High-precision short-wave infrared hyperspectral imaging soil plastic in-situ detection method and device

Abstract

The invention discloses a high-precision short-wave infrared hyperspectral imaging soil plastic in-situ detection method and a high-precision short-wave infrared hyperspectral imaging soil plastic in-situ detection device. The device is carried on an unmanned aerial vehicle or an indoor detection platform, large-range short-wave infrared hyperspectral imaging is carried out on soil, data are input into a multi-scale three-dimensional convolution instance segmentation network, space-spectrum dual-mode feature extraction is fused in the network, pixel-level classification of a plastic target under the complex soil background can be achieved, and the detection accuracy is improved. Even if the plastic is partially blocked or fragmented, the plastic can be effectively identified. And performing statistics on types, densities, spatial distribution and material traceability information of plastics in the soil area, and generating a plastic distribution space diagram for environmental plastic recovery path planning. Through the recognition technology, the efficient sorting capacity of the plastic packaging waste is improved, and the defect that in-situ classification recognition precision of mixed, colored or polluted plastic is poor in a traditional recycling method is overcome.

Description

Technical Field

[0001] The present invention belongs to the field of environmental resource recycling and protection, and specifically relates to a high-precision short-wave infrared hyperspectral imaging soil plastic in-situ detection method and device. Background Art

[0002] The EU-funded Polymark and HolyGrail 2.0 projects both aim to advance the efficiency and intelligence of plastics recycling systems and promote resource recycling by improving plastic packaging identification and sorting technologies. The Polymark project focuses on developing a chemical-based identification system that adds unique identifiers to plastic packaging that can be detected during the recycling process. This allows for precise sorting of different plastic types, improving recycling efficiency and the purity of recycled materials. The HolyGrail 2.0 project, on the other hand, uses digital watermarking technology to embed invisible information onto the packaging surface, enabling rapid identification of key attributes such as packaging material and purpose. This information can be read by intelligent camera systems on the sorting line, significantly improving the accuracy of automated sorting of plastic packaging, including colored plastics. Both projects share the same core objective: to improve the efficient sorting of plastic packaging waste through marking and identification technologies, addressing the technical bottlenecks of traditional recycling methods in handling mixed, colored, or contaminated plastics. However, the Polymark project and the HolyGrail 2.0 project require the use of chemical markers and digital watermarks to mark plastics. Marking plastics will increase the cost and expenses of plastic production. In addition, the marks may be difficult to identify and recycle due to stains and damage during use, which is not conducive to large-scale promotion and use.

[0003] To overcome the shortcomings of chemical markers and digital watermarks used in the Polymark and HolyGrail 2.0 projects, this paper proposes a label-free technical solution for classifying and identifying plastic types on complex soils based on the intrinsic absorption spectrum of plastic materials. This solution helps to promote the efficiency and intelligence of plastic recycling systems and is of great significance for promoting the recycling and reuse of resources. Summary of the Invention

[0004] To overcome the problems of the prior art, the present invention aims to provide a high-precision short-wave infrared hyperspectral imaging method and device for in-situ detection of soil plastics. The present invention can detect in-situ soil plastics using short-wave infrared hyperspectral imaging. The short-wave infrared hyperspectral imaging data can be segmented into pixel instances using a plastic 3D convolutional instance segmentation deep network model. Based on the instance segmentation results, the types and number of individual instances of various plastics in the soil are independently identified, and finally, the types, density, and number of individual instances of plastics in the soil of the detection area are counted. Short-wave infrared hyperspectral imaging can avoid the interference of visible light on color, obtain the intrinsic absorption spectrum of plastics, and thus can identify more than ten different types of plastics in different colors.

