Tower base hyperspectral semantic change detection method considering illumination change

Abstract

The application relates to a semantic change detection method, in particular to a tower base hyperspectral semantic change detection method considering illumination change. The method comprises the following steps: S1, determining a target scene according to the position and light path coverage range of a tower base remote sensing platform, and pre-setting a hyperspectral imager; moving at least three gray scale targets with gradient difference to the target scene; S2, collecting image data of the target scene and all the gray scale targets by using the hyperspectral imager, obtaining the radiance values of all the gray scale targets according to the image data of all the gray scale targets, and performing radiation correction on the target scene by using the radiance values of all the gray scale targets and the known reflectivity values of the gray scale targets, so as to obtain the radiance values of the target scene; and S3, performing semantic change detection analysis by using the radiation-corrected hyperspectral data. The application can realize radiation correction of the target scene and obtain more accurate ground object reflectivity data.

Description

TECHNICAL FIELD

[0001] The application relates to a semantic change detection method, in particular to a tower base hyperspectral semantic change detection method considering illumination changes. BACKGROUND

[0002] In the field of remote sensing monitoring, hyperspectral imaging technology is widely used in ground object identification, environmental monitoring and resource exploration due to its ability to provide rich spectral information. However, due to changes in illumination conditions such as solar elevation angle, atmospheric conditions and cloud coverage, the collection of hyperspectral data will be significantly affected, resulting in differences in spectral reflectance of the same ground object under different illumination conditions. This difference may be misjudged as a change in the ground object itself, thereby affecting the accuracy of change detection.

[0003] Existing hyperspectral change detection methods usually rely on comparing hyperspectral image data obtained at different time points to identify changes in ground objects. However, these methods often do not fully consider the impact of illumination changes on spectral data, resulting in false positives or false negatives in change detection results under large changes in illumination conditions. In addition, since the tower base hyperspectral imaging system is usually fixed at a specific location, its observation conditions are relatively stable, but in the long-term monitoring process, changes in illumination conditions are still inevitable.

[0004] In order to improve the accuracy of change detection, it is necessary to develop a tower base hyperspectral change detection method that can consider the impact of illumination changes. SUMMARY

[0005] The purpose of the present application is to solve the technical problem that the existing detection method is prone to spectral differences caused by illumination changes and actual changes in ground objects, resulting in inaccurate detection results, and to provide a tower base hyperspectral semantic change detection method considering illumination changes.

[0006] To achieve the above purpose, the technical solution adopted by the present application is:

[0007] A tower base hyperspectral semantic change detection method considering illumination changes, characterized in that it comprises the following steps:

[0008] S1, determine the target scene according to the position and optical path coverage range of the tower base remote sensing platform, and pre-set a hyperspectral imager which can rotate 360°, adjust the hyperspectral imager to ensure that the imaging clarity meets the set requirements; at the same time, move at least three gray scale targets with gradient differences to the target scene so that the hyperspectral imager can simultaneously acquire image data of the target scene and the gray scale targets;

[0009] S2, using a hyperspectral imager to collect image data of a target scene and all gray scale targets, obtaining radiance values of all gray scale targets according to the image data of all gray scale targets, and using the radiance values of all gray scale targets and the known reflectivity values thereof to perform radiometric correction on the target scene to obtain radiance values of the target scene, i.e. hyperspectral data;

[0010] S3, using the hyperspectral data after radiometric correction to perform semantic change detection analysis

[0011] S3.1, first pre-processing the hyperspectral data after radiometric correction;

[0012] S3.2, inputting the pre-processed hyperspectral data into the trained twin neural network, learning and comparing the hyperspectral data at two different time points through the twin neural network to obtain a BCD graph and semantic graphs at the two different time points;

[0013] S3.3, obtaining an SCD graph according to the BCD graph and the semantic graphs at the two different time points, and completing the hyperspectral semantic change detection.

[0014] Further, in S2, using the radiance values of all gray scale targets and the known reflectivity values thereof, first calculate the transformation coefficients A and C through the following formula:

[0015] L a = Gain x DN + Bias

[0016] L a = A x p a + C

[0017] Wherein: L a is the radiance value of the gray scale target, Gain is the gain of the hyperspectral imager, DN is the digital quantization value output by each detection element of the hyperspectral imager, Bias is the bias value of the hyperspectral imager, p a is the surface reflectivity of the gray scale target;

[0018] Then, using the calculated transformation coefficients A and C, the radiance value L tar of the target scene is calculated through the following formula:

[0019] L tar = A x p tar + C

[0020] Wherein, p tar is the surface reflectivity of the target scene.

