On The Location-Independent Reconstruction Of Photosynthetically Active Radiation In The Water Column Using Neural Networks

Accurate reconstruction of photosynthetically active radiation (PAR) in aquatic environments is critical for understanding primary production and ecosystem dynamics. This study evaluates the generalization abilities of artificial neural networks (ANNs) for location-independent PAR reconstruction using data from BGC-Argo floats. The proposed ANN model is trained on datasets from multiple geographic regions and validated against independent test data from diverse oceanic locations. Comparisons with multiple linear regression (MLR) and regression tree (RT) models demonstrate that the ANN consistently achieves superior predictive accuracy, with R² values exceeding 0.97 in most test cases. The results indicate that neural networks can effectively generalize across different marine environments, even in regions with distinct optical properties. Notably, the ANN outperforms alternative models except in one test case, highlighting the potential influence of regional environmental factors. This study underscores the potential of machine learning techniques to enhance bio-optical sensor configurations and reduce the necessity for dedicated PAR sensors.