Wetland mapping method and system based on pre-training model and characteristic distillation model

Abstract

The invention relates to the technical field of remote sensing information processing, in particular to a wetland mapping method and system based on a pre-training model and a characteristic distillation model. The method comprises the following steps: acquiring a plurality of remote sensing images with different remote sensing parameters, and acquiring pixels from the remote sensing images; combining different remote sensing parameters of the pixels at the corresponding positions to obtain a parameter sequence; inputting the parameter sequence into a first model; converting the parameter sequence into a multi-dimensional image through the first model, carrying out semantic segmentation on the multi-dimensional image through a pre-training model and a lightweight model to obtain a pixel ultimate feature, and classifying the pixel ultimate feature through a classifier to obtain a wetland category of the pixel; the lightweight model is obtained by performing knowledge distillation on the pre-training model based on spectral features and spatial features; and importing the wetland category and the position of the corresponding pixel into a geographic information system to make a wetland type diagram. The technical effect of accurately classifying the wetland landscape by using fewer samples is realized.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing information processing technology, and in particular to a wetland mapping method and system based on a pre-training model and a feature distillation model. Background Art

[0002] Wetland monitoring and mapping are crucial prerequisites for wetland conservation, ecological restoration, wetland management, and ultimately sustainable development. Traditional wetland survey methods rely primarily on manual visits, which are time-consuming, labor-intensive, and inherently dangerous. With the advancement of geographic information science and remote sensing technology, remote sensing has enabled wetland mapping and monitoring of its dynamics. Supervised classification methods, such as random forests, can extract abstract features from multi-source imagery and geographic information data, and autonomously learn and optimize models, enabling large-scale automated mapping. However, there is still significant room for improvement in mapping accuracy and efficiency.

[0003] Currently, deep learning methods have shown good results in wetland remote sensing image classification and mapping. However, deep learning requires a large amount of labeled data, or the AI ​​model structure is complex, resulting in high computational costs, which brings huge cost and time pressures for long-term wetland mapping. Therefore, how to quickly classify wetland images using fewer samples is an urgent problem that needs to be solved. The present invention provides a wetland mapping method and system based on a pre-trained model and a feature distillation model, which realizes the classification of wetland landscapes based on fewer samples and improves the accuracy of wetland landscape classification. Summary of the Invention

[0004] The purpose of the present invention is to address the technical problem in the prior art that it is difficult to quickly classify wetland images using fewer samples. The embodiments of the present invention provide a wetland mapping method and system based on a pre-trained model and a feature distillation model. By designing a pre-trained self-supervised learning model to extract deep features, the obtained deep features are input into a lightweight model obtained by feature distillation based on the pre-trained model, and supervised wetland landscape classification is performed in a linear multi-layer perceptron of the lightweight model, thereby achieving the technical effect of classifying wetland landscapes using fewer samples.

[0005] In order to achieve the above-mentioned object of the invention, the present invention provides the following technical solutions: Wetland mapping method based on pre-trained model and feature distillation model, including: Acquiring remote sensing images of a wetland at different times, the remote sensing images including images of the wetland with different remote sensing parameters; performing regional classification on the remote sensing images based on terrain relief in the remote sensing parameters to obtain a core area representing the wetland; and acquiring pixels from the core area; Combining different remote sensing parameters of the pixels at corresponding positions at the same time to obtain a parameter sequence; inputting the parameter sequence into a first model; The first model converts the parameter sequence obtained after the combination into a multidimensional image through multiple functions; the multidimensional image is semantically segmented by a pre-trained model and a lightweight model to obtain the wetland category for the pixel representing the wetland landscape classification; the lightweight model is obtained by performing knowledge distillation on the pre-trained model based on spectral features and spatial features; The wetland categories and the corresponding pixel locations are imported into a geographic information system to produce a wetland type map.

[0006] As a preferred technical solution of the present invention, after acquiring remote sensing images of multiple different remote sensing parameters, the method further includes: performing regional classification on the remote sensing images based on the terrain relief in the remote sensing parameters, including: At least one first threshold and at least one second threshold are preset, wherein the first threshold is smaller than the second threshold; the pixel whose terrain relief is smaller than the first threshold is recorded as a core area representing a wetland; and the pixel whose terrain relief is greater than or equal to the second threshold is recorded as a mountainous area; The pixels whose terrain relief is greater than or equal to the first threshold and less than the second threshold are recorded as buffer zones representing interlaced wetlands and mountainous areas.

[0007] As a preferred technical solution of the present invention, after the regional classification, it also includes: performing binary classification on the buffer zone through the first model, dividing the buffer zone into wetland pixels and mountain pixels; and adding the wetland pixels in the buffer zone to the core area.

[0008] As a preferred technical solution of the present invention, at least one interval threshold is preset between the first threshold and the second threshold to obtain a plurality of buffer zones.

[0009] As a preferred technical solution of the present invention, the combination of different remote sensing parameters of the pixels at corresponding positions includes: Using a geographic information system, the remote sensing images of different spectral bands are synthesized into a true color image for representing spectral characteristics, and an RGB value of the remote sensing image is obtained; For all the pixels, the RGB value, slope index, aspect index, normalized vegetation index NDVI, normalized water index NDWI and visible atmospheric resistance index VARI of the pixels are combined into the parameter sequence; For the pixel in the buffer zone, calculating a first minimum distance from the pixel to an inner boundary, where the inner boundary is a plurality of the pixels corresponding to the first threshold; calculating a second minimum distance from the pixel to an outer boundary, where the outer boundary is a plurality of the pixels corresponding to the second threshold; calculating a ratio of the first distance to the second distance to obtain a distance ratio; and adding the terrain relief and the distance ratio to the parameter sequence; Extracting the slope index and aspect index of the remote sensing image based on DEM data using a geographic information system; Using a geographic information system to calculate based on the spectral bands and reflectance of the remote sensing image, the normalized vegetation index NDVI, the normalized water index NDWI, and the visible atmospheric resistance index VARI are obtained; The terrain relief is obtained by calculating the altitude of the remote sensing image using a geographic information system.

