Improving wheat yield estimation using hyper-temporal lidar and spectral data fusion

Overview

This solution improves wheat yield estimation by combining hyper-temporal (near-daily) lidar and multi-spectral data through machine learning. The core insight is that crop yield results from the combined influence of external plant structure and internal physiological traits. By fusing structural features (captured by lidar) with spectral features (captured by multi-spectral sensors), estimation accuracy is improved across all growth stages. Early-stage estimation is dominated by spectral features, while later stages benefit more from stable structural features. The approach has demonstrated a maximum prediction accuracy of 0.891 using all extracted phenotypic traits, outperforming structural-only (0.819) and spectral-only (0.876) models.

Technical specifications

Key features:

• Hyper-temporal data collection: Near-daily lidar and multi-spectral measurements from germination to maturity
• Structural trait extraction: 27 structural traits derived from lidar to characterize three-dimensional plant morphology
• Spectral trait extraction: 14 spectral traits derived from multi-spectral data to capture physiological differences
• Machine learning fusion: 4 different ML models used to integrate structural and spectral predictors for yield prediction
• Stage-specific optimization: Spectral features dominate early-stage estimation; structural features provide stability in later stages
• Current data source: Fieldscan system (Phenospex, Netherlands) providing ultra-high resolution data
• Future deployment: UAV-based lidar and spectral sensors for scalable, cost-effective large-scale yield estimation

Benefits for partners:

• Improved yield prediction accuracy compared to single-source approaches
• Scalable methodology adaptable from high-end sensors to affordable UAV platforms
• Applicable to precision agriculture, crop insurance, food security planning, and farm management decisions

Technology readiness level

The technology has been validated through a complete hyper-temporal data collection cycle from germination to maturity, with 27 structural and 14 spectral traits extracted and tested across 4 machine learning models. Prediction accuracy has been quantified, demonstrating that fused features outperform single-source approaches at all growth periods. The next validation phase aims to replicate these findings using UAV-mounted lidar and spectral sensors, which are more affordable and scalable than the current Fieldscan system. This follow-up study is estimated to take 1 year and requires equipment for multi-source UAV data collection, computing resources, and funding for key personnel.