Estimating plant height, nitrogen uptake and above-ground biomass using UAV multispectral imaging coupled with machine learning in industrial hemp (Cannabis sativa L.)

Industrial hemp ( Cannabis sativa L.) is known for its high fiber production with lower ecological footprint. Nitrogen (N) status and stem biomass (SB) and total above-ground biomass (AGB) of the crop highly influence fiber quantity and quality. Conventional monitoring practices are labour intensive and time consuming. Unmanned Aerial Vehicles (UAVs) with imaging sensors can be a promising tool for mitigating these challenges. This study evaluated the performance of multispectral camera-equipped UAV in predicting key agronomic parameters, i.e., plant height (PH), Leaf Nitrogen Uptake (LNU) and SB and AGB. Field trials were conducted at UF/IFAS West Florida Research and Education Centre, Jay, FL during the years 2021 and 2022 consisting of two cultivars and six N treatments. The PH was estimated through Crop Height Model, yielding an R 2 of 0.87 at full crop maturity (90 days after planting). Twenty-seven Vegetation Indices (VIs) were extracted and features, including PH and VIs, were selected through Recursive Feature Elimination with adjusted Variance Inflation Factor (VIF<10) to develop machine learning models for the estimation of yield components. The LNU prediction was best with Support Vector Machine model with R 2 , RMSE and nRMSE % value of 0.364, 34.55 kg N ha −1 and 68.48 respectively. Random Forest Regressor predicted the SB and total AGB most accurately with R 2 , RMSE and nRMSE % value of 0.752 and 0.707, 890.70 and 1492.73 kg ha −1 , 48.86 and 43.05 respectively. The results demonstrate the potential of UAVs to generate more reliable estimates of PH, SB and total AGB whereas it remained unreliable for LNU. • The UAV multispectral imaging-derived crop height model is efficient in estimating hemp plant height (PH). • The UAV-PH is highly correlated with Leaf Nitrogen Uptake (LNU), stem biomass (SB) and total above-ground biomass (AGB). • The Support Vector Machine algorithm with ‘rbf’ kernel performed superior in predicting LNU uptake. • The Random Forest model predicted SB and total AGB with the highest accuracy among the tested models.