[0010] As a preferred technical solution of the present invention, converting the parameter sequence into a multidimensional image includes: Performing Grammi angle field conversion on the parameter sequence, converting the values ​​in the obtained two-dimensional Grammi angle field matrix into angles to obtain a first channel map; Performing Markov transition field conversion on the parameter sequence, and using the obtained transfer matrix as the second channel graph; Performing a recursive graph conversion on the parameter sequence, and using the obtained recursive matrix as the third channel graph; The first channel map, the second channel map, and the third channel map are combined into an HSV color space, one of the channel maps is converted into an H channel, one of the remaining two channel maps is converted into an S channel or a V channel, and the last channel map is converted into a V channel or an S channel; after converting the HSV space into an RGB color space, the multidimensional image is obtained.

[0011] As a preferred technical solution of the present invention, after acquiring a plurality of remote sensing images with different remote sensing parameters, the method further includes: performing standardization on the remote sensing images based on the remote sensing parameters, including: For the remote sensing images collected by different sensors, the remote sensing images with the same spectral band are screened out to obtain multispectral remote sensing images; If the multispectral remote sensing image has a panchromatic spectral band, fusing the panchromatic spectral band remote sensing image with the multispectral remote sensing image to improve the resolution; if the multispectral remote sensing image does not have a panchromatic spectral band, improving the resolution of the multispectral remote sensing image by k-nearest neighbor interpolation; The radiometric resolution of the multispectral remote sensing image is adjusted to a preset number of quantization bits.

[0012] As a preferred technical solution of the present invention, the semantic segmentation includes: The deep features of the multidimensional image are extracted through a pre-trained model, the deep features are enhanced through a lightweight model, and the ultimate features of the pixels are extracted; the parameter sequence is fused with the ultimate features of the pixels through a residual network to obtain new ultimate features of the pixels, and the new ultimate features of the pixels are used as the input of the classifier.

[0013] As a preferred technical solution of the present invention, when training the first model, before inputting the parameter sequence into the first model, the method further includes: classifying the plurality of parameter sequences based on spatial distance and / or feature similarity, the method including: Traversing the pixels corresponding to all the parameter sequences to obtain the current pixel; Acquire a plurality of pixels adjacent to the current pixel, recording them as adjacent pixels; traverse the adjacent pixels, respectively combine them with the current pixel to form a pair of paired pixels, and classify the obtained plurality of paired pixels as spatially adjacent positive samples; Based on the parameter sequence of the current pixel, feature similarity is calculated with other pixels, and a plurality of pixels whose feature similarity is greater than a preset similarity threshold are obtained, and recorded as similar pixels; the similar pixels are traversed, and are respectively combined with the current pixel to form a pair of paired pixels, and the obtained plurality of paired pixels are classified as feature adjacent positive samples; A plurality of pixels whose feature similarity is less than or equal to the similarity threshold are obtained and recorded as negative pixels; the negative pixels are traversed and respectively combined with the current pixel to form a pair of paired pixels, and the obtained plurality of paired pixels are classified as negative samples.

[0014] As a preferred technical solution of the present invention, the training of the first model includes: The paired pixels are respectively obtained from the spatially adjacent positive samples, the feature-adjacent positive samples, and the negative samples. The first model receives the paired pixels as input, converts the parameter sequences of the paired pixels into the multidimensional images, extracts features from the paired pixels through the pre-trained model and the lightweight model, and constructs a sample loss function based on the feature similarity of the paired pixels. The sample loss functions of different sample categories are summed to obtain the loss function L of the first model, which is as follows: ; in, For the paired pixels, i=1, 2, 3; when i=1, is the sample loss function of spatially adjacent positive samples; when i=2, is the sample loss function of the feature adjacent positive class sample; when i=3, is the sample loss function of negative samples; is the weight coefficient; is a hyperparameter.

[0015] As a preferred technical solution of the present invention, the method for constructing the sample loss function includes: Extracting depth features from the paired pixels using the pre-trained model, calculating the category similarity between the depth features, and obtaining a first decision feature; The depth feature is enhanced by the lightweight model, the ultimate features of the paired pixels are respectively extracted, and the category similarity between the ultimate features of the pixels is calculated to obtain a second decision feature; The sample loss function of the lightweight model is constructed based on the first decision feature and the second decision feature, so that the more similar the first decision feature and the second decision feature are, the smaller the value of the sample loss function is.

[0016] The present invention also provides a wetland mapping system based on a pre-trained model and a feature distillation model, comprising: A preprocessing module is configured to obtain a plurality of remote sensing images with different remote sensing parameters, obtain pixels from the remote sensing images, combine the different remote sensing parameters of the pixels at corresponding positions to obtain a parameter sequence, and input the parameter sequence into a first model; A semantic segmentation module, configured to convert the parameter sequence into a multidimensional image using the first model, semantically segment the multidimensional image using a pre-trained model and a lightweight model to obtain pixel ultimate features, and classify the pixel ultimate features using a classifier to obtain the wetland category of the pixel; the lightweight model is obtained by performing knowledge distillation on the pre-trained model based on spectral features and spatial features; The mapping module imports the wetland categories and the corresponding pixel locations into a geographic information system to produce a wetland type map.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. The wetland mapping method provided by the present invention is based on remote sensing image pixels. By upgrading the pixel parameter sequence from one dimension to a high dimension, the pre-trained self-supervised learning model is input to extract deep features from the pixels. The deep features are enhanced with the help of a lightweight model distilled from the pre-trained model, thereby reducing the input data required for semantic segmentation. The initial pixel parameter sequence is fused with the enhanced features through a residual network, thereby achieving the technical effect of accurately classifying wetland landscapes using fewer samples, which is conducive to better capturing various patterns and relationships in the image, improving the expressive power of the model, and providing more accurate modeling results.

