Target Identification And Classification Technology For Massive Heterogeneous Remote Sensing Data

The systematization of remote sensing series of satellites not only makes the acquisition of multi-source satellite remote sensing image data information gradually mature,but also promotes the further development of remote sensing image data feature analysis technology.In recent years,the continuous refinement of deep learning methods has provided the research basis for the rapid development of information interpretation technology of remote sensing image data,and this method has been widely used in the fields of military target identification and classification,crop area identification and prediction,urban planning analysis,disaster area monitoring,etc.Under the requirements of the new era of modern military with strong science and technology,it is of great research significance and prospect to extract the target features in remote sensing image data efficiently and accurately.In this paper,using the massive heterogeneous remote sensing images as the data base,a detection method with high efficiency and accurate results in recognition and classification is proposed for different feature target data.The main research contents of the paper are four points as follows.First,for the massive image data features,a Gaussian image pyramid is constructed,and the feature libraries of different levels within the pyramid are selected for different resolutions of the data to be classified,which is a great improvement to improve the target recognition and classification efficiency;for the heterogeneous image data features,various pre-processing methods of denoising and enhancement are proposed,and in order to effectively improve the accuracy of subsequent remote sensing image target recognition and classification,the pre-processing methods are visually selected through the GUI interface.The pre-processing methods are selected through the GUI interface in a visual manner.Next,the preprocessed images are used as the data input source,and a multilayer convolutional neural network with a fused attention mechanism is built to ensure that the training accuracy reaches the optimal value for comparison experiments.The Conv layer extracts the feature maps of the target features in the data,gradually searches for the target of interest in the data,and the MaxPooling layer extracts the feature maps of the target of interest and normalizes them.Finally,the corresponding features of the target of interest are passed to the fully connected layer,and the excitation functions are defined and passed down to the Softmax classifier and the Boxbounding regressor,respectively,to obtain the classification result of the current target of interest and its coordinate position frame.Again,combining the multilayer deep convolutional neural network and the specific labeling coding method,a fusion method of target object features for multiple objects and multiple sizes of data in the same place and an improved neural network fusion method for heterogeneous remote sensing image data in the same region are proposed.In the former,the regional and global features of the target of interest are extracted by the constructed feature extraction network,processed by the feedback and reconstruction layers,and finally input to the fully connected layer for feature vector integration;in the latter,an improved neural fusion network with the introduction of small convolutional kernel and L2 regularization is constructed to effectively fuse the training features of same-area data acquired by different loads.Finally,a multilayer convolutional neural network with improved Faster-RCNN algorithm is proposed.The target trajectories are first extracted using SSIM structural similarity for continuous multi-frame point target data,and then input into the training framework for target feature extraction;the target features are extracted directly for high-resolution single-frame salient target data;the target recognition and classification results are analyzed and summarized.