Assessing maize nitrogen content from hyperspectral data through MLRA: preliminary results in a Precision Farming framework

The increasing demand of food production pushes for the development of more sustainable cropping system, to reduce farm costs and the environmental impact of fertilisation. In this context, the assessment of crop nitrogen status, from optical remote sensing data, is a key information for site specific fertilisation management, in a precision agriculture (PA) workflow. Data provided by new hyperspectral satellite sensors, from both future international initiative (e.g. DLR-EnMap, ESA-CHIME, NASA-HyspIRI) and ongoing mission (ASI-PRISMA, launched on March 2019), are expected to improve the retrieval of vegetation biophysical variables (BVs). However, handling the high dimensionality of these data requires efficient retrieval approaches. In this framework, a field campaign was conducted in July 2018, in a rural area in Tuscany, Italy (Grosseto site - Lat 42.83, Lon 11.07), in conjunction with airborne HyPlant-DUAL imagery acquisitions (ESA FLEX and ESA CHIME experimental activities). Measurements of Leaf Area Index (LAI), Chlorophyll (Cab), Nitrogen (N%) and biomass were performed in 88 plots of about 10 × 10 m2, distributed in two maize fields. BVs maps were generated testing different Machine Learning Regression Algorithms (MLRAs), through the ARTMO toolbox. According to cross validation analysis, Gaussian process regression (GPR) resulted one of the best methods in the estimation of LAI (RMSE=0.48, nRMSE=9%, R2=0.89), Cab (RMSE=4.01, nRMSE=17%, R2=0.58), Chlorophyll Canopy Content, CCC (RMSE=25.9, nRMSE=9%, R2=0.84) and N% (RMSE=0.45, nRMSE=26%, R2=0.44), with the added value of providing also maps of uncertainties. Maize nitrogen uptake (NU) was then estimated following several approaches based on BVs combinations. Results showed that derived NU estimates are significantly correlated with independent ground measurements (R2>0.7). These findings are promising for an estimation of crop nutritional status, as a quantitative information for sustainable agro-practices based on variable rate nitrogen fertilisation. Further developments will be devoted to i) investigate the use of hybrid approaches for parameter retrieval and ii) test the PRISMA data, acquired in 2019 on the same study area.