yield. The insights gained from the attention weights can inform data collection strategies and enhance our understanding of the factors influencing yield.        

In the figure above: (A) Vanilla stacked LSTM with early fusion of both concatenated features from VNIR, LiDAR and weather data. (B) LSTM with attention mechanism, with early fusion of the modalities. (C) Multi-modal network with separate networks for each RS modality concatenated with weather, adding a late fusion module.

AI in Agriculture Podcast

What's on

If you'd like to receive the regular 'AI in Agriculture' newsletter in your inbox, simply add your email to my mailing list.

Join over 8,900 readers who enjoy weekly updates on AI advancements in agriculture!

To get more information about maize phenotyping capabilities with mobile - ask Petiole Pro