[0005] The technical solution for achieving the purpose of the present invention is as follows: A high-precision short-wave infrared hyperspectral imaging in-situ detection method for soil plastics comprises the following steps: S1. Use a shortwave infrared hyperspectral imager mounted on an unmanned aerial vehicle (UAV) or indoor testing platform to perform hyperspectral imaging of the soil area to be tested, obtaining a hyperspectral data cube with a wavelength range of 900-1800 nm and ≥400 spectral channels. S2. Performing radiation calibration, background noise suppression, spectral smoothing, and absorption peak normalization on the hyperspectral data cube to obtain preprocessed data reflecting the intrinsic absorption spectral characteristics of the plastic; S3. Input the preprocessed data obtained in step S2 into a pretrained Multi-Scale 3D Convolutional Instance Segmentation Network for Soil and Plastic (MS3D-ISNet). By integrating a spatial-spectral dual-modal feature extraction module, pixel-level instance segmentation results of soil and plastic are obtained. The instance segmentation results include the location of plastic pixels, type labels, and individual instance IDs. S4. Based on the instance segmentation results, the types, density, spatial distribution, and material traceability information of plastics in the tested soil area are collected for pollution source tracking and recycling path planning, forming a quantitative analysis report on soil plastic pollution.

[0006] The multi-scale three-dimensional convolutional instance segmentation network includes a three-dimensional convolutional feature extraction module, a spatial-spectral dual-stream feature fusion module and an instance segmentation decoding module; the three-dimensional convolutional feature extraction module performs spatial-spectral multimodal feature extraction on the input short-wave infrared hyperspectral data; the feature fusion module uses the residual channel attention mechanism to dynamically weighted fuse feature maps of different scales; the instance segmentation decoding module realizes pixel-level mask generation, bounding box prediction and material traceability classification of plastic targets.

[0007] The described method can simultaneously and quickly and accurately identify common plastics in soil, including PE, PP, PVC, PET, PS, PA, PMMA, PHC, and PVA.

[0008] Step S4. The specific volume of soil sample obtained from the soil in situ or detected by drone is (x, y, z). The plastic individuals in the first surface layer of the soil sample are segmented by type and instance using the soil plastic 3D convolutional instance segmentation deep network; the number of types of plastics in the first surface layer of the current soil sample is obtained. T i , where n is the number of types of plastics in the soil; the method for estimating the mass of plastics in the soil of the soil sample in the entire survey area is as follows:

[0009] P i is the specific volume of plastic type, i is the distribution density, V i Estimate the total survey area volume for soils.

[0010] The UAV is equipped with a short-wave infrared hyperspectral imager device, which performs a large-scale scan of the area to be tested according to a preset flight mode and path, and realizes real-time in-situ positioning and classification detection of soil plastics.

[0011] The method, step S1. The short-wave infrared hyperspectral imager includes an imaging lens, a slit, a collimating lens, a prism-near-infrared grating-prism combination, a Dove prism, a focusing lens and a short-wave infrared camera arranged in sequence, and a first near-infrared light source arranged around the first lens and a second near-infrared light source arranged around the second lens; the imaging lens images the target image of the plastic in the soil at the slit position, and the linear area light beam passing through the slit is collimated into a parallel light beam by the collimating lens; the parallel light beam is separated by wavelength after passing through the prism-near-infrared grating-prism combination to form light beams with different diffraction angles, and then focused on the short-wave infrared camera by the focusing lens to form a hyperspectral image.

[0012] The method described uses short-wave infrared hyperspectral imaging technology to obtain spectral information of soil samples in real time, realize in-situ detection and classification identification of plastic pollutants in the soil, and is used for sorting plastic packaging waste.

[0013] A device used according to the method comprises an imaging lens, a slit, a collimating lens, a prism-near-infrared grating-prism combination, a dove prism, a focusing lens and a short-wave infrared camera, which are arranged in sequence, and a first near-infrared light source and a second near-infrared light source arranged around the imaging lens; the imaging lens forms a target image of plastic in soil at the position of the adjustable slit, and the linear area light beam passing through the slit is collimated into a parallel light beam by the collimating lens; the parallel light beam is split according to wavelength after passing through the prism-near-infrared grating-prism combination to form light beams with different diffraction angles; the dove prism rotates the image by a certain angle, and then focuses the image on the short-wave infrared camera by the focusing lens to form a hyperspectral image.

[0014] Beneficial effects of the present invention: This invention discloses a method and apparatus for investigating colored plastics in soil using a shortwave infrared hyperspectral imager. This method can detect plastics in situ in soil using shortwave infrared hyperspectral imaging. The shortwave infrared hyperspectral imaging data can be segmented into pixel instances using a deep 3D convolutional instance segmentation model for plastics. Based on the instance segmentation results, the types and number of individual instances of various plastics in the soil are independently identified. Finally, the type, density, and number of individual instances of plastics in the soil within the detection area are counted. Shortwave infrared hyperspectral imaging avoids the interference of visible light on color, deriving the intrinsic absorption spectrum of plastics, and thus enabling the identification of different types of plastics in different colors.