[0018] 2. There is an interlaced transition zone between mountainous areas and wetlands, which is difficult to distinguish clearly. In order to capture the distribution of wetland landscapes in this transition zone, the remote sensing image is divided into three distinct blocks: mountainous areas, buffer zones, and core zones. The pixels in the buffer zones are reclassified, and the wetland parts in the buffer zones are merged into the core zones. The remote sensing images of the core zones are used to classify the wetland landscapes, which increases the expression of wetland characteristics and helps improve the integrity and accuracy of wetland mapping.

[0019] 3. While each pixel undergoing dimensionality increase has improved its ability to express features, this operation still simply transforms and combines the original features, making it difficult to directly reflect deep features. After dimensionality increase, the pixel classification task becomes a classification task of the RGB image representing the pixel. The pre-trained DINOv2 model already possesses general feature extraction capabilities. Using the DINOv2 model to extract deep features from multidimensional images reduces reliance on large amounts of annotated data, improves model stability and resistance to overfitting, and is suitable for small-sample learning scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A schematic structural diagram of the wetland mapping system provided by the present invention; Figure 2 A schematic flow chart of the wetland mapping method provided by the present invention; Figure 3 This is a schematic diagram of the results of regional classification of remote sensing images; Figure 4 Remote sensing image maps corresponding to multiple input features; Figure 5 is a structural diagram of the first model; Figure 6 Schematic diagram of the process of semantic segmentation for the first model; Figure 7 Schematic diagram of the process of performing feature dimensionality upgrade on a one-dimensional sequence of pixels; Figure 8 Schematic diagram of the lightweight model. DETAILED DESCRIPTION

[0021] To make the purpose, technical solutions and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them.

[0022] Therefore, the following detailed description of the embodiments of the present invention is not intended to limit the scope of the claimed invention, but merely represents some embodiments of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort shall fall within the scope of protection of the present invention.

[0023] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features and technical solutions therein may be combined with each other.

[0024] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings.

[0025] In the description of the present invention, it should be noted that the terms "upper" and "lower" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, or the orientations or positional relationships in which the inventive product is typically placed when in use, or the orientations or positional relationships commonly understood by those skilled in the art. Such terms are intended solely to facilitate the description of the present invention and simplify the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limitations on the present invention. Furthermore, the terms "first" and "second" and the like are used solely for distinction and should not be construed as indicating or implying relative importance.

[0026] For example Figure 1 As shown, the wetland mapping system based on the pre-trained model and feature distillation model provided in this embodiment includes: A preprocessing module is configured to obtain a plurality of remote sensing images with different remote sensing parameters, obtain pixels from the remote sensing images, combine the different remote sensing parameters of the pixels at corresponding positions to obtain a parameter sequence, and input the parameter sequence into a first model; A semantic segmentation module, configured to convert the parameter sequence into a multidimensional image using the first model, semantically segment the multidimensional image using a pre-trained model and a lightweight model to obtain pixel ultimate features, and classify the pixel ultimate features using a classifier to obtain the wetland category of the pixel; the lightweight model is obtained by performing knowledge distillation on the pre-trained model based on spectral features and spatial features; The mapping module imports the wetland categories and the corresponding pixel locations into a geographic information system to produce a wetland type map.

[0027] In this embodiment, for example Figure 2 As shown, the wetland mapping system is used to execute the wetland mapping method based on the pre-training model and the feature distillation model provided in this embodiment, including: Acquiring remote sensing images of a wetland at different times, the remote sensing images including images of the wetland with different remote sensing parameters; performing regional classification on the remote sensing images based on terrain relief in the remote sensing parameters to obtain a core area representing the wetland; and acquiring pixels from the core area; Combining different remote sensing parameters of the pixels at corresponding positions at the same time to obtain a parameter sequence; inputting the parameter sequence into a first model; The first model converts the parameter sequence obtained after the combination into a multidimensional image through multiple functions; the multidimensional image is semantically segmented using a pre-trained model and a lightweight model to obtain the wetland category for the pixel used to represent the wetland landscape classification. In this embodiment, the wetland categories include: river-lake, swamp, swampy meadow, meadow, meadow-shrub-tree, bare sand, building-road, and mountain. The lightweight model is obtained by performing knowledge distillation on the pre-trained model based on spectral features and spatial features; The wetland categories and the corresponding pixel locations are imported into a geographic information system such as ArcGIS software to produce a wetland type map.

[0028] In this embodiment, the study area of ​​the remote sensing image is set as the Ruoergai Wetland Nature Reserve, whose administrative area includes Sichuan Province and Gansu Province; among them, Sichuan Province includes Aba County, Ruoergai County, Hongyuan County and Songpan County, and Gansu Province includes Maqu County and Luqu County.

[0029] In this embodiment, the research time series of remote sensing images is set to 30 years. Due to the long time series, the types of remote sensing images available for selection are relatively small; at the same time, clouds have a greater impact on the quality of remote sensing images, and it is necessary to screen images with fewer clouds in the core study area as input data. Optionally, it also includes interpolating the pixels corresponding to the cloud layer in the remote sensing image through the k-nearest neighbor algorithm based on the pixels corresponding to the surface cover adjacent to the cloud layer, so as to achieve the purpose of removing the cloud layer. Taking into account the image shooting time, imaging quality and stability, spectral consistency and data acquisition cost, Landsat surface reflectance data is used, as shown in Table 1. Landsat surface reflectance data has different types of data collected due to the different sensors of different satellites. Therefore, it is necessary to standardize the Landsat surface reflectance data.

