Machine Learning Benchmarking under Data-Limited Conditions: A Case Study of Environmental Associations with Loligo spp. Catch Characteristics

Abstract

The application of machine learning (ML) in small-scale fisheries is often constrained by limited data availability, posing challenges for model development and reliability. This study presents a ML benchmarking framework for data limited artisanal fisheries, aiming to evaluate the predictive performance of supervised algorithms in assessing the effects of environmental variability on Loligo spp. catch size and quantity. The dataset comprised 278 fishing-trip observations collected from artisanal fishing vessels between October 2024 and September 2025. Catch records were integrated with environmental variables, including sea surface temperature (SST), sea surface salinity (SSS), dissolved oxygen (DO), and air temperature. Given the limited data typically available in small-scale fisheries, the study emphasizes model performance and interpretability as key considerations for practical fisheries applications. Three classification models—Decision Tree (DT), k-Nearest Neighbors (k-NN), and Logistic Regression (LR)—were comparatively evaluated to assess their predictive performance. Results identified SSS as an important environmental variable associated with squid catch characteristics. For catch size classification (≤10 cm vs. >10 cm mantle length), DT achieved the highest accuracy (89.04%) among the evaluated models. In contrast, catch-quantity classification across three categories showed lower performance overall, with DT achieving the best accuracy (57.14%). These findings demonstrate that, under data-limited conditions typical of artisanal fisheries, simpler and interpretable models such as DT can provide strong predictive performance. This study highlights the practical applicability of ML in small scale fisheries and provides a methodological reference for low-data environments, supporting decision-making and adaptive fisheries management under environmental variability.