Estimation of biophisical parameters in rice cropping system from Sentinel-2 data and hybrid approach: perspective for precision agriculture application

The availability of information able to assess crop nutritional status in space and time is a crucial issue to support sustainable agriculture in smart farming framework. Remote sensing techniques have become popular methods to support precision farming activities by producing spatial variability maps of crop conditions. In this framework, an experimental activity has been conducted to estimate Leaf Area index (LAI) and test potentiality of Canopy Nitrogen Content (CNC) retrieval, due to the importance of these parameters for assessing crop nutritional status. This study focuses on rice and retrievals have been conducted using a hybrid approach based on Radiative Transfer Model (RTM) simulations and Machine Learning Regression Algorithms (MLRA). A Look Up Table of rice spectra, with a cardinality of 2000 samples, was generated ranging crop parameters as input to the PROSPECT-PRO RTM. Simulations were resampled to 8 bands Sentinel 2-like configuration and were then used to train Gaussian Process Regression (GPR) and Neural Network (NN) MLR algorithms, also testing a sample selection procedure based on Active Learning (AL). Cross-validation results showed good performance for LAI retrieval using both the standard hybrid model (GPR: R2 ~ 0.78 , NN: R2 ~ 0.72) and AL approach (GPR: R2 ~ 0.71, NN: R2 ~ 0.67). Preliminary tests conducted to estimate CNC revealed promising results for plant nutritional status assessment. The within-field spatial variation of estimated CNC from Sentinel-2 (S2) data in a precision farming experiment resulted coherent with the observed heterogeneity in the field and to corresponding prescription maps used to manage the fertilisation.