摘要
高光谱降维是农作物遥感分类的关键环节,通过去除冗余信息保留有效特征信息,从而提升分类准确率。波段选择作为高光谱降维的一种核心手段,通常依赖于特定的优化算法和遥感任务来定量评估波段的重要性,并据此挑选出最佳波段子集。然而,现有的波段选择算法存在收敛速度慢、易陷入局部最优等问题。为此,提出了一种基于改进河马算法的高光谱波段选择算法,旨在为具有相似光谱特征的农作物精细分类提供一种高效的波段选择方法。首先,提出了一种离散化策略,并据此对HO算法进行了修正,从而形成了适用于波段选择的DHO算法。其次,分析了HO算法在波段选择中全局搜索效率不高和局部精细搜索能力不足的问题,并针对这些问题提出了自适应防御策略和非线性逃离策略,同时将这些策略与离散化策略相结合,形成了IDHO算法,以提升DHO算法在波段选择中的性能。最后,基于农作物精细分类的应用场景,以Jeffries-Matusita距离为核心变量创建适应度函数,并使用改进后的河马算法(DHO、IDHO)与其他元启发式优化算法(ACO、GA)进行波段选择,再将优选的波段子集输入到U-Net模型进行分割。结果表明,DHO和IDHO在收敛速度和精度方面均优于ACO和GA。特别是IDHO,在不显著增加计算成本的前提下,实现了比DHO更高的收敛精度,展现了其更强的优化能力。此外,由IDHO筛选出的波段子集在进行分类时也达到了更高的精度。
Hyperspectral dimensionality reduction is a crucial step in the remote sensing classification of crops,as it enhances classification accuracy by removing redundant information and retaining effective feature information.Band selection,a core method of hyperspectral dimensionality reduction,typically relies on specific optimization algorithms and remote sensing tasks to quantitatively evaluate the importance of bands and thereby select the optimal subset of bands.However,the existing band selection algorithms suffer from problems such as slow convergence speed and susceptibility to local optima.Therefore,this paper proposes a hyperspectral band selection algorithm based on an improved Hippopotamus Optimization algorithm,aiming to provide an efficient band selection method for the fine classification of crops with similar spectral characteristics.Firstly,a discretization strategy is proposed,leading to modifications of the HO algorithm to create the DHO algorithm tailored for band selection.Secondly,the issues of low global search efficiency and insufficient local detailed search capability of the HO algorithm in band selection are analyzed.To address these problems,adaptive defense and nonlinear escaping strategies were proposed.These strategies,in combination with the discretization approach,formed the IDHO algorithm,aiming to enhance the performance of the DHO algorithm in band selection.Finally,based on the application scenario of fine classification of crops,a fitness function is created with Jeffries-Matusita distance as the core variable.The improved Hippopotamus algorithm(DHO,IDHO)and other metaheuristic optimization algorithms(ACO,GA)are used for band selection.The optimal subset of bands is then input into the U-Net model for segmentation.The results show that DHO and IDHO are better than ACO and GA in terms of convergence speed and accuracy.Especially,IDHO achieves higher convergence accuracy than DHO without significantly increasing the computational cost,demonstrating its stronger optimization ability.Moreover,the bands selected by IDHO also achieve higher accuracy in classification.
作者
黄祥
王克晓
蒲晓君
周蕊
吴园
HUANG Xiang;WANG Kexiao;PU Xiaojun;ZHOU Rui;WU Yuan(Institute of Agricultural Science and Technology Information,Chongqing Academy of Agricultural Sciences,Chongqing 401329,China)
出处
《测绘地理信息》
2025年第5期84-91,共8页
Journal of Geomatics
基金
重庆市农业科学院市级财政科研项目(cqaas2023sjczqn007,cqaas2023sjczhx003)
重庆市技术创新与应用发展专项(CSTB2023TIAD-ZXX0049)。
关键词
河马优化算法
高光谱图像
波段选择
图像分割
Hippopotamus optimization algorithm
Hyperspectral image
Band selection
Image segmentation
