Abstract
This paper proposes an adaptive multi-threshold image segmentation method named IBKA-OTSU to address the limitations of existing deep learning-based image segmentation methods, particularly their heavy reliance on large-scale annotated datasets and high computational complexity. The proposed algorithm significantly enhances the capability of complex remote sensing scenarios by systematic improvements to core algorithm components, including population initialization strategy, attack behavior patterns, migration mechanisms, and opposition-based learning strategy. The improved intelligent optimization algorithm is innovatively integrated with the OTSU threshold method to establish a multi-threshold segmentation model specifically designed for remote sensing imagery. Experimental validation using representative samples from the ISPRS Potsdam benchmark dataset demonstrates that our IBKA-optimized OTSU multi-threshold segmentation method outperforms traditional IBKA-optimized pulse coupled neural network (PCNN) approaches in remote sensing image analysis. Quantitative evaluations reveal substantial improvements in the dice coefficient across six randomly selected remote sensing images, achieving performance enhancements of 7.76%, 11.99%, 30.75%, 22.91%, 44.37%, and 18.55%, respectively. This research provides an effective technical solution for intelligently interpreting remote sensing imagery in resource-constrained environments, demonstrating significant theoretical value and practical application potential in engineering implementations.