首页|期刊导航|四川环境|基于机器学习的多源卫星遥感成都地表温度降尺度研究

基于机器学习的多源卫星遥感成都地表温度降尺度研究OA

Research on Downscaling of Land Surface Temperature in Chengdu Using Multi-Satellite Remote Sensing Based on Machine Learning

中文摘要英文摘要

地表温度是最重要的环境参数之一,能较好地反映陆面过程.为更好地开展成都城市热环境等相关研究,研究融合Landsat-8/9高空间分辨率(30 m)数据与MODIS高时间分辨率(逐日)数据,采用随机森林机器学习降尺度方法,构建成都市地表温度降尺度模型,得到兼具高空间-高时间分辨率的地表温度数据集,并与MODIS卫星地表温度(LST)产品对比,结果表明:首先采用单窗分裂法反演Landsat-8/9地表温度,该步骤能得到低时间-高空间分辨率的LST,与MODIS的LST产品相关性R2可达0.73,为后续降尺度研究奠定可靠的数据基础;其次,基于双因子机器学习构建地表温度降尺度模型,可以得到研究区100 m分辨率的逐日地表温度数据,平均绝对误差和均方根误差分别为0.932、1.197,整体效果较好.相较于传统单一卫星1 000 m空间分辨率或16 d时间分辨率的LST数据,该结果能准确表达每日地表温度的空间分布特征,纹理精度相较MODIS地表温度产品显著提高,并且能一定程度排除云干扰,为成都城市热效应动态精细化监测与评估提供支撑.

Land surface temperature(LST)is one of the most critical environmental parameters and effectively reflects land surface processes.To better support research on the urban thermal environment of Chengdu and related studies,high spatial resolution(30 m)from Landsat-8/9 data was integrated with high temporal resolution(daily)data from MODIS,and the random forest machine learning downscaling method was adopted in this study to construct a land surface temperature(LST)downscaling model for Chengdu.An LST dataset with both high spatial and high temporal resolution was generated and compared with the MODIS LST product.The results showed the following:First,the split-window algorithm was used to retrieve LST from Landsat-8/9 imagery.This step yielded LST data with low temporal but high spatial resolution,and the correlation with the MODIS LST product reached an R2 of 0.73,laying a reliable data foundation for the subsequent downscaling study.Second,a dual-factor machine learning-based LST downscaling model was constructed,producing daily LST data at 100 m resolution over the study area.The mean absolute error(MAE)and root mean square error(RMSE)were 0.932 and 1.197,respectively,indicating satisfactory overall performance.Compared with conventional LST data from a single satellite,which typically offer either 1 000 m spatial resolution or 16-day temporal resolution,the downscaled results accurately capture the spatial distribution characteristics of daily LST,exhibit markedly improved textural detail relative to the MODIS LST product,and are able to mitigate cloud interference to a certain extent.This provides support for the dynamic,fine-scale monitoring and assessment of the urban heat effect in Chengdu.

李晓敏;潘媞;冯晓;王晨曦

四川省气象数据中心,四川成都 610072四川省气象服务中心,四川成都 610072四川省气候中心,四川成都 610072四川省气候中心,四川成都 610072

资源环境

多源卫星遥感地表温度降尺度机器学习随机森林

Multi-satellite remote sensingland surface temperature downscalingmachine learningrandom forest

《四川环境》 2026 (4)

31-40,10

10.14034/j.cnki.schj.2026.04.005

评论