Research On The Analysis And Compensation Methods For Geometric Positioning Errors Of Remote Sensing Images

Geometric positioning of remote sensing images is a scientific technique to locate the geographic space captured by stereo images based on the strict geometric relationship between the coordinates of each image point and its corresponding geographic space point in remote sensing images,which is the basis of remote sensing mapping information processing and the key technology to determine the accuracy of subsequent geographic information products.However,restricted by the embargo on the export of high-performance sensor equipment,there have been technical bottlenecks in the geometric positioning,and there is an urgent need to realize high-precision,high-robustness,and high-efficiency geometric positioning algorithms for remote sensing images under the conditions of limited data.In the paper,through the in-depth analysis of the classical model of geometric positioning and the detailed investigation of the current research situation at home and abroad,we sorted out the three key influencing factors in geometric positioning and their corresponding difficult problems: satellite attitude error,stereoscopic image matching error,and the systematic error of the bundle adjustment model.By exploring the mathematical characteristics of the three geometric positioning errors,the corresponding error compensation methods are designed using machine learning and deep learning,which effectively improve the accuracy of remote sensing geometric positioning,and realize the high-precision remote sensing image geometric positioning algorithm under the condition of no ground control points.The specific research contents are as follows:First,for the existing satellite attitude determination algorithms with nonlinear approximation errors,weak adaptability to data noise changes,and other problems,the paper proposes a satellite attitude compensation method based on the extreme learning machine.The method consists of three parts.1)Simplified adaptive extended Kalman filter module,which can achieve higher accuracy satellite attitude estimation under different prior accuracy conditions of the attitude sensors,and provide accurate initial values for the subsequent attitude fusion process.2)Extreme Learning Machine(ELM)-based satellite attitude error compensation network,which can accurately predict the approximation error of the linear filtering algorithm for a nonlinear system and output the attitude error compensation value,through the key filtering parameters in the simplified adaptive extended Kalman filter module.3)Adaptive weighted attitude fusion module,which adds a learnable weighted fusion factor when fusing the satellite attitude initial value and the attitude error compensation value,avoiding the reintroduction of attitude observation noise in the subsequent attitude smoothing process,to keep the final attitude determination result with high accuracy output.After the verification of satellite attitude data with different orbit parameters,the satellite attitude error compensation method proposed in the paper can detect the errors in the satellite attitude determination process and correctly compensate for them,thus providing more accurate input data,improving the accuracy of the subsequent geometric positioning model establishment,and realizing the accuracy improvement of the geometric positioning results.Subsequently,to address the problems of inaccurate modeling of stereo matching errors in existing remote sensing images and relying on manual methods for mismatch removal,the paper proposes a stereo matching error compensation method based on feature error metrics and confidence perception.The method consists of four parts.1)An attention-driven multiscale feature extractor,which can extract and mine the global depth feature of remote sensing multi-view images,significantly express the information differences between multi-view images,and provide an effective basis for the subsequent geometric positioning.2)A feature metric error perception module,which uses the multi-layer perceptron as the backbone network,and detects the confidence of the matched pairs obtained from the stereo matching process by comparing the feature differences between multi-view images,and predicts the corresponding pixel compensation values.3)Confidence-aware bundle adjustment,which assigns the matched pairs of points with weights according to the confidence and corrects the matched pairs using the pixel compensation values,so that the geometric positioning model has the capability of correcting the input data,thus realizing the high-precision 3D spatial point solving.4)RFM-based forward re-projection module,which enhances the detection and compensation ability of the proposed method for stereo-matching errors utilizing iterative updating.Through the verification of the remote sensing multi-view image dataset,the stereo matching error compensation method proposed in this paper can accurately detect and compensate for the matching error according to the input matching points,realizing the automatic correction of the stereo matching error in the geometric positioning process,and improving the geometric positioning accuracy based on accurate satellite attitude.Finally,for the problems of poor numerical stability and systematic errors of the existing bundle adjustment model,the paper proposes a systematic error compensation method based on a deep recurrent network.This method consists of three parts.1)A geometric and feature metrics joint extractor,which obtains the geometric information and features from the multi-view images and uses them as constraints to limit the solution domain of the parity,to improve the stability of the geometric positioning model.2)A correlation decoupled fusion metric constructor,which uses the correlation normalization to remove the magnitude difference between the geometric and feature information,and unifies the geometric positioning model to prevent it from obtaining the local optimal solution.3)Implicit recurrent optimizer for geometric positioning in remote sensing,which makes use of the property that deep recurrent networks can accurately fit arbitrary functions,reduces the systematic error of the original geometric positioning bundle adjustment model and improves the accuracy of the final 3D point clouds.In addition,this method avoids the process of large-scale matrix inverse in the traditional geometric positioning model,which improves the stability of the bundle adjustment model.On the remote sensing multi-view image datasets,the systematic error compensation method of the bundle adjustment model proposed in the paper further improves the accuracy of geometric positioning,which proves that it effectively eliminates the systematic error of the bundle adjustment model and achieves the stable solutions of geometric positioning under a variety of scenario conditions.In summary,the paper effectively solved the difficult problems in geometric positioning error compensation,improved the overall accuracy of geometric positioning of remote sensing images,and realized the high-precision geometric positioning method of remote sensing images without the support of ground control points,which can provide an innovative method for the realization of integrated and automated remote sensing image geometric positioning technology,and it can contribute innovative scientific research power to the business of China’s smart city and 3D real scene.