[0021] Further, the size of the hyperspectral data is adjusted to match the input requirements of the twin neural network, and the adjusted hyperspectral data is normalized to complete the preprocessing of the radiation corrected hyperspectral data.

[0022] Further, in S3.2, according to the hyperspectral data at two different time points, the deep BCD feature and the deep semantic feature at two different time points are obtained through the twin neural network.

[0023] According to the deep semantic features at two different time points, the semantic graphs at two different time points are obtained, and the deep semantic features at two different time points are fused with the deep BCD feature to obtain the BCD graph.

[0024] Further, in S3.3, according to the BCD graph and the semantic graphs at two different time points, the SCD graph is obtained through the mask operation.

[0025] Further, S4 is further included, which analyzes and verifies the SCD graph to evaluate the accuracy of the tower base hyperspectral semantic change detection.

[0026] The beneficial effects of the present application are:

[0027] The tower base hyperspectral semantic change detection method considering illumination change provided by the present application significantly improves the accuracy of hyperspectral semantic change detection through accurate radiation correction and advanced twin neural network. Especially for the inconsistency problem of ground object radiation under different time and weather conditions, the radiation correction of the target scene is realized, so that more accurate ground object reflectivity data is obtained, effectively reducing false positives and false negatives, and enhancing the reliability of remote sensing monitoring results. The rich dimension and continuous imaging time of the collected data enable the present application to realize long-term automatic unattended observation, meet the data demand of different time scales, and significantly improve the accuracy of change detection and the continuity of monitoring. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 is a radiation correction method schematic diagram of the tower base hyperspectral semantic change detection method considering illumination change of the present application;

[0029] Figure 2 is a flowchart of the tower base hyperspectral semantic change detection method considering illumination change of the present application;

[0030] Figure 3 is a radiation correction flowchart of the tower base hyperspectral semantic change detection method considering illumination change of the present application;

[0031] Figure 4 is a twin neural network flowchart in the tower base hyperspectral semantic change detection method considering illumination change of the present application. DETAILED DESCRIPTION

[0032] In order to make the objects, advantages and features of the present application more clearly, a kind of tower base hyperspectral semantic change detection method considering illumination change is further described in detail below in conjunction with the drawings and specific embodiments.The advantages and features of the present application will be clearer according to the following detailed description.

[0033] Referring to Figure 2 , the embodiment of the present application is a kind of tower base hyperspectral semantic change detection method considering illumination change, specifically including the following steps:

[0034] S1, referring to Figure 1 , the positioning ability of tower base remote sensing platform is used to determine the target scene according to the position and light path coverage range of tower base remote sensing platform, and the hyperspectral imager is adjusted to ensure that the imaging clarity meets the set requirements, to clearly capture the target scene within 10 kilometers range;At the same time, at least three gray scale targets with gradient difference are moved to the target scene, so that the hyperspectral imager can simultaneously acquire image data of the target scene and the gray scale target, to provide basic data for subsequent analysis and processing.

[0035] S2, given all gray scale target reflectivity, image data of target scene and all gray scale targets are collected using hyperspectral imager, and radiance values of all gray scale targets are obtained;Using the radiance values of all gray scale targets and their known reflectivity values, first calculate the conversion coefficients A and C by the following formula:

[0036] L a = Gain × DN + Bias

[0037] L a = A·ρ a +C

[0038] Wherein: L a It is the radiance value of gray scale target, Gain is the gain of hyperspectral imager, DN is the digital quantization value output by each detection element of hyperspectral imager, Bias is the bias value of hyperspectral imager, ρ a It is the surface reflectivity of gray scale target.

[0039] Then, using the calculated conversion coefficients A and C, the radiance value L tar of the target scene is radiometrically corrected by the following formula:

[0040] L tar = A·ρ tar +C

[0041] Wherein, ρ tar It is the surface reflectivity of target scene.

[0042] Radiance value of the aforementioned target scene, i.e. hyperspectral data.

[0043] S3、Referring to Figure 3 , semantic change detection analysis is performed on the hyperspectral data after radiation correction

[0044] S3.1, the hyperspectral data after radiation correction is preprocessed

[0045] The size of the hyperspectral data is adjusted to match the input requirements of the twin neural network, and the adjusted hyperspectral data is normalized to complete the preprocessing of the hyperspectral data after radiation correction.

[0046] S3.2, the preprocessed hyperspectral data is input into the trained twin neural network, the twin neural network can learn and compare the hyperspectral data at two different time points, i.e. T1 image and T2 image, and identify the differences between them.

