<![CDATA[Fisheries Management Area (WPP) 573 is one of the upwelling regions of the Indian Ocean supporting both neritic small pelagic and oceanic pelagic fisheries, and is affected by complex climate variability. This research modeled spatial probabilities of fish pelagic habitat using an optimized shallow architecture of a Feedforward Network, specifically a Multi-layer Perceptron integrated with oceanographic satellite data. This research utilized 492 validated fishing coordinates from Global Fishing Watch (GFW) collected throughout 2025 as presence labels. At the same time, oceanographic variables included sea surface temperature, distance to the nearest coral reef, chlorophyll-a, salinity, sea surface elevation, current velocity, and bathymetry. To prioritize model generalization and mitigate overfitting, this research implemented the MLP with a single hidden layer containing three neurons, optimized via the L-BFGS algorithm with strong regularization. The final model with spatially- constrained pseudo-absences, environmental distance filtering obtained a True Skill Statistic (TSS) of 0.68, a 5-fold cross-validation accuracy of 83.95% and a presence precision of 95%, indicating that the model possess a good ability to discriminate among habitats and it is robust to variation. The permutation importance highlighted SST as the strongest predictor (0.1386), followed by salinity (0.0976) and current velocity (0.0866). Dynamic habitat probability maps generated reveal fish catch potential zones that correspond to the biophysical conditions of the water mass. In summary, this study demonstrates that MLP-based prediction, reinforced by stringent parameter optimization, is a very suitable approach to identify pelagic fish habitats, thus assisting in fishing operations and sustainable fishery management in WPP 573.]]>
Copyrights © 2026
