| Interferometric synthetic aperture radar measurement technology is an important means of earth observation,which is widely used in terrain mapping,disaster rescue,ocean current monitoring,and other fields.Interferometric phase filtering and unwrapping,as key steps in interferometric phase processing,are affected by phase noise,and it is usually difficult to achieve a balance between processing accuracy and speed.In recent years,deep learning methods represented by convolutional neural networks have been applied to interferometric phase processing and achieved remarkable performance,but limited by the size of the convolution receptive field,their ability to extract global features is weak,and the self-similarity of the interference fringes is not fully exploited,which affects the measurement accuracy of the interferometric synthetic aperture radar.To this end,this thesis conducts research on interferometric phase filtering and unwrapping based on deep learning.the main contents and innovations of this thesis are as follows:（1）The principle of synthetic aperture radar interferometry is briefly described,and several typical traditional interferometric phase filtering and unwrapping methods are studied.The realization principle and process of each method are described in detail,and the advantages and disadvantages of each method are analyzed.The application of deep learning methods in interferometric phase filtering and unwrapping is introduced,paving the way for the comparison of methods in the following sections.（2）An interferometric phase filtering method based on Transformer network is proposed.In view of the self-similarity of the interferometric phase map,and considering the advantages of the Transformer network in long-distance dependency modeling or nonlocal feature information aggregation,the Transformer network is applied to the interferometric phase filtering task,and an interferometric method based on the Transformer network is proposed.Aiming at the variability of the interference phase fringes and the possible shortcomings of the Transformer network in local features extraction,a residual deformation convolution module is built,which can adaptively learn the phase shape.By introducing it into the Transformer network,further enhance the network’s ability to extract phase local features.On the simulation data,compared with three traditional and one deep learning-based phase filtering methods,the method proposed in this thesis reduces the mean square error by 23%～56% and the number of residual points by 56%～94%.On the measured data,the method in this thesis can better suppress the residual point and restore the interference phase details in the dense fringe area.（3）A joint interferometric phase filtering and unwrapping method based on Transformer-CNN network is studied.In traditional and deep learning-based methods,phase filtering and unwrapping are usually performed in two stages or two networks successively,which may lead to error accumulation and high time cost.In view of the above problems,this thesis is based on the Transformer network,adopts the hard parameter sharing method in multi-task learning,and completes phase filtering and wrapping number estimation simultaneously.And in order to balance the training weights between different tasks,the uncertainty loss weighting method is used to let the network automatically adjust the loss weights between the phase filtering and phase wrapping number estimation tasks,improving the prediction accuracy of the network.On the simulated data,compared with three traditional and one deep learning-based phase unwrapping methods,the method in this thesis reduces the root mean square error by58%～86%,and the total processing time was reduced by 16% to 87%.In the measured data,the method in this thesis can obtain better unwrapping results under the condition of strong phase noise,and the rewrapping phase is closer to the original wrapping phase. |