[0030] Table 1 Data sources and corresponding years ; In one embodiment, remote sensing images are standardized based on remote sensing parameters of Landsat surface reflectance data, including but not limited to spectral band type, spatial resolution, and radiometric resolution. The number of spectral band types determines the data volume of different remote sensing image types; spatial resolution measures the ground extent corresponding to a single pixel in a remote sensing image and is a key factor in determining the usability of classification results. Improving the spatial resolution of an image can reduce interference from mixed pixels; and radiometric resolution reflects the sensor's ability to distinguish differences in radiation intensity.

[0031] In a specific embodiment, for Landsat data with different sensor parameters, remote sensing images with the same spectral bands are selected. The spectral bands of remote sensing images are different, and the corresponding data sizes are different. Furthermore, it also includes selecting remote sensing images with the same or similar wavelength parameters based on the sensor parameters of Landsat surface reflectance data. Among them, the sensor parameters of Landsat5 and Landsat7 are exactly the same, the sensor parameters of Landsat8 and Landsat9 are exactly the same, and the wavelength values ​​between Landsat5 / 7 and Landsat8 / 9 are very close. Therefore, for different types of Landsat surface reflectance data of all years, remote sensing images of the six spectral bands shared by Landsat5 / 7 / 8 / 9 are retained to obtain multispectral remote sensing images, including blue light spectral band B1, green light spectral band B2, red light spectral band B3, near infrared spectral band B4, short wave infrared 1 spectral band B5, and short wave infrared 2 spectral band B6.

[0032] In another specific embodiment, the resolution of multispectral remote sensing images with the same spectral bands is improved. By adding a panchromatic spectral band to Landsat 7 / 8 / 9 remote sensing images and fusing the panchromatic spectral band with the multispectral remote sensing images, the spatial resolution can be increased from 30m to 15m. Since the Landsat 5 sensor design does not include a panchromatic spectral band, in order to ensure that all input remote sensing images have the same spatial resolution, a super-resolution operation with a scaling factor of 2 is performed on the Landsat 5 remote sensing images using the k-nearest neighbor interpolation method, increasing the spatial resolution of Landsat 5 from 30m to a multispectral remote sensing image with a resolution of 15m.

[0033] In another specific embodiment, the radiometric resolution of the multispectral remote sensing images is unified to a preset number of quantization bits. In this embodiment, the quantization bits of all multispectral remote sensing images are converted to 16-bit.

[0034] In this embodiment, after acquiring multiple remote sensing images with different remote sensing parameters, the method further includes: performing regional classification on the remote sensing images based on the terrain relief in the remote sensing parameters. In this embodiment, for example Figure 3 As shown in (a), multiple classification results are marked and distinguished using multiple different colors, where gray represents mountainous areas, yellow represents buffer areas, and green represents core areas. Regional classification methods include: The terrain relief is calculated based on the altitude of the remote sensing image using a geographic information system such as ArcGIS software; at least one first threshold and at least one second threshold are preset, the first threshold is less than the second threshold, and the pixels with the terrain relief less than the first threshold are recorded as the core area representing the wetland; the pixels with the terrain relief greater than or equal to the first threshold and less than the second threshold are recorded as the buffer zone representing the interlaced wetland and mountainous area. The buffer zone diagram is as follows: Figure 3 (b) and Figure 3 As shown in (d); the pixels whose terrain relief is greater than or equal to the second threshold are recorded as mountainous areas.

[0035] Optionally, at least one interval threshold is preset between the first threshold and the second threshold to obtain a plurality of buffer zones.

[0036] In another embodiment, for example Figure 3 (c) and Figure 3 As shown in (e), after the regional classification, it also includes: inputting the remote sensing image corresponding to the buffer zone into the first model, performing binary classification based on the terrain undulation, recording the pixels whose terrain undulation is less than a preset third threshold as wetland pixels, and adding the wetland pixels in the buffer zone to the core area; recording the pixels whose terrain undulation is greater than or equal to the third threshold as mountainous pixels, and adding the mountainous pixels in the buffer zone to the mountainous area. Figure 3 middle, Figure 3 (c) Figure 3 (b) Schematic diagram after regional classification; Figure 3 (e) Figure 3 (d) Schematic diagram after regional classification.

[0037] There is an interlaced transition zone between mountainous areas and wetlands, which is difficult to distinguish clearly. In order to capture the distribution of wetland landscapes in this transition zone, the remote sensing image is divided into three distinct blocks: mountainous areas, buffer zones, and core zones. The pixels in the buffer zones are reclassified, and the wetland parts in the buffer zones are merged into the core zones. The remote sensing images of the core zones are used to classify the wetland landscapes, which increases the expression of wetland characteristics and helps improve the integrity and accuracy of wetland mapping.

[0038] Furthermore, the accuracy of pixel-based remote sensing image segmentation depends on the characteristics of each pixel. The input feature data type of the first model in this embodiment is shown in Table 2, and different remote sensing parameter combinations are used to obtain parameter sequences for the pixels classified in different regions.

[0039] Table 2 Input features and their meanings ; Specifically, for example Figure 4 As shown in (a), the remote sensing images of 6 spectral bands are synthesized into true color images using ArcGIS software to represent the spectral characteristics of the remote sensing images.

[0040] Secondly, since vegetation growth and water content are the core characteristics of wetlands, the normalized vegetation index NDVI is calculated based on spectral bands and reflectance using ArcGIS software to obtain vegetation images, such as Figure 4 (b) shows that the normalized vegetation index is [-0.18, 0.63]; the normalized water index NDWI is calculated by ArcGIS software to obtain the water index image, for example Figure 4 (c) shows that the normalized water index is [-0.88, 0.17]; the visible atmospheric resistance index VARI is calculated by ArcGIS software to obtain the atmospheric image, for example Figure 4 (d) shows that the atmospheric drag index is [-0.57, 0.19]. Vegetation images, water index images, and atmospheric images are used to represent the surface cover characteristics of wetlands.

