Improved Yield Prediction of Winter Wheat Using a Novel Two-Dimensional Deep Regression Neural Network Trained via...

Area of Science:

• Agricultural Science
• Remote Sensing
• Machine Learning

Background:

• Accurate crop yield estimation is crucial for food security and agricultural management.
• Remotely sensed data and machine learning have shown promise in predicting crop yields.
• Existing methods often overlook the spatial variability within crop fields.

Purpose of the Study:

• To enhance crop yield prediction accuracy using Convolutional Neural Networks (CNNs).
• To introduce a novel CNN architecture, Hyper3DNetReg, for precise yield mapping.
• To leverage multi-channel raster data for early-season winter wheat yield prediction.

Main Methods:

• Developed Hyper3DNetReg, a CNN architecture processing multi-channel raster inputs to output pixel-wise yield predictions.
• Utilized eight rasterized features including nitrogen rate, precipitation, topography, and Sentinel-1 radar data.
• Employed early-season (March) data to predict harvest-season (August) winter wheat yields.

Main Results:

• Hyper3DNetReg generated a yield prediction map by aggregating overlapping prediction patches.
• The proposed CNN method outperformed five traditional machine learning and regression models.
• Leave-one-out cross-validation demonstrated superior prediction accuracy for rain-fed winter wheat.

Conclusions:

• CNNs, particularly Hyper3DNetReg, offer a significant advancement in spatial crop yield prediction.
• Early-season remote sensing data combined with advanced CNNs can reliably forecast harvest yields.
• The methodology provides a valuable tool for precision agriculture and farm management.