| Snow cover affects climate and environmental changes,and snow cover area and fractional snow cover are input factors for many hydrological and climate models.Monitoring the change of snow cover area can calculate the melting amount and melting speed of snow,and monitoring the change of snow cover can summarize the change pattern and distribution state of snow,so as to achieve the purpose of dynamic monitoring of ecology,hydrology and climate change.Optical remote sensing is the main technical means of snow monitoring,among which the Moderate-resolution Imaging Spectroradiometer(MODIS)with a spatial resolution of 500 m is widely used for snow monitoring.However,due to the low spatial resolution of MODIS data,the phenomenon of the mixed end members can lead to a decrease in snow cover monitoring accuracy.Although a large number of research results have been achieved for global and regional snow cover monitoring and related algorithms for snow products,there are still the following problems to be solved to achieve continuous,highly accurate snow cover monitoring with high spatial and temporal resolution:(1)The representative spatiotemporal fusion method of multisource remote sensing data can provide snow remote sensing images with high spatiotemporal resolution,but the existing spatiotemporal fusion results exist obvious phenomenon of spectral distortion when snowfall or snowmelt occurs during the observation period,which affects the quality of the spatiotemporal fusion image.(2)Snow monitoring based on optical remote sensing data is susceptible to cloud occlusion,and the accuracy of existing cloud-snow detection methods needs further improvement when clouds and snow coexist.(3)The linear regression model based on snow index is commonly used to invert fractional snow cover(FSC),and the snow index should be obtained by considering how to use all the spectral band information to fully extract snow features,and the inversion accuracy of fractional snow cover still needs to be improved.With the rapid development of deep learning technology,deep learning is widely used in remote sensing field for its powerful learning ability,feature extraction ability and the ability to deal with nonlinear mapping relationships.This thesis combines deep learning technology and spatiotemporal fusion technology to achieve highly accurate and high spatiotemporal resolution snow cover monitoring in the observation area by using the proposed method of the improved spatiotemporal fusion algorithm for monitoring daily snow cover changes with high spatial resolution,the method of cloud snow detection based on hybrid feature network and the method of snow cover inversion based on deep feature snow index.The innovative research work of this thesis consists of the following three parts.(1)SMPG Spatiotemporal Fusion Model for Monitoring Daily Snow Cover Changes with High Spatial ResolutionTo achieve high precision and high spatial and temporal resolution snow cover monitoring,spatiotemporal fusion technology has been applied in the field of snow remote sensing.To address the problem that traditional spatiotemporal fusion methods are prone to have the problem of spectral distortion,an effective STDFA-Matching-Pix2pix-Generative Adversarial Network(SMPG)model combining the unmixing-based method,deep learning method,pre-matching and post-matching module is proposed to reduce the spectral distortion of fusion images and improve fusion images quality.The experimental results indicate that the proposed SMPG model can obtain high quality and high spatial and temporal resolution fusion images.Compared with the validation data,the correlation coefficient accuracy(CC)of the fusion image of SMPG model can reach 0.952.In addition,the Normalized Differential Snow Index(NDSI)of the fusion images can be calculated to obtain the day-by-day snow cover area with a spatial resolution of 30 m,and the error between this result and the validation data is less than 0.84%,which can realize the effective monitoring of snow cover area.(2)Cloud and Snow Detection Based on Hybrid Feature NetworkTo address the problem that the coexistence of clouds and snow affects the effective monitoring of snow,this thesis proposes a cloud and snow detection method based on hybrid feature network(CSD-HFnet).First,the local binary pattern(LBP),gray-level co-occurrence matrix(GLCM),and superpixel segmentation are used to extract the shallow features of cloud and snow.Then,the deep feature extraction network is proposed to extract the deep heatmap feature of cloud and snow.Finally,the original spectral bands,the shallow features,and the deep heatmap are fused to form hybrid features by a concatenation function,and the hybrid features are sent to a cloud and snow detection network equipped with a multi-feature long short-term memory(LSTM)network to extract the cloud and snow.The results indicate that CSD-HFnet with hybrid features can detect cloud and snow in multi-spectral remote sensing images and Red-Green-Blue(RGB)images of various spatial resolutions under the condition when cloud and snow coexist.The overall accuracy(OA)of cloud and snow detection obtained by CSD-HFnet can reach up to 95.35%.(3)Fractional Snow Cover Estimation Method Using Deep Feature Snow IndexTo address the problems of low spatial resolution of existing FSC products and the need to improve the accuracy of FSC inversion based on conventional snow index,this thesis proposes a fractional snow cover inversion method based on Deep Feature Snow Index(DFSI).Based on the all-band information of optical remote sensing images,this thesis firstly uses the proposed Light-wide Res Net model to obtain the DFSI,then establishes a linear regression fractional snow cover inversion model based on the DFSI,and finally obtains the fractional snow cover inversion results with higher accuracy.The experimental results of fractional snow cover inversion based on day-by-day spatiotemporal fusion images show that the R~2 of the proposed DFSI-based linear regression model on fractional snow cover inversion can reach 0.89,which is better than that of the conventional snow index.The RMSE between estimation result of DFSI-based linear regression model and reference FSC within 0.11.In summary,the SMPG model is firstly used to obtain multispectral fusion images of high spatial and temporal resolution,which provides the basis of high-quality data for snow cover monitoring.Secondly,the CSD-HFnet method is proposed to achieve high accuracy on cloud and snow detection under cloud-snow coexistence and effectively reduce the influence of clouds on snow monitoring.Finally,the proposed DFSI-based linear regression fractional snow cover inversion model is used to obtain high accuracy and high spatial and temporal resolution snow cover results for high quality and long time series snow cover monitoring at regional scale. |
PDF Full Text Request