A method and system for extracting elements of an alpine pasture

Abstract

The application relates to the technical field of remote sensing information processing and geographic information system, and discloses a mountain pasture element extraction method and system, and aims to solve the problems of insufficient local detail perception and multi-scale information fusion capability of existing remote sensing image classification. The application comprises the following steps: remote sensing image preprocessing; extracting and converting multi-dimensional spectral-spatial joint features into multi-channel two-dimensional images; constructing and training an improved Swin-UNet model integrated with a local perception enhancement module and a cross-layer feature fusion module; using the model for mountain pasture element classification; and dynamically evaluating the grass-livestock balance. By using the above technical scheme, the application can significantly improve the classification accuracy, spatial continuity and robustness, and effectively support the dynamic evaluation and scientific management of the grass-livestock balance.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing information processing and geographic information system technology, and specifically relates to a method and system for extracting elements of mountain pastures. Background Technology

[0002] To achieve scientific assessment and dynamic monitoring of grassland-livestock balance, accurately acquiring pasture ecological elements, especially the spatiotemporal distribution information of grassland area, is a fundamental and crucial prerequisite. Within the existing technological system, classification methods based on high-resolution remote sensing imagery have been widely applied to ground feature extraction. Traditional remote sensing image classification largely relies on pixel-level classification of ground features based on their spectral characteristics, supplemented by the extensive prior knowledge and practical experience of the classifiers. These methods played a positive role in early ground feature classification, enabling the identification of spectral differences among different ground features, thereby achieving preliminary regional delineation.

[0003] However, with the continuous improvement of the spatial resolution of remote sensing images and the increasing demand for refined and automated information extraction in application scenarios, traditional methods, due to their heavy reliance on human experience and limitations in handling complex spectral confusion and ground fragmentation, have gradually become unable to meet the current requirements for high-precision and large-scale information extraction based on remote sensing data.

[0004] Therefore, how to fundamentally optimize the existing Swin-UNet model structure to more effectively enhance the extraction and representation of local detailed features, while improving the ability to fuse multi-scale contextual information, thereby significantly improving classification accuracy, reducing misclassification, and enhancing spatial continuity and robustness in the task of extracting complex mountain pasture ecological elements from high-resolution remote sensing images, has become a key challenge and an urgent technical problem for those skilled in the art. Summary of the Invention

[0005] To achieve the above objectives, this application provides the following technical solution:

[0006] According to a first aspect of the present invention, the present invention claims protection for a method for extracting elements of mountain pastures, comprising the following steps:

[0007] S1: Acquire remote sensing image data and preprocess it to generate a multi-temporal, high-precision remote sensing image dataset with uniform radiometric and geometric characteristics.

[0008] S2, Based on the preprocessed remote sensing image dataset, extract spectral features, terrain features, texture features and regional aggregation features to construct a comprehensive spectral-spatial joint feature dataset;

[0009] S3, perform a one-dimensional sequence to two-dimensional image conversion on the spectral-spatial joint feature dataset to generate two-dimensional image input data with multi-channel features;

[0010] S4. Construct and train an improved Swin-UNet model. The improved Swin-UNet model includes an encoder, a bottleneck module, a decoder, and several skip connections. The encoder integrates a local perception enhancement module, and the decoder integrates a cross-layer feature fusion module.

[0011] S5. The improved Swin-UNet model, after training, is used to classify the mountain pasture elements in the two-dimensional image input data to obtain a high-precision pasture element classification map.

[0012] S6. Based on the pasture element classification map, conduct a dynamic assessment of the grass-livestock balance to determine the grass-livestock balance status level of the mountain pasture.

[0013] Further, step S1 includes the following sub-steps:

[0014] S11, acquire high-resolution remote sensing images, the image data including images of the dry grass period and images of the lush grass period, wherein the high-resolution remote sensing images are images from the Gaofen-2 satellite and images from the Gaofen-7 satellite.