[0041] In addition, since topography and landforms also have an impact on the distribution of wetland landscapes, this embodiment also obtains ALOS 12.5-meter resolution DEM remote sensing images, and calculates the slope index based on the DEM remote sensing images using ArcGIS software to obtain a slope index image, for example Figure 4 (e) shows that the slope index is [-1 degree, 359.87 degrees]. The slope index is calculated by ArcGIS software to obtain the slope index image, for example Figure 4 (f) shows that the aspect index is [0 degrees, 84.74 degrees]. Slope index image and aspect index image are used to represent the spatial characteristics of terrain elements.

[0042] Since wetlands and mountains are interwoven in the buffer zone, the mountainous areas and the wetland core areas have distinct geomorphological features and are spatially adjacent to each other. Therefore, for the remote sensing images in the buffer zone, the terrain relief RDLS and the distance ratio to the buffer zone boundary PBBI are added, corresponding to the relief image and the distance ratio image, for example Figure 4 (g) and Figure 4 (h) is used to represent the buffer space characteristics.

[0043] The terrain relief RDLS is calculated by calculating the difference between the highest point and the lowest point in a specific area. The calculation formula is as follows: ; Among them, H max Represents the highest point altitude, H min Represents the lowest point altitude. In this embodiment, for example Figure 4 As shown in (g), the terrain relief value is [0 m, 1307 m], where m is the unit of terrain relief: meter.

[0044] The distance to the buffer zone boundary ratio PBBI is calculated as follows: ; Among them, d inner represents the first minimum distance from the buffer pixel to the inner boundary, where the inner boundary is the plurality of pixels corresponding to the first threshold; d outer Indicates the second minimum distance from the buffer pixel to the outer boundary, and the outer boundary is a plurality of pixels corresponding to the second threshold. In this embodiment, for example Figure 4 As shown in (h), the distance ratio is [0, 1265].

[0045] It should be noted that, in this embodiment, Figure 4 (b) – Figure 4 In (h), different colors are used to represent different values ​​of the corresponding remote sensing parameters.

[0046] In an optional embodiment, the input feature data corresponds to all pixels in the remote sensing image, and the RGB values, slope index, aspect index, normalized difference vegetation index (NDVI), normalized water index (NDWI), and visible atmospheric drag index (VARI) of the pixels are combined into the parameter sequence, recorded as the first parameter sequence. For the pixels in the buffer zone, the terrain relief and the distance ratio are added to the parameter sequence, recorded as the second parameter sequence.

[0047] Preferably, in order to reduce the amount of data of the first model, since the present invention only classifies wetland landscapes, it is preferred that the input feature data corresponds to the pixels in the core area, the first parameter sequence is used for the pixels in the core area, and the second parameter sequence is used for the pixels in the buffer area.

[0048] In this embodiment, the structural diagram of the first model is as follows: Figure 5 As shown, it includes: a high-dimensional mapping network for feature dimensionality upgrading; the pre-trained model that has been pre-trained and has frozen parameters; the lightweight model obtained by performing knowledge distillation on the pre-trained model based on spectral features and spatial features, a residual network for fusing the parameter sequence and the ultimate features of the pixel, and the high-dimensional mapping network, the pre-trained model and the lightweight model are connected in series.

[0049] The first model performs semantic segmentation, e.g. Figure 6 As shown, the steps include: The deep features of the multidimensional image are extracted through a pre-trained model, and the deep features are enhanced through a lightweight model to extract the ultimate features of the pixel; the parameter sequence is fused with the ultimate features of the pixel through a residual network to obtain a new ultimate feature of the pixel, and the new ultimate feature of the pixel is used as the input of the classifier to obtain the wetland category of the pixel.

[0050] In this embodiment, for example Figure 7 (a) is a schematic diagram of a pixel parameter sequence. The method for converting the parameter sequence into a multidimensional image includes: using Matlab software to increase the dimension of the pixel parameter sequence through the Gramian Angular Field (GAF) method, outputting a Gramian Angular Field matrix with a dimension of H x W, and visualizing the Gramian Angular Field matrix through Matlab with H, W and GAF ​​values ​​as coordinate axes. The result is, for example Figure 7 (c) As shown. Since the GAF value range is [-1, 1], the GAF value multiplied by 2π can be used as the H channel in the HSV color space. Therefore, in this embodiment, the values ​​in the obtained two-dimensional Gramian angle field matrix ‌ are converted into angles to obtain the first channel map. Matlab software is used to increase the dimension of the parameter sequence of the paired pixels through the Markov Transition Field (MTF) method, and a transfer matrix with a dimension of H x W is output. The transfer matrix is ​​visualized by Matlab with H, W and MTF values ​​as coordinate axes. The result is as follows: Figure 7 As shown in (d). Since the MTF value range is [0, 1], it can be used as the S channel in the HSV color space. Therefore, in this embodiment, the obtained transfer matrix is ​​used as the second channel diagram. Matlab software is used to increase the dimension of the parameter sequence of the paired pixels through the recurrence plot (RP) method based on the similarity measurement function, and a recurrence matrix with a dimension of H x W is output. The recurrence matrix is ​​visualized by Matlab with H, W and RP values ​​as coordinate axes. The result is as follows: Figure 7 As shown in (e). Since the RP value range is [0, 1], it can be used as the V channel in the HSV color space. Therefore, in this embodiment, the obtained recursive matrix is ​​used as the third channel map. The HSV color space obtained by combining the first channel map, the second channel map, and the third channel map is converted into the RGB color space to obtain a multi-dimensional image input to the pre-trained model, for example Figure 7(b) shows the height and width of the input multidimensional image, where H and W are the height and width of the input multidimensional image. The pixel parameter sequence is a one-dimensional feature expression, which has a low degree of expression of pixel characteristics, making it difficult for the model to capture this characteristic expression of the pixel. Therefore, the one-dimensional sequence is converted to a high-dimensional sequence, so that the feature expression can include or highlight more features.

