基于卷积神经网络的土壤水分反演研究OA
Research on Soil Moisture Retrieval Based on Convolutional Neural Network
研究基于 Sentinel-1 SAR 与 Sentinel-2 光学影像数据,经水云模型校正后消除植被冠层干扰,构建四维时空特征向量并输入卷积神经网络模型,实现了对 0～20 cm 表层土壤水分反演.结果表明:①模型反演精度较高,在测试集上 PCC=0.75,NRMSE=0.24,SMAPE=11.40％;②空间上,西北部地区土壤含水量整体上较东南部地区高;时间上,2024 年 10 月至 2025 年 1 月土壤含水量整体呈下降趋势,但受 2024 年 11 月灌溉事件和 2025 年 2 月积雪融水影响,部分站点有所回升.研究提出的卷积神经网络模型,有效提高了云贵高原覆盖区土壤含水量反演精度,可运用于精细化农业干旱预警、水资源管理中.
This study utilizes Sentinel-1 SAR and Sentinel-2 optical imagery data.After correcting for vegetation canopy interference using the Water Cloud Model,a 4D spatiotemporal feature vector was constructed and input into a Convolutional Neural Network(CNN)model to retrieve surface soil moisture(0-20 cm depth).The results indicate:1)The model achieved high retrieval accuracy on the test set,with PCC=0.75,NRMSE=0.24,and SMAPE=11.40％;2)Spatially,SM levels in the northwestern region were generally higher than those in the southeastern region.Temporally,overall SM exhibited a declining trend from October 2024 to January 2025,although localized increases occurred at specific monitoring sites due to an irrigation event in November 2024 and snowmelt in February 2025.The proposed CNN model effectively enhances the accuracy of soil moisture retrieval over the Yunnan-Guizhou Plateau region and holds significant potential for application in refined agricultural drought early warning and water resource management.
黄浩然;易珍言;刘蛟;李雪
西南科技大学环境与资源学院,四川 绵阳 621010西南科技大学环境与资源学院,四川 绵阳 621010西南科技大学环境与资源学院,四川 绵阳 621010西南科技大学环境与资源学院,四川 绵阳 621010
天文与地球科学
土壤水分反演卷积神经网络时空特征水云模型
soil moisture retrievalconvolutional neural networkspatiotemporal featurewater cloud model
《云南水力发电》 2026 (7)
24-27,4
西南科技大学素质类教改(青年发展研究)专项(22SZJG15)西南科技大学人才引进项目(21zx7154)
评论