[0015] S12, perform standardized preprocessing on the high-resolution remote sensing image, the standardized preprocessing includes radiometric correction, atmospheric correction, geometric correction, image fusion, image mosaicking, histogram matching and spatial alignment.

[0016] S13, the preprocessed remote sensing images are divided into a grassy season image dataset and a grassy dry season image dataset according to the acquisition time. The grassy season is set to May to September each year, and the grassy dry season is set to October to April of the following year.

[0017] Furthermore, S2 extracts spectral features, topographic features, texture features, and regional aggregation features based on the preprocessed remote sensing image dataset to construct a comprehensive spectral-spatial joint feature dataset;

[0018] The spectral features include raw four-band data, soil-modified vegetation index (SAVI), and first principal component (PCA1). The raw four-band data includes blue light, green light, red light, and near-infrared light.

[0019] The terrain feature is slope;

[0020] The texture feature is the mean value of the gray-level co-occurrence matrix (GLCM).

[0021] The region aggregation feature is the mean of the first principal component of the multi-scale segmented region.

[0022] Furthermore, the soil-regulated vegetation index (SAVI) is calculated as follows: SAVI = (NIR - Red) / (NIR + Red + L) * (1 + L), where NIR is the near-infrared reflectance, Red is the red light reflectance, and L is the soil regulation coefficient, which ranges from 0 to 1.

[0023] The first principal component, PCA1, was obtained by performing principal component analysis on the original four-band data.

[0024] The slope is obtained by calculation from the digital elevation model (DEM) data;

[0025] The mean value of the gray-level co-occurrence matrix (GLCM) is obtained by calculating the gray-level image of the remote sensing image. The mean value of the GLCM is the sum of the products of all element values ​​in the GLCM matrix and their corresponding frequencies.

[0026] The mean value of the first principal component of the multi-scale segmented region is obtained by calculating the mean value of the first principal component PCA1 within each segmented object after performing multi-scale segmentation processing on the remote sensing image.

[0027] Further, step S3 converts the spectral-spatial joint feature dataset from a one-dimensional sequence to a two-dimensional image, generating two-dimensional image input data with multi-channel features, including the following sub-steps:

[0028] S31, for each pixel in the spectral-spatial joint feature dataset, its spectral feature sequence and spatial feature sequence are concatenated to form a one-dimensional feature vector;

[0029] S32, the one-dimensional feature vector is taken as input and converted into three two-dimensional feature matrices through Gram angle field transformation, Markov transition field transformation and recursive graph transformation respectively.

[0030] S33, the three two-dimensional feature matrices are spliced ​​together in the channel dimension to generate a two-dimensional image input data with multi-channel features;

[0031] The Gram angle field transformation maps the elements in the one-dimensional feature vector to a polar coordinate system and calculates the cosine values ​​of the sum or difference between different time points, generating a symmetric matrix representing time correlation.

[0032] The Markov transition field transformation is based on a first-order Markov chain, which calculates the transition probabilities between states in the one-dimensional feature vector to generate a two-dimensional matrix representing the state transition mode.

[0033] The recursive graph transformation generates a symmetric matrix representing the dynamic behavior of a time series by calculating the Euclidean distance between points in the one-dimensional feature vector and determining whether state recursion has occurred based on a preset threshold.

[0034] Furthermore, S4 constructs and trains an improved Swin-UNet model, as follows:

[0035] The encoder consists of four stages, each stage including a block merging layer and one or more sliding window transformer modules (STBs), and each STB is connected to a local perception enhancement module (LPEM).

[0036] The bottleneck module is located between the encoder and the decoder, and consists of a block merging layer and one or more sliding window transformer modules.

[0037] The decoder consists of four stages, each stage including a block expansion layer and one or more sliding window transformer modules, and each decoder stage includes a cross-layer feature fusion module (CLFFM).

[0038] The plurality of skip connections are respectively used to transmit the feature map of each stage of the encoder to the corresponding stage of the CLFFM of the decoder;

[0039] The output layer of the improved Swin-UNet model consists of a 1x1 convolutional layer and a Softmax activation function layer.

