Machine learning for enhancing wheat yield prediction accuracy with drone images

Overview

This solution combines soil electrical conductivity (EC) mapping with high-quality drone-based multispectral and thermal imagery to train machine learning models that deliver more accurate in-season wheat yield estimates. By accounting for both soil variability and plant physiological status, the approach overcomes limitations of previous remote sensing efforts that relied on imagery alone or used lower-quality cameras. The result is a yield prediction capability suited to precision agriculture management and to the rapid screening of thousands of experimental lines in wheat breeding programs each year.

Technical specifications

• Soil variability quantification using high spatial resolution EC maps generated with a Veris MSP3 across multiple study sites
• Multitemporal drone imagery collected throughout the growing season, including 10 narrow spectral bands in the multispectral camera and calibrated thermal data
• Post-processing kinematic (PPK) capabilities integrated with swappable camera payloads to eliminate the need for laborious ground control points while improving image quality
• Machine learning algorithms such as Partial Least Squares Regression trained on fused soil EC, multispectral, thermal, and heat unit (physiological status) inputs to estimate wheat yield
• Ground reference data including grain yield and supporting measurements supplied by the K-State wheat breeding lab for model training and validation
• Multi-site validation across three study sites in collaboration with the K-State wheat breeding program

Technology readiness level

The approach is at an active development and validation stage. Soil EC mapping, drone-based multispectral and thermal image collection, and machine learning model training are being carried out in collaboration with the K-State wheat breeding lab across three study sites. The current effort focuses on demonstrating that fusing soil variability and plant physiological metrics with high-quality multitemporal imagery improves yield estimation accuracy over prior remote sensing methods, with the goal of enabling practical deployment for both precision agriculture and high-throughput breeding screening.