Assessing rice genotype dependency in remote sensing: challenges in nitrogen discrimination and yield forecasting

Background Rice ( Oryza sativa L.) is a staple crop that accounts for 8% of the global primary crop production. This sector faces an environmental challenge driven by climate change, rising water scarcity, and the need for more resilient and adaptive production systems. Although Remote Sensing (RS) offers solutions for optimizing inputs, most current models are calibrated for specific varieties, limiting their application across diverse cultivars. This study evaluates the transferability of RS models for monitoring nitrogen (N) fertilization status and predicting yield across a highly heterogeneous dataset of rice genotypes, locations and seasons. Materials and methods Six field trials were conducted across three locations in Spain (Valencia and Tarragona) during seasons 2022 and 2023. The study analysed over 170 cultivars, including commercial Japonica and Indica varieties, and a selection of 170 non-commercialized-under development varieties, both subjected to low (100 kg N/ha) and high (200 kg N/ha) fertilization regimes. Multispectral UAV imagery (MAIA S2) was normalized using Accumulated Growing Degree Days (GDD) to align phenological stages across sites. Random Forest (RF) classifiers were employed to analyse the capacity of RS to identify whether rice paddies are under- or over-fertilized. The transferability of N models between rice genotypes was also assessed. Furthermore, the previously established MS3 + yield regression model, originally developed for the JSendra variety, was evaluated against a multi-variety dataset. Results Random Forest classifiers effectively discriminated between nitrogen application rates across diverse genotypes, with several sites exceeding an 85% validation accuracy. A consistent trend emerges when analysing spectral importance: visible (VIS) bands take importance during the early season stages, whereas near-infrared (NIR) and red-edge (RE) reflectance provide critical diagnostic information throughout the entire crop cycle. Notably, 82% of the evaluated varieties demonstrated a high compatibility with the global model. Conversely, the yield model showed limited transferability between varieties. While it performed poorly on the global dataset, it successfully predicted yields for 48.8% of commercial varieties (residues within ± 1 tons per hectare), specifically those with genetic and structural similarities to the training variety. Conclusion The study concludes that N-status monitoring via RS classifiers is robust across varying rice genetics, whereas yield prediction models exhibit strong genotype dependency.