[0015] This invention overcomes the defect of traditional low-precision spectral resolution that cannot accurately identify plastic types. It can quickly and accurately identify more than ten common plastic types, such as PE, PP, PVC, in the soil, as well as their density and spatial distribution information, providing more reliable data support for soil environmental surveys and promoting the development of plastic recycling and environmental protection. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flow chart of the high-precision shortwave infrared hyperspectral imaging soil plastic in situ detection method.

[0017] Figure 2 It is a device that uses high-precision short-wave infrared hyperspectral imaging to detect soil plastic in situ.

[0018] Figure 3 This is a schematic diagram of the MS3D-ISNet network structure.

[0019] Figure 4 It is the short-wave infrared hyperspectral curve used by the present invention to identify different types of plastics.

[0020] Figure 5 It is a rendering of different types of plastics recognized by the present invention.

[0021] Figure 2Icons in the figure: imaging lens 1, slit 2, collimating lens 3, prism-near-infrared grating-prism 4, Dove prism 5, focusing lens 6, short-wave infrared camera 7, first lens 8, first near-infrared light source 9, second lens 10, second near-infrared light source 11. DETAILED DESCRIPTION

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0023] like Figure 1 As shown, a high-precision short-wave infrared hyperspectral imaging in-situ detection method for soil plastics includes the following steps: S1. Use a shortwave infrared hyperspectral imaging device mounted on a drone or indoor testing platform to perform hyperspectral imaging of the soil area to be tested, generating a hyperspectral data cube with a wavelength range of 900-1800 nm and ≥400 spectral channels. The wavelength range of 900-1800 nm is characteristic for plastics and avoids interference from color on visible light absorption. ≥400 spectral channels provide more detailed and rich data, enabling high-precision plastic identification.

[0024] S2. Performing radiometric calibration, background noise suppression, spectral smoothing, and absorption peak normalization on the hyperspectral data cube to obtain preprocessed data reflecting the intrinsic absorption spectral characteristics of the plastic.

[0025] S3. The preprocessed data obtained in step S2 is fed into a pre-trained soil-plastic multi-scale 3D convolutional instance segmentation network (MS3D-ISNet). By integrating the spatial-spectral dual-mode feature extraction module, pixel-level instance segmentation results of soil plastic are obtained. The instance segmentation results include the location of plastic pixels, type labels, and instance individual identifiers (IDs). The detailed architecture and workflow of the proposed MS3D-Net model are shown in the figure. Figure 3As shown in Figure 1. First, the three-dimensional hyperspectral data cube is preprocessed through a convolutional layer to initially extract its spatial-spectral features. The data then enters the encoder module designed in this paper, which consists of two parts: a deep convolutional neural network and a dilated convolution. The ResNet-based DCNN contains multiple residual modules (ResBlocks). These modules utilize shortcut connections to allow features to propagate directly across layers, effectively overcoming the vanishing gradient problem and supporting efficient training of deeper neural networks. Each residual module in the DCNN extracts hierarchical and deeper spatial-spectral features step by step, effectively capturing the inherent correlations within the hyperspectral data. The feature maps processed by the DCNN are then fed into atrous convolutions (also known as dilated convolutions) with different dilation rates to further capture multi-scale contextual information. Dilated convolutions can significantly expand the network's receptive field without reducing spatial resolution, effectively preserving the detailed information necessary for accurate segmentation.

[0026] S4. Based on the instance segmentation results, the type, density, spatial distribution, and material traceability of plastics within the tested soil area are collected for pollution source tracking and recycling route planning, generating a quantitative analysis report on soil plastic pollution. This technology overcomes the limitations of traditional low-precision spectral resolution, which prevents accurate identification of plastic types. It can simultaneously and accurately identify more than ten common plastics in soil, including PE, PP, and PVC. This identification technology improves the efficient sorting of plastic packaging waste and addresses the technical bottlenecks of traditional recycling methods when dealing with mixed, colored, or contaminated plastics.

