Research On Quantitative Simulation Of Hyperspectral Scenes And Targets Based On Deep Learning

Hyperspectral reflectance data can reflect the intrinsic properties of the corresponding target and improve human’s ability to recognize the world.Its huge application potential and value have attracted more and more attention in various fields.The application of hyperspectral imagery requires a large amount of data for a priori study.However,there are some factors such as high cost,difficulty in obtaining specific scenarios and so on,which lead to data shortage.To solve this problem,the important path is to obtain simulation data through quantitative simulation of scene targets.The existing hyperspectral simulation schemes.However,the existing hyperspectral simulation have some problems,such as cumbersome modeling,low accuracy,slow speed,single scene.This dissertation establishes a high-precision intelligent simulation model on the link of "Simulation of spatial structure→Simulation of spectrum characteristics" to solve the above problems.The key problems faced in the research process include: 1)There are shortage of data,low simulation accuracy,entanglement of features between subclasses,violation of scene targets and sexual fusion problems in intelligent spatial structure simulation.2)The absence of near infrared range,isochromatic spectrum and direct few-to-many mapping in the simulation of spectral characteristics lead to complex problems with low simulation accuracy.3)Difficulty in training spectral characteristics simulation model due to too deep network.The main research contents and innovations are as follows:1)Aiming at the shortage of spatial structure database,according to the primary and secondary characteristics of wing and propeller fuselage color on remote sensing satellites,this dissertation constructs a two-level 14-type aircraft model dataset.Furthermore,a joint pre embedded adaptive weighting condition is proposed to generate the spatial structure simulation model of the countermeasure network.Combined with the loss function containing semantic style,it solves the problems of feature entanglement between classes,low modeling accuracy,and inconsistent fusion of scene and target in hyperspectral spatial structure simulation.A large number of experiments have been carried out on three datasets.Experiments conducted on three datasets demonstrate that the simulation model mining FID of sample spatial capability is 7.5,7.86% greater than the sub-optimal algorithm,and the SSIM index of the scene target mapping capability to be 0.82,1.67% higher than the sub-optimal algorithm.Finally,according to the needs of users,intelligently simulate the target scenes with multiple postures,types and shadows of hyperspectral spatial dimension.2)To solve the problems of missing near-infrared range and low accuracy of metamerism simulation,panchromatic features were introduced as a priori and hyperspectral data(400～1000 nm)containing near-infrared were collected for simulation model learning.Furthermore,a simulation network and three-step training method of joint codec and multi-stage transmission module are proposed,which can obtain stronger point-to-point spectral space expression by converting the few to many spectral simulations into one-to-one mapping,and solve the problem of low accuracy caused by the direct few to many mappings of spectral characteristics simulation.Experiments were carried out on two public datasets and the self-built true and false target datasets,proving that the proposed simulation model and panchromatic prior can realize the high-precision spectral simulation of visible near-infrared(400～1000 nm),and the spectral accuracy on the self-built true and false target datasets averaged 92.62%,10.44% higher than the simulation accuracy without panchromatic prior,and 3.91%higher than the simulation accuracy of the suboptimal algorithm.3)To overcome the problem of large parameters and long training time of the simulation model of spectral characteristics caused by too deep network,starting from the principle of multispectral imaging,it is proposed to use external learnable vectors to simulate white balance,color mapping and gamma correction to realize lightweight design of network architecture.To improve the feature extraction capabilities of lightweight models,global and local non-linear features are obtained using Transformer modules and external dynamic vectors.The pressure test is carried out under a variety of lighting conditions.Compared with the most advanced 1.62 M lightweight model,the proposed lightweight spectral characteristics simulation model uses only one parameter of 3090 GPU and 0.17 M to achieve 89.48% optimal spectral accuracy,with an average improvement of 9.66%,and the training time is shortened by about 8%.This model can better support the high-precision simulation of a large number of hyperspectral remote sensing data with limited computing resources.