[0040] Furthermore, the structure of the Local Perception Enhancement Module (LPEM) includes:

[0041] A feature input interface is provided for receiving feature maps from the sliding window transformer module;

[0042] A multi-branch dilated convolutional unit contains multiple parallel convolutional branches, each of which employs a 3x3 depthwise separable convolution and sets different dilation rates to capture local contextual information at different scales.

[0043] The adaptive spatial feature fusion unit utilizes a spatial attention mechanism to dynamically emphasize locally important spatial regions in the feature map;

[0044] The channel splicing layer is used to splice the output of the multi-branch dilated convolutional unit with the output of the adaptive spatial feature fusion unit in the channel dimension;

[0045] A 1x1 convolutional layer is used to compress the channel dimension and fuse the concatenated features, and serve as the feature output of the LPEM;

[0046] The structure of the cross-layer feature fusion module CLFFM includes:

[0047] The encoder feature input interface is used to receive feature maps from the corresponding stage of the encoder;

[0048] The decoder upsampled feature input interface is used to receive the feature map upsampled from the previous stage of the decoder;

[0049] The channel alignment module adjusts the number of channels of the feature map from the encoder's feature input interface using a 1x1 convolutional layer to match the number of channels of the feature map from the decoder's upsampled feature input interface.

[0050] The hollow space pyramid pooling ASPP module receives the channel-aligned encoder feature map and the decoder upsampled feature map as input, and aggregates multi-scale context information through multiple parallel branches;

[0051] The channel splicing layer is used to splice the outputs of the various branches of the ASPP module along the channel dimension.

[0052] A 1x1 convolutional layer is used to perform channel dimension compression and feature fusion on the concatenated features;

[0053] The weighted summation module is used to perform a weighted summation of the features fused by the ASPP module and the upsampled feature map of the decoder.

[0054] The channel-spatial attention module receives the weighted summed features as input, generates a refined feature map, and uses it as the feature output of the CLFFM.

[0055] Further, step S5 uses the trained and improved Swin-UNet model to classify mountain pasture features from the two-dimensional image input data to obtain a high-precision pasture feature classification map, including the following sub-steps:

[0056] S51, For the pasture image dataset during the lush pasture season, the pasture feature classification map output by the improved Swin-UNet model contains six types of land features: grassland, shrubs, river-lake, cloud, bare land, and shadow.

[0057] S52, For the dry season image dataset, the pasture feature classification map output by the improved Swin-UNet model contains five types of land features: grassland, shrubs, river-lake-cloud-snow, bare land, and shadow.

[0058] After the pasture element classification map is generated, it is post-processed using morphological operations or Conditional Random Field (CRF) to eliminate the salt and pepper effect.

[0059] Further, step S6, based on the pasture element classification map, performs a dynamic assessment of the grass-livestock balance to determine the grass-livestock balance status level of the mountain pasture, including the following sub-steps:

[0060] S61, Calculate the area of ​​grassland elements based on the pasture element classification map;

[0061] S62, Based on the grassland area and field survey data, calculate the theoretical carrying capacity R and the actual carrying capacity A of the mountain pasture;

[0062] S63, Calculate the grass-livestock balance index BGLI of the mountain pasture based on the theoretical carrying capacity R and the actual carrying capacity A;

[0063] S64, Determine the grass-livestock balance status level of the mountain pasture based on the value of the grass-livestock balance index (BGLI);

[0064] The theoretical carrying capacity R is calculated as follows: R=(Y*E) / (I*D), where R is the reasonable carrying capacity of the grassland during the grazing period, Y is the total grass yield of the grassland, E is the grassland utilization rate, I is the daily feed intake of sheep, and D is the grazing duration.

[0065] The formula for calculating the grass-livestock balance index (BGLI) is: BGLI = (AR) / R;

[0066] The grass-livestock balance status is divided into five levels:

[0067] Extreme overloading, when the BGLI is greater than 150%; severe overloading, when the BGLI is greater than 80% and less than or equal to 150%; overloading, when the BGLI is greater than 20% and less than or equal to 80%; grass-livestock balance, when the BGLI is greater than -20% and less than or equal to 20%; underloading, when the BGLI is less than or equal to -20%.

