High-resolution Image Target Identification Based On Quantitative Remote Sensing And Deep Learning

At present,high resolution optical remote sensing small target identification methods are mostly inherited from deep learning models in the field of computer vision(Computer Vision,CV),but the rich quantitative information in remote sensing targets has not been mined and applied to deep learning models.Compared with ordinary RGB images,remote sensing images also contain unique quantitative information such as spectrum and radiation in addition to common features such as shape and size.Combining quantitative and image feature information to carry out target identification belongs to the category of remote sensing image fusion interpretation,which can be divided into three levels: pixel level,feature level and decision level.In order to fully combine the effective information of quantitative remote sensing and CV,and further improve the remote sensing target identification effect of high-resolution satellite images,in this paper,the quantitative processing method of atmospheric radiometric correction remote sensing information is taken as the entry point,and takes wind farm as an example.From the three aspects of remote sensing image fusion interpretation,the fusion target identification strategy of quantitative remote sensing information and deep learning model systematically is studied.The specific research work is as follows:(1)The QUantitative and Automatic Atmospheric Correction(QUAAC)link of highresolution images is constructed to obtain the quantitative information of high-resolution images,which provides a processing tool for the subsequent high-resolution target fusion identification of quantitative remote sensing and deep learning.By taking advantage of the aerosol sensitive spectral band and high temporal and spatial resolution of the new generation of stationary satellites,the atmospheric aerosol product inversion of stationary Himawari-8 satellite is used to provide the same space and near real-time atmospheric aerosol parameters for high-resolution images,so as to support the construction of automatic quantitative atmospheric correction link for high-resolution images.It has been proved that QUAAC can better restore the real spectral information of the surface,and the corrected spectral curve of ground objects is consistent with the measured spectral trend,which is better than the widely used ENVI/FLAASH results.(2)A fusion recognition strategy based on depthology model is proposed to realize target recognition in high-resolution surface reflectance images.When QUAAC is introduced into the GF-2 image preprocessing process,the model recognition accuracy is increased from 90.5% to92.7%,and the improvement effect is more obvious.The input of quantitative remote sensing information such as spectrum is increased to the network,and the pixel-level fusion of quantitative remote sensing and deep learning is promoted.Based on the fan data set processed by quantitative atmospheric correction,CBAM attention mechanism is introduced into the feature pyramid part of YOLOv5 model,so that the network can give more weight to the fan target.The accuracy of image recognition increased from 92.7% to 93.2%,and the fusion improvement effect at the feature level was limited.(3)Based on the QUAAC-corrected surface reflectance fan data set and the YOLOv5-CBAM network model,a multi-feature accurate recognition method based on quantitative remote sensing information is proposed.According to the preliminary results of model recognition,the false detection objects of low confidence targets are eliminated by using the quantitative spectrum and image feature information of high confidence fan targets.Before discriminating the low confidence target threshold,the threshold range will vary according to the fan characteristic information in different images.It has been proved that the correct recognition rate increases from 93.2% to97.4%,and the false alarm rate decreases from 6.1% to 1.6%.This method shows excellent fusion recognition effect at the decision level.Aiming at optical high-resolution remote sensing target identification,the cross research of CV and remote sensing is carried out.In this paper,a fusion identification method strategy of quantitative remote sensing and deep learning from multiple aspects of remote sensing image fusion interpretation is provided,and the positive enhancement effect of quantitative remote sensing information is revealed on deep learning methods,which provides ideas for other optical remote sensing target identification research.