Machine learning cluster analysis of haplochromine schools and pelagic serial deduction for improved fish biomass estimates using Lake Victoria’s acoustic data

Haplochromines in Lake Victoria, most of which are extinct or threatened, represent a diverse phylogenetic group of cichlids with relatively narrow genetic and morphological variation. They play a key ecological role as primary prey for Nile perch and important grazers of phytoplankton. However, biomass estimation from annual acoustic surveys has been constrained by methodological limitations and classification uncertainty. This study presents an unsupervised machine learning approach to estimate haplochromine biomass using echosounder data collected from stock assessment surveys conducted between 1999 and 2022. A serial deduction framework was applied to progressively classify fish schools and isolate species-specific backscatter, with an average detection of 30.14 schools per kilometre (IQR: 13–57). Schools were first classified as Dagaa or non-Dagaa using supervised methods, after which non-Dagaa schools were clustered using K-means based on acoustic features including backscattering intensity, depth, height, and corrected school dimensions, producing four assemblage clusters. Clusters consistent with haplochromine characteristics, validated using independent trawl and acoustic data, were assigned as haplochromines, while others were attributed to Nile perch. Biomass was estimated by converting classified backscatter using a TS-to-mass conversion factor, revealing lower estimates for haplochromines, Dagaa, and Nile perch compared to conventional acoustic methods. This underestimation may partly reflect exclusion of non-schooling haplochromines. The approach improves acoustic classification resolution and provides a more robust framework for biomass estimation in Lake Victoria to support ecosystem-based management.