B037-0012
Mapping Canopy Nitrogen Concentration across Ryegrass and Barley Crop using Random Forest Regression

Wednesday, 9 December 2020
Poster
Manish Kumar Patel1,2, Dongryeol Ryu2, Andrew William Western1, Glenn Fitzgerald3,4, Eileen M. Perry3,5, Helen Suter6 and Iain Young7, (1)University of Melbourne, Department of Infrastructure Engineering, Parkville, VIC, Australia, (2)The University of Melbourne, Department of Infrastructure Engineering, Parkville, VIC, Australia, (3)Agriculture Victoria, Horsham, VIC, Australia, (4)University of Melbourne, Centre for Agricultural Innovation, Faculty of Veterinary and Agricultural Sciences, Parkville, VIC, Australia, (5)University of Melbourne, Department of Infrastructure Engineering, School of Engineering, Parkville, VIC, Australia, (6)The University of Melbourne, School of Agriculture and Food, Parkville, VIC, Australia, (7)The University of Sydney, School of Life and Environmental Sciences, Sydney, NSW, Australia
Abstract:
Canopy nitrogen concentration (CNC) is an important crop yield regulator. Mapping CNC using hyperspectral remote sensing has the potential to markedly improve the management of economically and environmentally costly nitrogen fertiliser through rapid, non-destructive and cost-effective monitoring. One important challenge is developing robust and transferrable models to interpret high dimensional data produced by such sensors.

Classification and Regression Tree (CART) based random forest regression is a commonly used machine learning predictive model with relatively few hyperparameters to tune. It is an intuitive algorithm suitable for predicting many biophysical variables, including canopy properties; which forms an ensemble of decorrelated high variance regression trees. However, in most past studies, the crop data is either limited to certain growth stages or a specific crop type. Therefore, the transferability of random forest models across growth stages and different crop types is unclear.

This study aims to investigate the robustness of the random forest model across two different crops; ryegrass and barley. Canopy reflectance was measured using an ASD HandHeld-2 field spectroradiometer (325 nm – 1075 nm) (Analytical Spectral Devices, Inc., Boulder, CO, USA), together with corresponding ground CNC data in ryegrass (winter (May-June 2018), summer (January-February 2019) at four growth stages in each season, as part of a sub-research project of the More Profit from Nitrogen Program) and barley (August-October 2019) at five growth stages till harvest. There were eight and three nitrogen treatments for ryegrass and barley respectively. The ryegrass was grown under irrigation while barley was rainfed. The random forest hyperparameters, which consist of number of trees, leaf node size and number of variables at each split (Mtry), were tuned using a grid search on out-of-bag error calculated for 10,000 randomly drawn samples. Results showed that the random forest algorithm worked well when trained on a specific crop (0.53≤R2≤0.86 for ryegrass; 0.56≤R2≤0.91 for barley test data) but that the predictive power dramatically dropped (-0.06≤R2≤0.27 on barley; -1.34≤R2≤-0.25 on ryegrass) when applied to the other crop. Variable importance in projection (VIP) analysis highlighted two promising spectral regions (within the visible to near infrared), blue for ryegrass and blue and near infrared (NIR) for barley. The findings of this study suggest that the reflectance characteristics of crops changes substantially between species. Shifting importance of specific bands and varying canopy morphology together make random forest models for CNC prediction from reflectance spectra that are trained on a specific crop unfit for generalization to other crops.

<< Previous Abstract | Next Abstract >>