[0047] The twin neural network uses an existing network, which is trained in advance to obtain a trained twin neural network. Specifically, see Figure 4 , the twin neural network includes a semantic encoder, a change encoder, a semantic decoder, and a change decoder.

[0048] According to the T1 image and the T2 image, the T1 deep semantic feature and the T2 deep semantic feature are obtained through the semantic encoder in the twin neural network, and the deep BCD feature is obtained through the change encoder in the twin neural network.

[0049] S3.3, according to the deep semantic features at two different time points, the T1 semantic map and the T2 semantic map are obtained through the semantic decoder in the twin neural network. At the same time, the T1 semantic map and the T2 semantic map are fused with the deep BCD feature through the twin neural network, and the fused feature is decoded through the change decoder in the twin neural network to obtain the BCD map (binary change map). Then, according to the BCD map and the deep semantic features at two different time points, the SCD map (semantic change map) is obtained through the mask operation, and the tower base hyperspectral semantic change detection is completed.

[0050] S4, the SCD map is analyzed and verified to evaluate the accuracy of the tower base hyperspectral semantic change detection.

Claims

1. A method for detecting hyperspectral semantic changes in a tower base considering illumination variations, characterized in that, Includes the following steps: S1. Determine the target scene based on the location and optical path coverage of the tower-based remote sensing platform, and adjust the hyperspectral imager to ensure that the imaging clarity meets the set requirements; at the same time, move at least three gray-level targets with gradient differences to the target scene so that the hyperspectral imager can acquire image data of the target scene and gray-level targets simultaneously. S2. Use a hyperspectral imager to acquire image data of the target scene and all gray-level targets. Based on the image data of all gray-level targets, obtain the radiance values ​​of all gray-level targets. Then, use the radiance values ​​of all gray-level targets and their known reflectance values ​​to perform radiometric correction on the target scene and obtain the radiance value of the target scene, i.e., hyperspectral data. S3. Using radiometrically corrected hyperspectral data for semantic change detection and analysis. S3.1 First, preprocess the radiometrically corrected hyperspectral data; S3.2 Input the preprocessed hyperspectral data into the trained Siamese neural network, and learn and compare the hyperspectral data at two different time points through the Siamese neural network to obtain the BCD map and the semantic map at the two different time points; the BCD map refers to the binary variation map. In S3.2, based on hyperspectral data from two different time points, deep BCD features and deep semantic features from the two different time points are obtained through a Siamese neural network. Based on the deep semantic features at two different time points, semantic maps at two different time points are obtained; at the same time, the deep semantic features at the two different time points are fused with deep BCD features to obtain a BCD map. S3.

3. Based on the BCD diagram and the semantic diagrams at two different time points, obtain the SCD diagram, which refers to the semantic change diagram, and complete the tower base hyperspectral semantic change detection.

2. The method for detecting hyperspectral semantic changes in a tower base considering illumination variations according to claim 1, characterized in that: In S2, using the radiance values ​​of all grayscale targets and their known reflectance values, the transformation coefficients A and C are first calculated using the following formula: L a =Gain×DN+Bias L a =A·ρ a +C Where: L a ρ represents the radiance value of the grayscale target, Gain represents the gain of the hyperspectral imager, DN represents the digital quantization value output by each detector element of the hyperspectral imager, Bias represents the bias value of the hyperspectral imager, and ρ represents the radiance value of the target. a The surface reflectance of a grayscale target; Using the calculated transformation coefficients A and C, the radiance value L of the target scene is calculated using the following formula. tar : L tar =A·ρ tar +C Where, ρ tar The surface reflectivity of the target scene.

3. The method for detecting hyperspectral semantic changes in a tower base considering illumination variations according to claim 1 or 2, characterized in that, S3.1 specifically refers to: The size of the hyperspectral data is adjusted to match the input requirements of the Siamese neural network, and the adjusted hyperspectral data is normalized to complete the preprocessing of the radiometrically corrected hyperspectral data.

4. The method for detecting hyperspectral semantic changes in a tower base considering illumination variations according to claim 3, characterized in that: In S3.3, the SCD graph is obtained through a masking operation based on the BCD graph and the semantic graphs at two different time points.

5. The method for detecting hyperspectral semantic changes in a tower base considering illumination variations according to claim 1, characterized in that: It also includes S4, which analyzes and validates the SCD image to evaluate the accuracy of detecting hyperspectral semantic changes in the tower base.

Images