[0051] In some other embodiments, the values ​​of the first channel map are transformed into a numerical range [0, 1], for example, by taking absolute values, and can be used as the S channel or the V channel. The second channel map can be directly used as the V channel; or the values ​​of the second channel map can be transformed into a numerical range [-1, 1], for example, by normalization, and can be used as the H channel. The values ​​of the third channel map can be transformed into a numerical range [-1, 1], for example, by normalization, and can be used as the H channel; or the third channel map can be directly used as the S channel.

[0052] The pixel parameter sequence is a one-dimensional feature expression, which has a low degree of representation of pixel characteristics, making it difficult for the model to capture this characteristic expression of the pixel. By upgrading the parameter sequence from one dimension to a higher dimension, the higher-dimensional feature expression can include or highlight more features, thereby helping to better capture various patterns and relationships in the image, improving the model's expressiveness and providing more accurate modeling results.

[0053] In this embodiment, when training the first model, before inputting the parameter sequence into the first model, the method further includes: classifying the plurality of parameter sequences based on spatial distance and / or feature similarity, the method including: Traversing the pixels corresponding to all the parameter sequences to obtain the current pixel; Wetland landscape classification has a certain area coverage, and its spatial distribution is usually patchy, that is, adjacent pixels may have the same category. Multiple pixels adjacent to the current pixel are obtained and recorded as adjacent pixels. The adjacent pixels are traversed and combined with the current pixel to form a pair of paired pixels. The obtained multiple paired pixels are classified as spatially adjacent positive samples. Based on the parameter sequence of the current pixel, feature similarity is calculated with other pixels, and a plurality of pixels whose feature similarity is greater than a preset similarity threshold are obtained, and recorded as similar pixels; the similar pixels are traversed, and are respectively combined with the current pixel to form a pair of paired pixels, and the obtained plurality of paired pixels are classified as feature adjacent positive samples; A plurality of pixels whose feature similarity is less than or equal to the similarity threshold are obtained and recorded as negative pixels; the negative pixels are traversed and respectively combined with the current pixel to form a pair of paired pixels, and the obtained plurality of paired pixels are classified as negative samples.

[0054] In a specific embodiment, the method for calculating feature similarity includes a K-nearest neighbor algorithm, which calculates the feature space distance between the current pixel and other pixels, such as the Euclidean distance, and takes the inverse of the feature space distance as the feature similarity. The smaller the feature space distance, the higher the feature similarity.

[0055] In another embodiment, the method for calculating feature similarity includes taking the parameter sequence as a vector to calculate cosine similarity, and taking the cosine similarity as the feature similarity. The greater the cosine similarity, the higher the feature similarity.

[0056] In this embodiment, the training of the first model includes: The paired pixels are obtained from the spatially adjacent positive samples, the characteristically adjacent positive samples, and the negative samples, respectively, and the first model receives the paired pixels as input. Preferably, to increase the generalization capability of the model, only one pair of the paired pixels is randomly selected from the characteristically adjacent positive samples as input. The parameter sequences of the paired pixels are respectively converted into the multidimensional images.

[0057] Specifically, the pre-trained model extracts depth features from the multidimensional image, including: obtaining the multidimensional image corresponding to the paired pixels; using the pre-trained model to extract depth features from the multidimensional image; calculating category similarity for the depth features corresponding to the paired pixels to obtain a first decision feature for indicating whether the paired pixels belong to the same category.

[0058] Specifically, in a preferred embodiment, the formula for extracting deep features using the pre-trained model is as follows: ; in, is a multidimensional image of paired pixels after feature dimension upgrading, O is a sample set consisting of spatially adjacent positive samples, feature-adjacent positive samples, and negative samples, C = 3 is the number of RGB channels, H and W are the height and width of the RGB image after dimension upgrading; M pt is a pre-trained model obtained by pre-training an existing model; l and m are the two output depth features corresponding to the paired pixels; vec() represents a function that flattens a matrix into a vector. For the paired pixels, the similarity between the depth features l and m indicates whether they belong to the same category. The formula for calculating category similarity is as follows: ; The larger the first decision feature is, the higher the probability that the paired pixels belong to the same category. ptThis is DINOv2 using ViT-g / 14 parameters. This was obtained by pre-training using a comparison method, selecting the model based on maximum classification accuracy from existing models: SimCLRv2, MoCov2, YOLOv5, YOLOv8, and DINOv2. Note that the parameters of the pre-trained model are frozen during training.

[0059] While each pixel undergoing dimensionality increase has enhanced its ability to express features, this operation still simply transforms and combines the original features, making it difficult to directly reflect deep features. After dimensionality increase, the pixel classification task becomes a classification task of the RGB image representing the pixel. The pre-trained DINOv2 model already possesses general feature extraction capabilities. Using the DINOv2 model to extract deep features from multidimensional images reduces reliance on large amounts of annotated data, improves model stability, and resists overfitting, making it suitable for small-sample learning scenarios.

[0060] Specifically, the lightweight model enhances the depth feature, including: inputting the depth feature corresponding to the paired pixels into the lightweight model, and extracting the corresponding pixel ultimate feature. The structure of the lightweight model is as follows: Figure 8 As shown in Figure 1, a multi-layer perceptron is composed of multiple linear layers, and ReLU activation connections are used between multiple linear layers, where B is the number of training batches. represents the dimension of the depth feature, N is the linear layer dimension, and Φ is the dimension of the ultimate pixel feature. The formula for extracting the ultimate pixel feature of the lightweight model is as follows: ; Among them, l, m are the deep features extracted by the pre-training model; p, q are the ultimate features of the pixel after being enhanced by the lightweight model; M light The lightweight model is described. The category similarity of the ultimate feature of the pixel is calculated to obtain the second decision feature, which is used to determine whether the paired pixels belong to the same category. The calculation formula for the second decision feature S is as follows: ; During the training process, a sample loss function of the lightweight model is constructed based on the first and second decision features. The more similar the first and second decision features are, the smaller the value of the sample loss function is, indicating that the process of extracting deep features by the pre-trained model and the process of extracting pixel ultimate features by the lightweight model are more consistent. The formula of the sample loss function is as follows: ; Where F is the first decision feature; S is the second decision feature; and β is a preset hyperparameter used to determine whether the paired pixels belong to the positive or negative class. For positive input samples, the setting of the hyperparameter β will make F-β a positive number. Minimizing the loss function will increase S, indicating that the ultimate feature similarity between positive samples has increased. For negative input samples, the setting of the hyperparameter β will make F-β a negative number. Minimizing the loss function will decrease S, indicating that the ultimate feature similarity between negative samples has decreased, ultimately achieving a clustering effect.