[0068] According to a second aspect of the present invention, the present invention claims protection for a mountain pasture element extraction system, comprising:

[0069] The remote sensing image preprocessing module is used to acquire and preprocess remote sensing image data to generate a multi-temporal, high-precision remote sensing image dataset with uniform radiometric and geometric characteristics.

[0070] The spectral-spatial feature extraction module is used to extract spectral features, topographic features, texture features, and regional aggregation features from the remote sensing image dataset output by the remote sensing image preprocessing module, so as to construct a comprehensive spectral-spatial joint feature dataset.

[0071] The feature 2D conversion module is used to convert the spectral-spatial joint feature dataset output by the spectral-spatial feature extraction module from a one-dimensional sequence to a two-dimensional image, generating two-dimensional image input data with multi-channel features;

[0072] An improved Swin-UNet model is used to receive two-dimensional image input data output by the feature 2D transformation module and output a mountain pasture feature classification map. The improved Swin-UNet model includes an encoder, a bottleneck module, a decoder and several skip connections. The encoder integrates a local perception enhancement module and the decoder integrates a cross-layer feature fusion module.

[0073] The grass-livestock balance dynamic assessment module is used to perform a grass-livestock balance dynamic assessment based on the pasture element classification map output by the improved Swin-UNet model, so as to determine the grass-livestock balance status level of the mountain pasture.

[0074] The aforementioned mountain pasture element extraction system is used to execute the aforementioned mountain pasture element extraction method.

[0075] This application relates to the field of remote sensing information processing and geographic information systems, and discloses a method and system for extracting elements from mountain pastures, aiming to solve the problems of insufficient local detail perception and multi-scale information fusion capabilities in existing remote sensing image classification. This invention preprocesses remote sensing images; extracts and converts them into multi-channel two-dimensional images with multi-dimensional spectral-spatial joint features; constructs and trains an improved Swin-UNet model integrating a local perception enhancement module and a cross-layer feature fusion module; uses this model to classify elements in mountain pastures; and dynamically evaluates the grassland-livestock balance. By adopting the above technical solutions, this application can significantly improve classification accuracy, spatial continuity, and robustness, effectively supporting the dynamic evaluation and scientific management of grassland-livestock balance. Attached Figure Description

[0076] Figure 1 This is a flowchart illustrating the process of a method for extracting elements from a mountain pasture, as claimed in an embodiment of the present invention.

[0077] Figure 2 This is a second workflow diagram of a method for extracting elements from a mountain pasture as claimed in an embodiment of the present invention.

[0078] Figure 3 A third workflow diagram of a method for extracting elements from mountain pastures, as claimed in this embodiment of the invention;

[0079] Figure 4 This is a structural block diagram of a mountain pasture element extraction system claimed in an embodiment of the present invention;

[0080] Figure 5A schematic diagram of the overall classification results of a study area for a method for extracting elements of a mountain pasture as claimed in an embodiment of the present invention;

[0081] Figure 6 This is a partial schematic diagram of the classification details of a mountain pasture element extraction method claimed in an embodiment of the present invention. Detailed Implementation

[0082] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0083] The terms "first," "second," and "third" in this application are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of this application are only used to explain the relative positional relationships and movements between components in a specific orientation (as shown in the figures). If the specific orientation changes, the directional indications also change accordingly. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0084] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0085] With the rapid development of deep learning technology, the Transformer architecture, especially its multi-head self-attention mechanism, has demonstrated powerful parallel computing capabilities and advantages in long-distance dependency modeling, achieving breakthroughs in natural language processing and image recognition. SwinTransformer, as an improvement on Transformer in the vision domain, effectively reduces computational complexity by employing a moving window-based self-attention mechanism, making it more suitable for processing large-scale remote sensing imagery. This architecture introduces a moving window into the self-attention mechanism, aiming to achieve global modeling while using a hierarchical construction method to better apply it to multi-dimensional image data. Building on this, the Swin-UNet model further integrates the SwinTransformer module into the classic U-shaped network structure, featuring encoders, decoders, bottlenecks, and skip connections, becoming an ideal backbone network for processing remote sensing images with limited spatial information. Its self-attention mechanism's global feature extraction capability demonstrates significant potential in extracting important features from remote sensing images.