[0027] like Figure 2 As shown, the short-wave infrared hyperspectral imager device includes an imaging lens 1, a slit 2, a collimating lens 3, a prism-near-infrared grating-prism combination 4, a dove prism 5, a focusing lens 6 and a short-wave infrared camera 7, which are arranged in sequence, and a first near-infrared light source 9 and a second near-infrared light source 11 arranged around the imaging lens 1; the imaging lens 1 images the target image of the plastic in the soil at the position of the adjustable slit 2, and the linear area light beam passing through the slit 2 is collimated into a parallel light beam by the collimating lens 3; the parallel light beam is separated according to wavelength after passing through the prism-near-infrared grating-prism combination 4 to form light beams with different diffraction angles, the dove prism 5 rotates the image by a certain angle, and then focuses on the short-wave infrared camera 7 by the focusing lens 6 to form a hyperspectral image.

[0028] The multi-scale three-dimensional convolutional instance segmentation network includes a three-dimensional convolutional feature extraction module, a spatial-spectral dual-stream feature fusion module and an instance segmentation decoding module; the three-dimensional convolutional feature extraction module performs spatial-spectral multimodal feature extraction on the input short-wave infrared hyperspectral data; the feature fusion module uses the residual channel attention mechanism to dynamically weighted fuse feature maps of different scales; the instance segmentation decoding module realizes pixel-level mask generation, bounding box prediction and material traceability classification of plastic targets.

[0029] The high-precision short-wave infrared hyperspectral imaging in-situ detection method for soil plastics overcomes the defect that traditional low-precision spectral resolution cannot accurately identify the type of plastic. It can simultaneously and quickly and accurately identify more than ten common plastics in the soil, including but not limited to PE, PP, PVC, PET, PS, PA, PMMA, PHC, and PVA. Figure 4 It is the short-wave infrared hyperspectral curve used by the present invention to identify different types of plastics, wherein the spectral curves of various types of plastics are different. Figure 5 This is a rendering of different types of plastics identified by the present invention. Different plastics are mixed in the soil. Using the method proposed by the present invention, they can be well segmented and identified and marked with different colors.

[0030] The high-precision short-wave infrared hyperspectral imaging soil plastic in-situ detection method is described. The specific volume of soil sample obtained from the soil in-situ or detected by drone is (x, y, z). The soil plastic 3D convolutional instance segmentation deep network is used to perform type and instance segmentation on the plastic individuals in the first layer of the soil sample. In this way, the number of types of plastic individuals in the first layer of the current soil sample is obtained. T i , where n is the number of types of plastic in the soil. The mass of plastic in the soil sampled in the area is estimated as follows:

[0031] P i is the specific volume of plastic type, i is the distribution density, V i Estimate the total survey area volume for soils.

[0032] The drone is equipped with a short-wave infrared hyperspectral imager device, which conducts a large-scale scan of the area to be tested according to the preset flight mode and path, and realizes real-time in-situ positioning and classification detection of soil plastics, significantly improving the efficiency of plastic environmental monitoring and reducing monitoring costs.

[0033] The drone is equipped with a short-wave infrared hyperspectral imager device, which uses short-wave infrared hyperspectral imaging technology to obtain the spectral information of soil samples in real time, realizing in-situ detection and classification identification of plastic pollutants in the soil; through this identification technology, the efficient sorting capacity of plastic packaging waste is improved, and the technical bottlenecks of traditional recycling methods in dealing with mixed, colored or contaminated plastics are solved.

[0034] The embodiments described above may be further combined or replaced, and the embodiments are merely descriptions of preferred embodiments of the present invention and do not limit the concept and scope of the present invention. Various changes and improvements made to the technical solutions of the present invention by persons of ordinary skill in the art without departing from the design concept of the present invention are within the scope of protection of the present invention. The scope of protection of the present invention is defined by the appended claims and any equivalent technical solutions.