[0061] After extracting features from the paired pixels using the pre-trained model and the lightweight model, a sample loss function is constructed based on the feature similarity of the paired pixels; based on the sample loss functions of the spatially adjacent positive samples, the feature-adjacent positive samples, and the negative samples, the three sample loss functions are weighted and summed to construct a model loss function L, as shown in the following formula: ; in, For the paired pixels, i=1, 2, 3; when i=1, is the sample loss function of spatially adjacent positive samples; when i=2, is the sample loss function of the feature adjacent positive class sample; when i=3, is the sample loss function of negative samples; is the weight coefficient; is a hyperparameter. During the training process, minimizing the model loss function of the first model means maximizing the similarity between positive samples and minimizing the similarity between negative samples. This balance relationship is given by and changes with the adjustment.

[0062] In this embodiment, the optimizer of the first model uses the Adam optimizer with a learning rate of 0.0001. The training batch size is set to 256, the validation batch size is set to 512, and the classifier uses the Softmax linear classifier. During model verification, the linear probe adopts a simple perceptron structure and uses the cross entropy loss function to train the linear projection of the ultimate feature to the category label. The optimization method also uses Adam with a learning rate of 0.001. All experiments are implemented using the PyTorch framework. The experiment uses an NVIDIA GeForce RTX 3080 GPU with a processor cache of 10GB. The training time takes nearly 12 hours. The specific hyperparameter settings of the first model are shown in Table 3: Table 3 Hyperparameter configuration of the first model ; Wherein, K1 is the number of spatially adjacent positive samples in this embodiment, K2 is the number of feature-adjacent positive samples, K3 is the number of negative samples; λ is the number of linear layers of the lightweight model.

[0063] During training, the present invention selected seven deep learning models, namely KNN, SimCLRv2, MoCov2, Swin-Transformer, DINOv2, IIC, and ResNet50, and conducted comparative experiments with the first model. The results of the F1 score, overall accuracy OA, and Kappa coefficient value of the wetland category are shown in Table 4: Table 4 Experimental results of multi-model wetland landscape classification ; After comparative experiments, the results showed that the classification results of the first model were better than those of other deep learning model algorithms, proving that the first model can be applied to wetland landscape mapping.

[0064] The present invention focuses on the subdivision of the same landform in the core area of ​​the wetland. For example, rivers, lakes and swamps can be roughly divided into water bodies. The degree of attention paid to the feature sequence carried by each pixel is greater than the edge features between pixels that reflect the relationship between each other. At the same time, since the remote sensing images studied in the long time series in the present invention are medium-resolution remote sensing images, the problem of inaccurate edges is prone to occur. The wetland mapping method provided by the present invention is based on remote sensing image pixels. By upgrading the parameter sequence of the pixel from one dimension to a high dimension, the pre-trained self-supervised learning model is input to extract the depth feature of the pixel, and the depth feature is enhanced with the help of a lightweight model distilled from the pre-trained model, thereby reducing the input data required for semantic segmentation. The initial pixel parameter sequence is fused with the enhanced feature through the residual network, thereby achieving the technical effect of accurately classifying the wetland landscape using fewer samples, which is conducive to better capturing various patterns and relationships in the image, improving the expression ability of the model, and providing more accurate modeling results.

[0065] Those skilled in the art will appreciate that all or some of the steps and systems in the method disclosed above can be implemented as software, firmware, hardware, and appropriate combinations thereof. Some physical components or all physical components can be implemented as software executed by a processor, such as a central processing unit, a digital signal processor, or a microprocessor, or implemented as hardware, or implemented as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, and the computer-readable medium can include computer storage media (or non-transitory media) and communication media (or temporary media). As known to those skilled in the art, the term computer storage media is included in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data) and is volatile and non-volatile, removable, and non-removable. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory, or other memory technology, CD-ROM, digital versatile disks (DVD), or other optical disk storage, magnetic cassettes, magnetic tapes, disk storage, or other magnetic storage devices, or any other medium that can be used to store desired information and can be accessed by a computer. Furthermore, as is well known to those skilled in the art, communication media typically embodies computer-readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media.

[0066] The above embodiments are only used to illustrate the present invention and are not intended to limit the technical solutions described in the present invention. Although this specification has described the present invention in detail with reference to the above embodiments, the present invention is not limited to the above specific implementation methods. Therefore, any modification or equivalent replacement of the present invention; and all technical solutions and improvements thereof that do not depart from the spirit and scope of the invention are included in the scope of the claims of the present invention.

Claims

1. A wetland mapping method based on a pre-trained model and a feature distillation model, characterized in that: include: Acquire multiple remote sensing images with different remote sensing parameters, and obtain pixels from the remote sensing images; combine the different remote sensing parameters of the pixels at corresponding positions to obtain a parameter sequence; and input the parameter sequence into a first model; The parameter sequence is converted into a multidimensional image using the first model, the multidimensional image is semantically segmented using a pre-trained model and a lightweight model to obtain pixel ultimate features, and the pixel ultimate features are classified using a classifier to obtain the wetland category of the pixel; the lightweight model is obtained by performing knowledge distillation on the pre-trained model based on spectral features and spatial features; The wetland categories and the corresponding pixel locations are imported into a geographic information system to produce a wetland type map.

