Method Research And System Implementation Of Urban Impervious Surface Extraction Based On Multi-source Remote Sensing Data Fusion

Impervious surface is a type of surface with extremely low water permeability,such as various types of buildings,roads,squares,etc.Remote sensing technology can obtain information on the distribution and changes of impervious surfaces in a wide range of cities,which is helpful for urban planning and environmental monitoring related research and business development.However,the use of optical image or SAR image alone for impervious surface extraction has certain defects.How to reasonably use the effective information in optical and SAR images to extract impervious surfaces has become a hot issue in current research.Based on the optical remote sensing image and polarimetric SAR image in Hong Kong,this thesis conducts research on the feature selection and fusion algorithm of multi-source remote sensing fusion,including the following parts:First,an optical feature fusion classification algorithm based on Xgboost is proposed for the extraction of urban impervious surfaces from optical remote sensing images.According to the optical remote sensing imaging mechanism and spectral characteristics,17 kinds of optical features are extracted.After research and comparison,the ensemble learning algorithm shows better classification performance.The Xgboost algorithm with the highest accuracy is used to achieve the fusion classification of optical features.The final impervious surface extraction result pixel accuracy rate is97.43%,and the Kappa coefficient is 0.9416.Second,a GBDT-based SAR feature fusion classification algorithm is proposed for the extraction of urban impervious surfaces from SAR images.According to the scattering mechanism and imaging principle of polarized SAR,19 kinds of SAR features are extracted,and the GBDT algorithm is used to realize the fusion classification of SAR features.The final impervious surface extraction result pixel accuracy rate is 84.59%,and the Kappa coefficient is 0.6847.Third,this thesis proposes a feature evaluation system to provide a basis for feature selection for multi-source remote sensing image classification algorithms.We use Pearson correlation coefficient,JS divergence and random forest algorithm to systematically analyze the relevance,separability and contribution of optical and polarization SAR features to land use classification tasks,and select the optimal feature set for multi-source Remote sensing image classification.The feature-level fusion classification algorithm has a final pixel accuracy rate of 97.82% and a Kappa coefficient of 0.9545.Fourth,a conditional evidence coding algorithm is proposed to improve the multisource remote sensing image classification algorithm based on decision-level fusion.This method can make full use of the complementary information of optical and polarized SAR to improve the classification accuracy of multi-source remote sensing fusion algorithms.The final impervious surface extraction result pixel accuracy rate of the decision-level fusion algorithm is 98.06%,and the Kappa coefficient is 0.9588.Fifth,we designed and constructed an urban impervious surface extraction system based on multi-source remote sensing fusion.The reliability of each functional module is verified through system testing,which provides technical support for urban impervious surface monitoring,and has high practical application value.