[0086] However, despite the numerous advantages demonstrated by the Swin-UNet model in handling remote sensing image classification tasks, it still faces some deep-seated and non-obvious inherent limitations when applied to the fine extraction of extremely complex and heterogeneous ecological elements, such as those found in mountain pastures on the Qinghai-Tibet Plateau. Specifically, while the original Swin-UNet model achieves multi-scale feature extraction through its hierarchical shifting window Transformer module, its ability to perceive and enhance local details and effectively fuse multi-scale contextual information is not optimal in pixel-level classification tasks of high-resolution remote sensing images, especially when accurately distinguishing features with subtle spectral differences and complex spatial boundaries, such as grasslands, shrubs, and bare land.

[0087] The reasons for this are twofold. Firstly, the Swin-UNet model primarily uses a self-attention mechanism for feature aggregation. While this captures long-range dependencies, it falls short compared to the feature extraction and texture perception capabilities of convolutional operations within the local receptive field. It may be less effective at capturing fine local texture features and accurately locating edge details. For the subtle spectral differences and irregular, blurred transition areas frequently encountered in mountain pastures between grasslands and shrubs, and between bare land and degraded grasslands, relying solely on the Transformer's self-attention mechanism is insufficient to completely overcome misclassification caused by high heterogeneity and severe spectral confusion. This can easily lead to the "salt and pepper effect," where numerous isolated pixels or messy small patches appear in the classification map, affecting the spatial continuity of land features. Under a pixel-centric classification approach, this phenomenon directly challenges the accuracy of final area estimation and grassland-livestock balance assessment. Secondly, while Swin-UNet's skip connection mechanism helps recover spatial information lost during encoder block merging, it primarily involves simple feature splicing or addition, lacking a more efficient and sophisticated mechanism to aggregate rich semantic contextual information from different scales. For pasture feature extraction that requires simultaneous consideration of local details and global semantics, especially under the combined influence of complex terrain and texture features, the original skip connections alone are insufficient to effectively fuse these heterogeneous multi-scale information. This limits the model's discriminative ability and robustness when handling ground features of diverse shapes and scales. In particular, in high-resolution imagery, subtle terrain undulations, local variations in vegetation density, and factors such as illumination and shadow can all have complex effects on the spectral response of ground features. Without a dedicated module to enhance local details and achieve broader multi-scale context fusion, the original Swin-UNet falls short in adapting to these complex scenes.

[0088] According to the first embodiment of the present invention, referring to Figure 1 This invention claims protection for a method for extracting elements from mountain pastures, comprising the following steps:

[0089] S1: Acquire remote sensing image data and preprocess it to generate a multi-temporal, high-precision remote sensing image dataset with uniform radiometric and geometric characteristics.

[0090] S2, Based on the preprocessed remote sensing image dataset, extract spectral features, terrain features, texture features and regional aggregation features to construct a comprehensive spectral-spatial joint feature dataset;

[0091] S3, perform a one-dimensional sequence to two-dimensional image conversion on the spectral-spatial joint feature dataset to generate two-dimensional image input data with multi-channel features;

[0092] S4. Construct and train an improved Swin-UNet model. The improved Swin-UNet model includes an encoder, a bottleneck module, a decoder, and several skip connections. The encoder integrates a local perception enhancement module, and the decoder integrates a cross-layer feature fusion module.

[0093] S5. The improved Swin-UNet model, after training, is used to classify the mountain pasture elements in the two-dimensional image input data to obtain a high-precision pasture element classification map.

[0094] S6. Based on the pasture element classification map, conduct a dynamic assessment of the grass-livestock balance to determine the grass-livestock balance status level of the mountain pasture.

