Method and device for dust identification utilizing multimodal neural network, and storage device
A method for dust identification includes following steps: collecting multi-source data related to dust; preprocessing the multi-source data to obtain processed data; constructing a training set for a dust identification model using the processed data; constructing the dust identification model, where the dust identification model includes a backbone network, an output network, and a fusion network; training the dust identification model based on the training set to obtain a final model; and identifying dust based on the final model. The dust identification method significantly improves a speed and accuracy of dust identification, and also partially improves continuity of the dust identification.
1 . A method for dust identification utilizing a multimodal neural network, comprising following steps:
S 1 : collecting multi-source data related to dust;
S 2 : preprocessing the multi-source data to obtain processed data;
S 3 : constructing a training set for a dust identification model using the processed data;
S 4 : constructing the dust identification model, wherein the dust identification model comprises a backbone network, an output network, and a fusion network;
S 5 : training the dust identification model based on the training set to obtain a final model; and
S 6 : identifying dust based on the final model;
wherein in the step S 4 , the backbone network has a U-net architecture; and the U-net outputs a confidence coefficient of a dust category to which each pixel belongs;
wherein the output network adopts an extreme Gradient Boosting (XGboost) tree; and the XGboost tree outputs a confidence coefficient indicating that a corresponding grid point of a to-be identified region is a dust region.
2 . The method according to claim 1 , wherein the multi-source data comprises: inversion data of an imaging spectrometer, data of a ground meteorological observation station, and image data of a dust event.
3 . The method according to claim 1 , wherein the preprocessing in the step S 2 specifically comprises: fusing the multi-source data to obtain fused data; and combining a normalized difference dust index and a thermal infrared dust index of the image data of the dust event to obtain a comprehensive dust distinguish index.
4 . The method according to claim 1 , wherein in the step S 3 , a process of constructing the training set of the dust identification model is as follows:
performing spectral analysis on a corresponding channel of the image data of the dust event to obtain a channel suitable for distinguishing dust; and
in the channel suitable for distinguishing the dust, manually marking corresponding image data of the dust event, distinguishing between a dust region and a non-dust region, and completing construction of the training set.
5 . The method according to claim 1 , wherein the fusion network adopts a Bayesian method to fuse an output of the U-net and an output of the XGboost tree to obtain a final result.
6 . The method according to claim 1 , wherein in the step S 5 , an Adam optimization algorithm is used to optimize model parameters during the training of the dust identification model.
7 . A storage device, wherein the storage device stores an instruction and data for implementing the method according to claim 1 .
8 . A dust identification device based on a multimodal neural network, comprising a processor and a storage device, wherein the processor loads and executes an instruction and data in the storage device to implement the method according to claim 1 .