Enhancing LAI estimation using multispectral imagery and machine learning: a comparison between reflectance-based and vegetation indices-based approaches

Authors: CHATTERJEE, SUMANTRA - Texas Agrilife Research; BAATH, GURJINDER - Texas Agrilife Research; SAPKOTA, BALA RAM - Texas Agrilife Research; Flynn, Kyle; Smith, Douglas

Submitted to: Computers and Electronics in Agriculture
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 12/4/2024
Publication Date: 12/18/2024
Citation: Chatterjee, S., Baath, G.S., Sapkota, B., Flynn, K.C., Smith, D.R. 2024. Enhancing LAI estimation using multispectral imagery and machine learning: a comparison between reflectance-based and vegetation indices-based approaches. Computers and Electronics in Agriculture. https://doi.org/10.1016/j.compag.2024.109790.
DOI: https://doi.org/10.1016/j.compag.2024.109790

Interpretive Summary: Leaf area index (LAI) is a critical growth parameter in precision agricultural applications, but estimating LAI across large and complex fields is challenging. Remote sensing approaches, particularly unmanned aerial vehicles (UAV)-based imagery remote sensing, can offer on demand, cost-effective solutions. While vegetation indices (VIs) are commonly used for LAI estimation with multispectral imagery, they often face issues such as inconsistency and saturation in dense crop canopies, especially in late growth stages of crops like corn. Alternatively, canopy spectral reflectance, if used directly with advanced statistical tools like machine learning, can potentially overcome these limitations of VI-based models and provide more consistent and accurate LAI predictions. Results demonstrated that reflectance-based models outperformed VI-based models, especially at mid-late vegetative growth and early reproductive stages. VI-based models performed better at early vegetative growth stages. Findings suggest that the direct use of canopy spectral reflectance with machine learning algorithms could enhance LAI predictions, benefiting precision agriculture applications in dense canopy crops like corn.

Technical Abstract: Leaf area index (LAI) is a critical growth parameter in precision agricultural applications, but estimating LAI across large and complex fields is challenging. Remote sensing approaches, particularly unmanned aerial vehicles (UAV)-based remote sensing, can offer on demand, cost-effective solutions. While vegetation indices (VIs) are commonly used for LAI estimation with multispectral imagery, they often face issues such as inconsistency and saturation in dense crop canopies, especially in late growth stages of crops like corn (Zea Mays L.). Alternatively, canopy spectral reflectance, if used directly with advanced statistical tools like machine learning, can potentially overcome these limitations of VI-based models and provide more consistent and accurate LAI predictions. This research used a large diverse dataset of LAI data collected from corn experimental plots treated with seven planting times in three different field locations from 2022-2023. A side-by-side performance comparison of reflectance-based models and VI based-models was conducted for LAI predictions across multiple datasets, categorized based on six different crop growth stages, three field locations and all combined, using five different machine learning algorithms. Machine learning modeling was executed for each dataset via k-fold cross-validation using 80% data, and the model performance was externally tested with the remaining 20% independent data. Results demonstrated that reflectance-based models outperformed VI-based models, especially at mid- to late vegetative growth (5 -15%) and silking stages (25%). The superior performance of reflectance-bands was regulated by red edge and near infrared bands due to their greater sensitivity to higher LAI. However, VI-based models performed better at early vegetative growth stages, primarily due to the effectiveness of soil-adjusted indices like Modified Soil Adjusted Vegetation Index (MSAVI). Machine learning algorithms such as K Neighbors Regressor and Extra Tree Regressor were most effective for modeling LAI, regardless of whether reflectance-based or VI-based approaches were used. Findings suggest that the direct use of canopy spectral reflectance with machine learning algorithms could enhance LAI predictions, benefiting precision agriculture applications in dense canopy crops like corn.