首页|结合贝叶斯特征选择的近海水深反演方法研究

结合贝叶斯特征选择的近海水深反演方法研究

扫码查看
针对现有近海水深数据获取周期长、危险性大,难以满足大范围近海水深数据高时间分辨率测量的问题.本文以辽东湾地区为研究区域,提出了一种结合贝叶斯特征选择的改进LightGBM近海水深反演方法.该方法首先依据贝叶斯特征选择方法对13个波段及波段比值特征进行优选;其次基于地理空间加权模块计算输入变量的权重;最后基于LightGBM反演模块对水深数据进行反演研究.本文选用辽东湾200个不同空间位置的数据样本作为测试集,验证本文方法和LightGBM模型在近海水深反演任务中的精度.结果表明,本文方法的皮尔森相关系数(r)为0.946,均方根误差(RMSE)为0.265 m,偏差(bias²)为0.017,平均绝对百分比误差(MAPE)为0.031.反演精度及稳定性均优于经典LightGBM模型,能够适应于近海水深反演.
Research on inshore depth inversion method combined with Bayesian feature selection
Aiming at the problems of long acquisition period and high risk of existing offshore water depth data,it is difficult to meet the high time resolution measurement of large-scale offshore water depth data.Taking the Liaodong Bay area as the research area,this paper proposes an improved LightGBM offshore water depth inversion method combined with Bayesian feature selection.Firstly,13 bands and band ratio features are selected according to Bayesian feature selection method.Secondly,the weight of input variables is calculated based on the geospatial weighting module.Finally,based on the LightGBM inversion module,the inversion of water depth data is studied.In this paper,200 data samples from different spatial locations in Liaodong Bay are selected as the test set to verify the accuracy of the method and LightGBM model in the offshore water depth inversion task.The results show that the Pearson correlation coefficient(r)value of this method is 0.946,the root mean square error(RMSE)is 0.265 meters,the deviation bias²is 0.017 and the mean absolute percentage error MAPE is 0.031.The inversion accuracy and stability are better than the Classic LightGBM model,which can be applied to the inversion of offshore water depth.

water depth inversionBayesian feature selectionLiaodong BayLightGBM

蓝歆玫、李佳、叶杨、朱洪波、薛国坤

展开 >

辽宁工程技术大学 测绘与地理科学学院,辽宁 阜新 123000

辽宁工程技术大学 地理空间信息服务协同创新研究院,辽宁 阜新 123000

大连黄渤海海洋测绘数据信息有限公司,辽宁 大连 116000

水深反演 贝叶斯特征选择 辽东湾 LightGBM

国家自然科学基金

42071343

2024

海洋通报
国家海洋信息中心 国家海洋局北海分局 国家海洋局东海分局 国家海洋局南海分局

海洋通报

CSTPCD北大核心
影响因子:1.07
ISSN:1001-6392
年,卷(期):2024.43(4)
  • 27