Hyperspectral data processing method based on deep learning

Abstract

The invention relates to the field of spectrum detection data processing, in particular to a hyperspectral data processing method based on deep learning, which comprises the following steps: step 1, acquiring a hyperspectral signal of a detection sample, and preprocessing the hyperspectral signal to obtain a plurality of effective pixels n; 2, each effective pixel n obtained in the step 1 is learned in a learning model, and each effective pixel n obtains a corresponding depth feature; step 3, performing Fourier transform on the plurality of depth features obtained in the step 2, and mixing each transformed depth feature with the hyperspectral signal to obtain a plurality of mixed signals; and 4, inputting all the mixed signals obtained in the step 3 into the random forest model, and identifying and detecting the detection sample. According to the invention, the accuracy, high efficiency and rapidity of a hyperspectral detection technology are improved.

Description

Technical Field

[0001] The present invention relates to the field of spectral detection data processing, and in particular to a hyperspectral data processing method based on deep learning. Background Art

[0002] Hyperspectral detection technology is an emerging detection technology. It obtains rich, detailed and continuous hyperspectral data in the ultraviolet, visible and infrared bands of the electromagnetic spectrum. The detection system based on this technology can easily obtain the spectral information of the test sample, and then obtain the spatial distribution map of the chemical characteristics of the test sample. However, the rich information imaging of hyperspectral also leads to difficulties in hyperspectral data processing. Although with the development of computer technology, the processing capability of hyperspectral data has been further improved, especially after the implementation of deep learning methods such as convolutional neural networks, deep learning methods rely to a large extent on the availability of a large amount of high-quality data, and in the process of spectral data processing, there are still well-known problems such as the large amount of information in spectral images and the presence of a large amount of redundant multicollinear information, which affects the accuracy, efficiency and speed of hyperspectral detection technology. Ensuring the accuracy and real-time performance of hyperspectral detection technology has always been an urgent problem to be solved by workers in the field of hyperspectral detection data processing.

[0003] In view of this, the inventors provide a hyperspectral data processing method based on deep learning to solve the above technical problems. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the above-mentioned prior art and propose a hyperspectral data processing method based on deep learning, which improves the accuracy, efficiency and speed of hyperspectral detection technology.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] The present invention provides a hyperspectral data processing method based on deep learning, comprising the following steps:

[0007] Step 1: obtaining a hyperspectral signal of a detection sample, and preprocessing the hyperspectral signal to obtain a plurality of effective pixels n;

[0008] Step 2: Each valid pixel n obtained in step 1 is learned in the learning model, and each valid pixel n obtains a corresponding depth feature;

[0009] Step 3, performing Fourier transform on the multiple depth features obtained in step 2, and connecting each transformed depth feature with the hyperspectral signal to obtain multiple mixed signals;

[0010] Step 4: Input all the multiple mixed signals obtained in step 3 into the random forest model to identify and detect the test samples.

[0011] Furthermore, in step 1, the preprocessing includes using an adaptive filter to adaptively learn each band information in the hyperspectral signal to obtain the effective pixel n contained in the effective band information.

[0012] Furthermore, in step 2, the learning model is a space-spectrum joint learning model, and the space-spectrum joint learning model includes a spatial deep learning model channel, a spectral deep learning model channel, a feature extraction layer and a joint learning layer.

[0013] Furthermore, each effective pixel n includes a spectral curve S n and the spatial neighborhood P n ,In step 2, the deep feature extraction process includes the following steps:

[0014] Step 2.1: The spectral curve S n After the spectral deep learning model channel and feature extraction layer, the spectral feature F1 (S n );

[0015] Step 2.2: The spatial neighborhood P n After the spatial deep learning model channel and feature extraction layer, the spatial feature F2 (P n );

[0016] Step 2.3, the spectral feature F1 (S n ) and spatial features F2(P n )The deep features are obtained through the joint learning layer.

[0017] Furthermore, in step 2.3, the joint learning layer includes a calculation formula for deep features, which is specifically as follows:

[0018] F n =f{W·(F1(S n )☉F2(P n ))+b}

[0019] Among them, W represents the weight; ⊙ represents the cascade operation; f{n} represents the nonlinear transformation of the calculation result; b represents the bias term, which is the adjustment coefficient.

[0020] Furthermore, in step 3, the transformed depth feature and the hyperspectral signal are connected through residual connection to obtain a mixed signal.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] The invention provides a method for processing hyperspectral data based on deep learning, which uses an adaptive filter to preprocess the spectral signal, can adaptively learn the effective band of the spectral signal, improve the readiness of feature extraction, effectively avoid processing a large amount of redundant multicollinear information, and improve the processing rate and accuracy; it also uses a spatial-spectral joint deep learning model to eliminate interference factors in the spectral data, improve the clarity and reliability of the data, and improve the robustness of data processing. The hyperspectral data processing method of the invention can provide a basis for real-time and accurate on-site detection of hyperspectral data, and the processing method of the invention can also be extended to other spectral detections, such as Raman detection, infrared spectral detection, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The accompanying drawings are incorporated in and constitute a part of this specification and, together with the description, serve to explain the principles of the present invention.

[0024] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0025] Figure 1 This is a flow chart of the hyperspectral data processing method based on deep learning of the present invention;

[0026] Figure 2 This is a schematic diagram of hyperspectral data processing based on deep learning in the present invention. DETAILED DESCRIPTION

[0027] Here, exemplary embodiments will be described in detail, and the embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Instead, they are only examples consistent with some aspects of the present invention described in detail in the appended claims.