[0095] Furthermore, referring to Figure 2 S1 includes the following sub-steps:

[0096] S11, acquire high-resolution remote sensing images, the image data including images of the dry grass period and images of the lush grass period, wherein the high-resolution remote sensing images are images from the Gaofen-2 satellite and images from the Gaofen-7 satellite.

[0097] S12, perform standardized preprocessing on the high-resolution remote sensing image, the standardized preprocessing includes radiometric correction, atmospheric correction, geometric correction, image fusion, image mosaicking, histogram matching and spatial alignment.

[0098] S13, the preprocessed remote sensing images are divided into a grassy season image dataset and a grassy dry season image dataset according to the acquisition time. The grassy season is set to May to September each year, and the grassy dry season is set to October to April of the following year.

[0099] In this embodiment, high-resolution remote sensing image data is acquired, strategically covering the two key growing seasons of the target area: the dry season and the wet season. Preferably, the acquired high-resolution remote sensing images are from the Gaofen-2 GF-2 satellite and the Gaofen-7 GF-7 satellite, with spatial resolutions of 0.8 meters and 0.65 meters respectively, providing sufficient detail for refined classification.

[0100] The Gaofen-2 satellite imagery includes four spectral bands: blue, green, red, and near-infrared. The Gaofen-7 satellite imagery includes panchromatic and multispectral blue, green, red, and near-infrared images. The time series of these image data covers at least three complete annual cycles to capture seasonal variations and interannual fluctuations in vegetation growth.

[0101] A standardized preprocessing workflow is used for high-resolution remote sensing imagery. This standardized preprocessing workflow aims to eliminate various errors and distortions in the raw image data, ensuring data quality and comparability.

[0102] The specific operations include: radiometric correction, which converts the digital values ​​(DN) of the image into the physical quantities of radiance or reflectance to eliminate sensor differences and atmospheric effects; atmospheric correction, which uses models such as the FLAASH or 6S to remove atmospheric scattering and absorption effects and obtain the true surface reflectance; geometric correction, which uses ground control point (GCP) and DEM data to accurately register the image to a unified geographic coordinate system, with the registration error controlled within 0.5 pixels; and image fusion, which fuses the panchromatic bands and multispectral bands of Gaofen-2 or Gaofen-7 to generate an image with high spatial resolution and... For multispectral images with rich spectral information, Gram-Schmidt fusion or Pan-sharpening methods can be used for fusion algorithms. Image stitching involves seamlessly stitching adjacent remote sensing images to cover a larger study area, using feathering or color balance techniques to eliminate stitching gaps. Histogram matching adjusts the grayscale distribution of images from different batches or different sensors to make them have similar statistical characteristics, thereby eliminating visual differences and spectral inconsistencies. Spatial alignment ensures that all images are precisely aligned with each other at the pixel level for subsequent temporal analysis and feature extraction.

[0103] The standardized preprocessed remote sensing image data was divided into two datasets based on its acquisition time: the lush grass season and the dry grass season. The lush grass season, taking into account geographical region and climate characteristics, is specifically defined as May to September each year, during which vegetation growth is vigorous and biomass reaches its peak. The dry grass season, on the other hand, is defined as October to April of the following year, during which vegetation is dormant or withered, and the surface features are mainly bare ground, withered grass, and snow cover. This seasonal division allows the model to accurately learn and classify ground cover features for different phenological periods.

[0104] Furthermore, S2 extracts spectral features, topographic features, texture features, and regional aggregation features based on the preprocessed remote sensing image dataset to construct a comprehensive spectral-spatial joint feature dataset;

[0105] The spectral features include raw four-band data, soil-modified vegetation index (SAVI), and first principal component (PCA1). The raw four-band data includes blue light, green light, red light, and near-infrared light.

[0106] The terrain feature is slope;

[0107] The texture feature is the mean value of the gray-level co-occurrence matrix (GLCM).

[0108] The region aggregation feature is the mean of the first principal component of the multi-scale segmented region.

