Data Assimilation Versus Machine Learning: Comparative Study Of Fish Catch Forecasting

Abstract

Data assimilation (DA) and machine learning (ML) are empirically compared for automatic daily fish catch forecasting (DFCF). ML would be a promising approach if large-scale data are available for training. Otherwise, DA would perform well, where prior knowledge on a monitoring target is incorporated into modeling. The present study aims to clarify the robustness of both approaches in DFCF with a small amount of data, and their evolution as the amount of training data increases. Experimental comparisons using catch and meteorological data demonstrate that a DA-based DFCF system yields a significant improvement over an ML-based systems with a small amount of data, and is comparable with ML-based systems with sufficient amount of data.