Detecting Intra-Field Variation in Rice Yield With Unmanned Aerial Vehicle Imagery and Deep Learning.

Area of Science:

• Agricultural Science
• Remote Sensing
• Machine Learning

Background:

• Unmanned Aerial Vehicles (UAVs) with multispectral sensors provide high-resolution data for early crop stress monitoring.
• Integrating UAV data with machine learning can enhance real-time agricultural management, but guidance is limited.

Purpose of the Study:

• To compare two deep learning strategies for early crop stress detection and yield prediction in irrigated rice using UAV imagery.
• To evaluate the effectiveness of 3D Convolutional Neural Networks (CNNs) and 2D-CNNs for predicting field-scale yield.

Main Methods:

• Utilized multitemporal UAV imagery throughout the growing season for irrigated rice in Arkansas.
• Implemented a 3D-CNN to capture spatial and temporal dimensions across multiple time points.
• Applied a 2D-CNN focusing on spatial relationships from single-day imagery.

Main Results:

• Both deep learning strategies surpassed traditional methods like linear regression and gradient boosted decision trees.
• The 3D-CNN achieved a Root Mean Squared Error (RMSE) of 8.8% of the mean yield.
• The 2D-CNN showed RMSEs ranging from 7.4% to 8.2% when images were taken during the booting stage or later.

Conclusions:

• Deep learning models, particularly convolutional autoencoders, show significant promise for UAV-based yield prediction in rice.
• The spatial denoising effect of these models improves pixel-level yield predictions based on surrounding data.
• These findings support the integration of advanced AI techniques with remote sensing for precision agriculture.