[0109] Furthermore, the soil-regulated vegetation index (SAVI) is calculated as follows: SAVI = (NIR - Red) / (NIR + Red + L) * (1 + L), where NIR is the near-infrared reflectance, Red is the red light reflectance, and L is the soil regulation coefficient, which ranges from 0 to 1.

[0110] The first principal component, PCA1, was obtained by performing principal component analysis on the original four-band data.

[0111] The slope is obtained by calculation from the digital elevation model (DEM) data;

[0112] The mean value of the gray-level co-occurrence matrix (GLCM) is obtained by calculating the gray-level image of the remote sensing image. The mean value of the GLCM is the sum of the products of all element values ​​in the GLCM matrix and their corresponding frequencies.

[0113] The mean value of the first principal component of the multi-scale segmented region is obtained by calculating the mean value of the first principal component PCA1 within each segmented object after performing multi-scale segmentation processing on the remote sensing image.

[0114] In this embodiment, based on the preprocessed remote sensing image dataset, spectral features, topographic features, texture features, and regional aggregation features are extracted to construct a comprehensive spectral-spatial joint feature dataset. The specific sub-steps are as follows.

[0115] Spectral features are extracted, including the original four-band data, the Soil-Adjusted Vegetation Index (SAVI), and the first principal component (PCA1). The original four-band data specifically refer to the blue light band (450-520nm), green light band (520-590nm), red light band (630-690nm), and near-infrared band (760-900nm) of GF-2 satellite imagery. These bands directly reflect the spectral reflectance characteristics of ground features. The calculation of the Soil-Adjusted Vegetation Index (SAVI) aims to reduce the influence of soil background on the vegetation index. Its calculation formula is: SAVI = (NIR - Red) / (NIR + Red + L) * (1 + L). Where NIR is the near-infrared reflectance, Red is the red light reflectance, and L is the soil adjustment coefficient, whose value can be adjusted between 0 and 1 according to the actual soil cover. In a preferred embodiment of this invention, the L value is set to 0.5 to accommodate medium-density vegetation cover and a moderately bare soil background.

[0116] The first principal component (PCA1) was obtained by performing principal component analysis on the original four-band data. PCA is a commonly used dimensionality reduction technique that can remove redundant information from the original bands and extract the information dimensions that contribute the most to land cover classification. Typically, the first principal component contains most of the variance in the original data and can effectively distinguish different land cover types.

[0117] Topographic features, specifically slope, are extracted. Slope information is obtained by calculating digital elevation model data, and its spatial resolution can be resampled based on the resolution of remote sensing imagery. Slope reflects the degree of undulation of the landform and has a significant impact on vegetation type, soil erosion, and water conditions in mountain pastures, helping to distinguish features at different elevation gradients and topographic locations.

[0118] Texture features are extracted, specifically the mean of the gray-level co-occurrence matrix (GLCM). This is achieved by statistically analyzing the co-occurrence frequency of gray-level values ​​between pixel pairs in a grayscale image of a remote sensing image at specified distances and directions to describe the image's texture information. In the calculation, eight directions can be set: 0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315°, and a distance of one pixel. These GLCMs are then averaged to extract the mean feature. The mean of the GLCM is the sum of the products of all element values ​​and their corresponding frequencies in the GLCM. This mean feature reflects the uniformity and precision of the image's gray-level distribution and plays a crucial role in distinguishing ground features with different spatial arrangements, such as grasslands, shrubs, and bare land.

[0119] Regional aggregation features are extracted, which are the mean of the first principal component (PCA1) of the multi-scale segmented regions. This feature is obtained as follows: First, the remote sensing image is segmented at multiple scales. In this embodiment, the segmentation scale is set to 100, which controls the size of the segmented objects; a larger value tends to generate larger objects. The shape factor is set to 0.1, which determines the weight of shape homogeneity in the segmentation process; a lower value means that color homogeneity is more important. The compactness factor is set to 0.5, which affects the compactness of the segmented objects. By finely setting these parameters, the image can be effectively segmented into semantically meaningful geographic objects, rather than isolated pixels. Second, based on the segmentation results, the mean of the first principal component (PCA1) within each segmented object is calculated. This regional aggregation feature can effectively reduce pixel-level noise, enhance the spatial continuity and integrity of ground objects, capture the overall spectral characteristics of ground objects at a certain spatial scale, and thus improve the robustness of classification.