Claims

1. A high-precision short-wave infrared hyperspectral imaging soil plastic in-situ detection method, characterized in that: The steps include: S1. Use a shortwave infrared hyperspectral imager mounted on an unmanned aerial vehicle (UAV) or indoor testing platform to perform hyperspectral imaging of the soil area to be tested, obtaining a hyperspectral data cube with a wavelength range of 900-1800 nm and ≥400 spectral channels. S2. Performing radiation calibration, background noise suppression, spectral smoothing, and absorption peak normalization on the hyperspectral data cube to obtain preprocessed data reflecting the intrinsic absorption spectral characteristics of the plastic; S3. Input the preprocessed data obtained in step S2 into a pretrained Multi-Scale 3D Convolutional Instance Segmentation Network for Soil and Plastic (MS3D-ISNet). By integrating a spatial-spectral dual-modal feature extraction module, pixel-level instance segmentation results of soil and plastic are obtained. The instance segmentation results include the location of plastic pixels, type labels, and individual instance IDs. S4. Based on the instance segmentation results, the types, density, spatial distribution, and material traceability information of plastics in the tested soil area are collected for pollution source tracking and recycling path planning, forming a quantitative analysis report on soil plastic pollution.

2. The method according to claim 1, characterized in that The multi-scale three-dimensional convolutional instance segmentation network includes a three-dimensional convolutional feature extraction module, a spatial-spectral dual-stream feature fusion module and an instance segmentation decoding module; the three-dimensional convolutional feature extraction module performs spatial-spectral multimodal feature extraction on the input short-wave infrared hyperspectral data; the feature fusion module uses the residual channel attention mechanism to dynamically weighted fuse feature maps of different scales; the instance segmentation decoding module realizes pixel-level mask generation, bounding box prediction and material traceability classification of plastic targets.

3. The method according to claim 1, characterized in that It can quickly and accurately identify common plastics in soil, including PE, PP, PVC, PET, PS, PA, PMMA, PHC, and PVA.

4. The method according to claim 1, wherein: Step S4. A specific volume (x, y, z) of soil samples obtained in situ or detected by drone is used to segment the plastic individuals in the first surface layer of the soil sample using a soil plastic 3D convolutional instance segmentation deep network. To obtain the number of types of plastic in the first layer of the current soil sample T i , where n is the number of types of plastics in the soil; the method for estimating the mass of plastics in the soil of the soil sample in the entire survey area is as follows: ； P i is the specific volume of plastic type, i is the distribution density, V i Estimate the total survey area volume for soils.

5. The method according to claim 1, wherein The UAV is equipped with a short-wave infrared hyperspectral imager device, which performs a large-scale scan of the area to be tested according to a preset flight mode and path, and realizes real-time in-situ positioning and classification detection of soil plastics.

6. The method according to claim 1, wherein: Step S1. The short-wave infrared hyperspectral imager comprises an imaging lens (1), a slit (2), a collimating lens (3), a prism-near-infrared grating-prism combination (4), a Dove prism (5), a focusing lens (6) and a short-wave infrared camera (7) which are arranged in sequence, and a first near-infrared light source (9) arranged around the first lens (8) and a second near-infrared light source (11) arranged around the second lens (10); the imaging lens (1) images the target image of the plastic in the soil at the position of the slit (2), and the linear area light beam passing through the slit (2) is collimated into a parallel light beam by the collimating lens (3); the parallel light beam is separated according to wavelength after passing through the prism-near-infrared grating-prism combination (4) to form light beams with different diffraction angles, and then focused on the short-wave infrared camera (7) by the focusing lens (6) to form a hyperspectral image.

7. The method according to claim 1, wherein: Short-wave infrared hyperspectral imaging technology is used to obtain spectral information of soil samples in real time, enabling in-situ detection and classification identification of plastic pollutants in the soil for the sorting of plastic packaging waste.

8. A device used in the method according to claim 1, characterized in that: The invention comprises an imaging lens (1), a slit (2), a collimating lens (3), a prism-near-infrared grating-prism combination (4), a Dove prism (5), a focusing lens (6) and a short-wave infrared camera (7) which are arranged in sequence, and a first near-infrared light source (9) and a second near-infrared light source (11) which are arranged around the imaging lens (1); the imaging lens (1) forms an image of a target image of plastic in the soil at the position of the adjustable slit (2); the linear area light beam passing through the slit (2) is collimated into a parallel light beam by the collimating lens (3); the parallel light beam is split according to wavelength after passing through the prism-near-infrared grating-prism combination (4) to form light beams with different diffraction angles; the Dove prism (5) rotates the image by a certain angle, and then focuses the image on the short-wave infrared camera (7) through the focusing lens (6) to form a hyperspectral image.

Images