Crop Traits Retrival from Prisma Imagery with Machine Learning: A Comparison of Data Driven and Hybrid Approach with a Multi-Year Multi-Crop Dataset

In the coming years, new operational hyperspectral missions, like Copernicus-CHIME and NASA-SBG, will deliver unprecedented spectroscopic data to support sustainable agriculture and food security. In this framework, efficient plant traits retrieval methods need development and testing on real spaceborne data. A multi-year (2020–2023), PRISMA and ground data collection, has been created for algorithm development and validation. A machine learning regression algorithm (MLRA) was developed and compared to a hybrid approach (HYB) and its variant with active learning (HAL). These methods were tested to evaluate crop traits such as Leaf Area Index (LAI), chlorophyll content at leaf (LCC) and canopy scale (CCC) as well as nitrogen content (LNC, CNC). The results showed that while retrieving foliar biochemical characteristics from space is challenging, canopy-level traits can be efficiently derived using MLRA and HAL approaches (R2 > 0.70 and rRMSE under 15%), emphasizing the importance of using real data for model training.