[0028] See also Figures 1-2 The present invention provides a hyperspectral data processing method based on deep learning, comprising the following steps:

[0029] Step 1, select pesticides in spinach as the target detection object, wash and dry the commercially purchased spinach, and evenly spray pesticides on the surface of the spinach as a detection sample, use a hyperspectral detector to collect hyperspectral data of the detection sample, obtain the hyperspectral signal of the tested sample, and use an adaptive filter to adaptively learn the information of each band in the spectral signal to obtain the effective pixel n in the effective band information;

[0030] Step 2: Each valid pixel n obtained in step 1 is learned in the learning model, and each valid pixel n obtains a corresponding depth feature;

[0031] Among them, the learning model is a space-spectrum joint learning model, and the space-spectrum joint learning model includes a spatial deep learning model channel, a spectral deep learning model channel, a feature extraction layer and a joint learning layer.

[0032] Step 3: The spectral curve S of each effective pixel n n After the spectral deep learning model channel and feature extraction layer, the spectral feature F1 (S n );

[0033] Step 4: The spatial neighborhood P of each valid pixel n n After the spatial deep learning model channel and feature extraction layer, the spatial feature F2 (P n );

[0034] Step 5: The spectral feature F1(S n ) and spatial features F2(P n ) The deep features are obtained through the joint learning layer;

[0035] Among them, the joint learning layer includes the calculation formula of deep features, which is as follows:

[0036] F n =f{W·(F1(S n )☉F2(P n ))+b}

[0037] Among them, W represents the weight, which is obtained for each effective pixel n during the training process of the spatial-spectral joint learning model; ⊙ represents the cascade operation; f{n} represents the nonlinear transformation of the calculation result; b represents the bias term, which is the adjustment coefficient.

[0038] Step 6: transform the multiple deep features obtained in step 5 into the frequency domain by Fourier transform, and obtain multiple mixed signals by residual connection with the hyperspectral signal obtained in step 1;

[0039] Step 7: Input all the multiple mixed signals obtained in step 6 into the random forest model, complete the identification and detection of the test samples through the feature decision process, and achieve accurate detection of pesticides in spinach.

[0040] In order to illustrate the effect of the present invention, the present invention carried out the following experiment:

[0041] In this experiment, imidacloprid pesticide on spinach, dichlorvos pesticide on spinach, and heavy metal arsenic on spinach were used as target detection objects, and conventional processing methods and the processing methods of the present invention were selected. The conventional processing methods included: residual network + fully connected model (model one), random forest classification (model two), noise reduction support vector + residual network + dual attention network (model three); the processing method of the present invention was: adaptive filtering + spatial-spectral joint learning model + Fourier transform + residual connection + random forest classification (model four). The data processing method was trained with the same set of spectral data to verify the accuracy and efficiency of the data processing method proposed by the present invention.

[0042] The experimental results are shown in Table 1. It can be seen from Table 1 that the accuracy of the prediction results of Model 1, Model 2, and Model 3 are all less than 90%, while the accuracy of the prediction results of Model 4 is greater than 90%.

[0043] Table 1 Comparison of the detection results of spinach by various data processing methods

[0044]

[0045] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention.

[0046] It should be understood that the present invention is not limited to what has been described above and that various modifications and changes may be made without departing from its scope. The scope of the present invention is limited only by the appended claims.

Claims

1. A hyperspectral data processing method based on deep learning, characterized in that: The following steps are involved: Step 1: obtaining a hyperspectral signal of a detection sample, and preprocessing the hyperspectral signal to obtain a plurality of effective pixels n; Step 2: Each valid pixel n obtained in step 1 is learned in the learning model, and each valid pixel n obtains a corresponding depth feature; Step 3, performing Fourier transform on the multiple depth features obtained in step 2, and mixing each transformed depth feature with the hyperspectral signal to obtain multiple mixed signals; Step 4: Input all the multiple mixed signals obtained in step 3 into the random forest model to identify and detect the test samples.

2. The method for processing hyperspectral data based on deep learning according to claim 1, characterized in that: In step 1, the preprocessing includes using an adaptive filter to adaptively learn each band information in the hyperspectral signal to obtain the effective pixel n contained in the effective band information.

3. The method for processing hyperspectral data based on deep learning according to claim 1, characterized in that: In step 2, the learning model is a space-spectrum joint learning model, and the space-spectrum joint learning model includes a spatial deep learning model channel, a spectral deep learning model channel, a feature extraction layer and a joint learning layer.

4. The method for processing hyperspectral data based on deep learning according to claim 3, characterized in that: Each effective pixel n includes a spectral curve S n and the spatial neighborhood P n ,In step 2, the deep feature extraction process includes the following steps: Step 2.1: The spectral curve S n After the spectral deep learning model channel and feature extraction layer, the spectral feature F1 (S n ); Step 2.2: The spatial neighborhood P n After the spatial deep learning model channel and feature extraction layer, the spatial feature F2 (P n ); Step 2.3, the spectral feature F1 (S n ) and spatial features F2(P n )The deep features are obtained through the joint learning layer.

5. The method for processing hyperspectral data based on deep learning according to claim 4, characterized in that: In step 2.3, the joint learning layer includes a calculation formula for deep features, which is as follows: F n =f{W·(F1(S n )☉F2(P n ))+b} Among them, W represents the weight; ⊙ represents the cascade operation; f{n} represents the nonlinear transformation of the calculation result; b represents the bias term, which is the adjustment coefficient.

6. The method for processing hyperspectral data based on deep learning according to claim 1, characterized in that: In step 3, the transformed deep features and the hyperspectral signal are connected through residual connection to obtain a mixed signal.

Images