2. The method according to claim 1, wherein After acquiring the remote sensing images of multiple different remote sensing parameters, the method further includes: performing regional classification on the remote sensing images based on the terrain relief in the remote sensing parameters, including: At least one first threshold and at least one second threshold are preset, wherein the first threshold is smaller than the second threshold; and the pixel whose terrain relief is smaller than the first threshold is recorded as a core area representing a wetland; The pixels whose terrain relief is greater than or equal to the first threshold and less than the second threshold are recorded as buffer zones representing the interlaced wetlands and mountainous areas; and the pixels in the core area are used as inputs of the first model.

3. The method according to claim 2, wherein The combining of different remote sensing parameters of the pixels at corresponding positions includes: synthesizing the remote sensing images of different spectral bands into a true color image for representing spectral characteristics, and obtaining RGB values ​​of the remote sensing image; For all the pixels, combining the RGB value, the slope index, the aspect index, the normalized vegetation index, the normalized water index and the visible atmospheric resistance index into the parameter sequence; For the pixel in the buffer zone, calculating a first distance from the pixel to an inner boundary, where the inner boundary is the pixel corresponding to the first threshold; calculating a second distance from the pixel to an outer boundary, where the outer boundary is the pixel corresponding to the second threshold; calculating a ratio of the first distance to the second distance to obtain a distance ratio; and adding the terrain relief and the distance ratio to the parameter sequence.

4. The method according to claim 1, wherein Converting the parameter sequence into a multi-dimensional image comprises: Performing Grammi angle field conversion on the parameter sequence, converting the values ​​in the obtained two-dimensional Grammi angle field matrix into angles to obtain a first channel map; Performing Markov transition field conversion on the parameter sequence, and using the obtained transfer matrix as the second channel graph; Performing a recursive graph conversion on the parameter sequence, and using the obtained recursive matrix as the third channel graph; The first channel image, the second channel image, and the third channel image are combined into an HSV color space, and converted into an RGB color space to obtain the multidimensional image.

5. The method according to claim 1, wherein After acquiring a plurality of remote sensing images with different remote sensing parameters, the method further includes: performing standardization on the remote sensing images based on the remote sensing parameters, including: For the remote sensing images collected by different sensors, the remote sensing images with the same spectral band are screened out to obtain multispectral remote sensing images; If the multispectral remote sensing image has a panchromatic spectral band, fusing the panchromatic spectral band remote sensing image with the multispectral remote sensing image to improve the resolution; if the multispectral remote sensing image does not have a panchromatic spectral band, improving the resolution of the multispectral remote sensing image by k-nearest neighbor interpolation; The radiometric resolution of the multispectral remote sensing image is adjusted to a preset number of quantization bits.

6. The method according to claim 1, wherein The semantic segmentation includes: The deep features of the multidimensional image are extracted through a pre-trained model, the deep features are enhanced through a lightweight model, and the ultimate features of the pixels are extracted; the parameter sequence is fused with the ultimate features of the pixels through a residual network to obtain new ultimate features of the pixels, and the new ultimate features of the pixels are used as the input of the classifier.

7. The method according to any one of claims 1 to 6, wherein: When training the first model, before inputting the parameter sequence into the first model, the method further includes: classifying the plurality of parameter sequences based on spatial distance and / or feature similarity, the method including: Traversing the pixels corresponding to all the parameter sequences to obtain the current pixel; Acquire a plurality of pixels adjacent to the current pixel, combine them with the current pixel respectively, and classify the obtained plurality of paired pixels into spatially adjacent positive samples; Based on the parameter sequence of the current pixel, feature similarity is calculated with other pixels, a plurality of pixels having feature similarity greater than a preset similarity threshold are obtained, the pixels are respectively combined with the current pixel, and the plurality of paired pixels obtained are classified as feature adjacent positive samples; A plurality of pixels whose feature similarity is less than or equal to the similarity threshold are obtained, and the pixels are respectively combined with the current pixel, and the obtained plurality of paired pixels are classified as negative samples.

8. The method according to claim 7, wherein The training of the first model comprises: The paired pixels are respectively obtained from the spatially adjacent positive samples, the feature-adjacent positive samples and the negative samples. The first model receives the paired pixels as input, converts the parameter sequences of the paired pixels into the multidimensional images, extracts features from the paired pixels through the pre-trained model and the lightweight model, and constructs a sample loss function based on the feature similarity of the paired pixels. The sample loss functions of different sample categories are summed to obtain the loss function of the first model.

9. The method according to claim 7, wherein The method for constructing the sample loss function includes: Extracting depth features from the paired pixels using the pre-trained model, calculating the category similarity between the depth features, and obtaining a first decision feature; The depth feature is enhanced by the lightweight model, the ultimate features of the paired pixels are respectively extracted, and the category similarity between the ultimate features of the pixels is calculated to obtain a second decision feature; The sample loss function of the lightweight model is constructed based on the first decision feature and the second decision feature, so that the more similar the first decision feature and the second decision feature are, the smaller the value of the sample loss function is.

10. A wetland mapping system based on a pre-trained model and a feature distillation model, characterized by: include: A preprocessing module is configured to obtain a plurality of remote sensing images with different remote sensing parameters, obtain pixels from the remote sensing images, combine the different remote sensing parameters of the pixels at corresponding positions to obtain a parameter sequence, and input the parameter sequence into a first model; A semantic segmentation module, configured to convert the parameter sequence into a multidimensional image using the first model, semantically segment the multidimensional image using a pre-trained model and a lightweight model to obtain pixel ultimate features, and classify the pixel ultimate features using a classifier to obtain the wetland category of the pixel; the lightweight model is obtained by performing knowledge distillation on the pre-trained model based on spectral features and spatial features; The mapping module imports the wetland categories and the corresponding pixel locations into a geographic information system to produce a wetland type map.

Images