Retrieval of Wheat Traits from Prisma Data Through Neural Networks with Different Training Setups

The rationalisation of farm operations is a fundamental task to improve sustainability and facilitate more effective responses to climate-related challenges. Hyperspectral data could prove a valuable asset in this regard, enhancing our ability to observe and monitor a range of processes occurring within the crop. In this study, we evaluated the performance of different artificial neural network architectures to predict key wheat biophysical parameters: leaf area index (LAI), canopy chlorophyll (CCC), and nitrogen content (CNC), using PRISMA hyperspectral imagery. The models were trained with different datasets: i) ground-sampled hyperspectral measurements, ii) PRISMA satellite spectra, and iii) reflectances simulated from radiative transfer model (PROSAIL-PRO). Results show satisfactory performance for the three training scenarios considered, with best results when ground data are used. The hybrid solution results are encouraging and may represent a potentially more transferable solution to other sites/seasons.