Change detection in remote sensing images based on multi-tree genetic programming

Abstract

Change detection in remote sensing images plays a crucial role in applications such as environmental monitoring, urban planning, and disaster management. Accurately identifying and distinguishing changed areas within complex image data poses significant challenges. Existing methods often struggle with high false-positive rates and limited adaptability. This paper introduces a novel approach using multi-tree genetic programming (GP) to automate the construction of ensembles for change detection in remote sensing images. The method employs a unique multi-tree GP representation comprising three distinct trees that utilize difference, spectral, and texture features to identify changes. These trees are combined into an ensemble using a majority voting strategy to make predictions. The approach integrates multi-tree crossover and mutation strategies to generate new individuals, which are evaluated based on a fitness function derived from classification accuracy. To validate its effectiveness, the proposed multi-tree GP approach is evaluated on four benchmark datasets (SZTAKI, EGY_BCD, LEVIR_CD+, and S2Looking) and compared with eight methods. In most cases, the proposed approach achieves higher maximum change detection accuracy. Notably, on the SZTAKI dataset (Img_10), it achieves an accuracy of 96.11%, representing a 5.55% improvement over the worst baseline (KNN) and a 0.55% gain over the best baseline (SpectralFormer). Experimental results demonstrate that the proposed approach outperforms standard GP, as well as several classic classifiers and neural network based methods, establishing it as an effective tool for remote sensing change detection. The method’s capability of to leverage diverse features and integrate them through ensemble learning underscores its potential in enhancing change detection accuracy using remote sensing imagery.