Comparison of Crop Trait Retrieval Strategies Using UAV-Based VNIR Hyperspectral Imaging

Area of Science:

• Agricultural remote sensing
• Hyperspectral imaging
• Machine learning applications in agriculture

Background:

• Unmanned aerial vehicles (UAVs) equipped with hyperspectral cameras offer sub-field scale crop monitoring.
• Illumination variations during the growing season can impact the accuracy of crop trait retrieval from UAV data.
• Understanding the influence of illumination variability on different retrieval methods is crucial for reliable crop monitoring.

Purpose of the Study:

• To compare the performance of four different crop trait retrieval methods using UAV-based hyperspectral data.
• To assess the impact of illumination variability on the accuracy of retrieving leaf area index (LAI), fractional vegetation cover (fCover), and canopy chlorophyll content (CCC).
• To identify the most promising retrieval method for hyperspectral data acquired by UAVs.

Main Methods:

• Four retrieval methods were evaluated: standard look-up table (LUTstd), regularized look-up table (LUTreg), hybrid methods, and random forest regression (RF) with and without exposure time (RFexp).
• The Soil-Leaf-Canopy (SLC) model was used for LUT-based and hybrid methods.
• Statistical methods (RF and RFexp) relied solely on in situ data, with RFexp incorporating exposure time as a variable.

Main Results:

• The RFexp method demonstrated the highest accuracy for LAI (5.36% NRMSE), fCover (5.87% NRMSE), and CCC (15.01% NRMSE), effectively mitigating illumination variability and cloud shadow effects.
• LUTreg outperformed LUTstd and hybrid methods, achieving NRMSEs of 9.18% for LAI, 10.46% for fCover, and 12.16% for CCC.
• RF generally showed higher accuracies than LUTreg for LAI and fCover, indicating the robustness of machine learning approaches.

Conclusions:

• Machine learning approaches, particularly random forest regression (RF) with exposure time (RFexp), are the most promising for retrieving crop traits from UAV-based hyperspectral data.
• RFexp's ability to reduce illumination variability effects makes it a reliable method for accurate crop trait estimation.
• The study highlights the potential of advanced statistical methods to overcome challenges in remote sensing data acquisition.