Abstract
This study presents a regionally trained version of the “CArbonate system and Nutrients concentration from hYdrological properties and Oxygen using a Neural network” (CANYON) method, named CANYON-PU, for estimating primary macronutrients (phosphates, silicates, and nitrates) in the Peruvian Upwelling System (PUS). Using a neural network approach, the model was trained using extensive biogeochemical data spanning between 2003 and 2021, collected by the Peruvian Institute of Marine Research (IMARPE). Variables representing the low-frequency variability related to ENSO were introduced in the training and significantly improved the performance of the algorithm. The performance of CANYON-PU was validated against independent datasets and demonstrated an improvement in accuracy over the global CANYON model that struggled to represent the nutrient distribution in the PUS mainly due to the lack of samples in its training. Therefore, CANYON-PU successfully captured nutrient variability across different spatial and temporal scales, showcasing its applicability to diverse datasets, including high-frequency data such as profiling floats or gliders. This work highlights the effectiveness of neural networks for representing the nutrient distribution within highly variable ecosystems like the PUS.