Random forest algorithm and regional applications of spectral inversion model for estimating canopy nitrogen concentration in rice

Abstract

Canopy Nitrogen Concentration (CNC) is a key indicator of crop yields. It is feasible to establish a realtime regional model to estimate CNC by upscaling the field-scale spectral model. This study focuses on monitoring the CNC in rice on a large scale in real-time. The Random Forest (RF) algorithm is used to establish the CNC spectral inversion model, and some vegetation indexes that are sensitive to nitrogen were selected as input parameters for the RF. CNC was selected as an output parameter. The hyperspectral and biochemical data were collected in a paddy in Changchun City, Jilin Province, China, and the data in Suzhou was used to test the model's universality and effectiveness. Two regional-scale models were developed by applying scale transformation based on the input and output variables respectively. The results show that the RFCNC model (CNC spectral inversion model based on the RF algorithm) performed accurately and significantly improved upon existing methods. R, used to validate method accuracy in Changchun and Suzhou, was 0.82 and 0.73 respectively. The regional application accuracy increased (R =0.81) through the two upscaling methods using hyperspectral remote sensing satellite images. This study suggests that this method is promising for estimating regional CNC in rice by upscaling a field-scale spectral model if the strategy is appropriately selected.