Multifactorial Evolutionary Algorithm based on Population Evolvability for Side-scan Sonar Image Segmentation

Abstract Evolutionary multitask optimization (EMTO) seeks to solve multiple optimization problems simultaneously by exploiting latent relationships between tasks. Among existing EMTO approaches, the Multifactorial Evolutionary Algorithm (MFEA) has shown considerable promise; however, its performance is highly dependent on task relatedness and may deteriorate when negative transfer occurs between dissimilar tasks. To address this limitation, this paper proposes MFEA-PE (Multifactorial Evolutionary Algorithm with Population Evolvability), a new EMTO framework that uses population evolvability derived from dynamic fitness landscape analysis to regulate inter-task knowledge transfer. The proposed method introduces an evolvability-guided asymmetric random mating probability (RMP) matrix and a direction-aware crossover strategy that adaptively controls both the intensity and direction of information exchange during the evolutionary search. The effectiveness of MFEA-PE is evaluated on nineteen benchmark multitasking optimization problems and compared with six representative algorithms: MFEA, MFEA-II, SOEA, MFEA-AKT, MFEA-DGD, and EMTO-AI. Among the seven algorithms, MFEA-PE obtains the second-best Friedman mean rank of 2.11 on the first benchmark suite and the best mean rank of 2.10 on the second benchmark suite. The pairwise Wilcoxon tests demonstrate clear improvements over the canonical MFEA and SOEA baselines in multiple cases, while showing that MFEA-PE remains statistically competitive with the three recent knowledge-transfer methods in most comparisons. Moreover, its cumulative runtime is comparable to the recent methods on the first benchmark suite and approximately 4.2%–6.9% lower on the second benchmark suite. The practical applicability of MFEA-PE is further demonstrated through multitask data clustering and side-scan sonar image segmentation. Experimental results show that MFEA-PE improves convergence speed by an average of 18.7% and solution quality by 12.4% compared with the canonical MFEA, while having less runtime overhead relative to MFEA-II. The proposed framework also demonstrates competitive performance in multitask clustering, particularly on heterogeneous task pairs where adaptive transfer is most beneficial. To investigate its practical applicability, MFEA-PE is further applied to side-scan sonar image segmentation formulated as a multitask clustering problem. On fifteen challenging SSS segmentation tasks, the method achieves a Dice coefficient of 0.84 ± 0.05, a Jaccard index of 0.82 ± 0.06, and a Boundary F1 score of 0.89 ± 0.04, outperforming representative segmentation baselines, including U-NET, Graph-Cut, Differential Evolution, and Particle Swarm Optimization. In addition, MFEA-PE operates 28% faster than U-NET and maintains robust performance under noisy imaging conditions, yielding accuracy improvements on degraded sonar imagery. These results demonstrate that evolvability-guided knowledge transfer provides an effective mechanism for reducing negative transfer and improving search efficiency. The proposed MFEA-PE framework therefore offers a computationally efficient and practically viable solution for both benchmark multitask optimization and real-world side-scan sonar image segmentation problems.