[0120] Furthermore, referring to Figure 3 S3 converts the spectral-spatial joint feature dataset from a one-dimensional sequence to a two-dimensional image, generating two-dimensional image input data with multi-channel features, including the following sub-steps:

[0121] S31, for each pixel in the spectral-spatial joint feature dataset, its spectral feature sequence and spatial feature sequence are concatenated to form a one-dimensional feature vector;

[0122] S32, the one-dimensional feature vector is taken as input and converted into three two-dimensional feature matrices through Gram angle field transformation, Markov transition field transformation and recursive graph transformation respectively.

[0123] S33, the three two-dimensional feature matrices are spliced ​​together in the channel dimension to generate a two-dimensional image input data with multi-channel features;

[0124] The Gram angle field transformation maps the elements in the one-dimensional feature vector to a polar coordinate system and calculates the cosine values ​​of the sum or difference between different time points, generating a symmetric matrix representing time correlation.

[0125] The Markov transition field transformation is based on a first-order Markov chain, which calculates the transition probabilities between states in the one-dimensional feature vector to generate a two-dimensional matrix representing the state transition mode.

[0126] The recursive graph transformation generates a symmetric matrix representing the dynamic behavior of a time series by calculating the Euclidean distance between points in the one-dimensional feature vector and determining whether state recursion has occurred based on a preset threshold.

[0127] Specifically, for each pixel in the spectral-spatial joint feature dataset, all extracted spectral feature sequences (original four bands, SAVI, PCA1) and spatial feature sequences (slope, GLCM mean, and PCA1 mean of multi-scale segmentation regions) are concatenated in a preset order to form a high-dimensional one-dimensional feature vector. For example, if there are 7 features, each pixel will correspond to a 7-dimensional one-dimensional feature vector.

[0128] Using a one-dimensional eigenvector as input, it is transformed into three different two-dimensional eigenma matrices through Gram angle field transformation, Markov transition field transformation, and recursive graph transformation, respectively.

[0129] Gram angle field transform maps the elements of a one-dimensional feature vector to a polar coordinate system, while Markov transition field transform, based on the principle of a first-order Markov chain, statistically analyzes the transition probabilities between states in the one-dimensional feature vector. First, the numerical range of the feature vector is divided into predefined discrete intervals, each interval representing a state. Then, the number of transitions from state i to state j is calculated and normalized to a transition probability, generating a two-dimensional matrix representing the feature state transition pattern. This matrix effectively captures the dynamic evolution of the feature sequence. Recursive graph transform calculates the Euclidean distance between any two points in the one-dimensional feature vector (i.e., the feature dimension) and determines whether state recursion has occurred (i.e., the two points are sufficiently close in the feature space) based on a preset distance threshold. If the distance is less than the threshold, the corresponding position in the two-dimensional matrix is ​​marked as 1; otherwise, it is marked as 0. The generated symmetric matrix intuitively represents the similarity or repetition patterns of the time series feature vector at different times, revealing its inherent dynamic behavior and periodicity.

[0130] The three-dimensional and two-dimensional feature matrices obtained after GAF, MTF, and RP transformations are concatenated along the channel dimension. For example, if each two-dimensional matrix is ​​NxN in size, the concatenation will result in an NxNx3 two-dimensional image input data with multi-channel features. This multi-channel two-dimensional image input data integrates information encoded from the original one-dimensional feature vectors from different angles, providing richer and more structured input for subsequent deep learning models, and effectively improving the model's ability to recognize complex patterns.

[0131] Furthermore, S4 constructs and trains an improved Swin-UNet model, as follows:

[0132] The encoder consists of four stages, each stage including a block merging layer and one or more sliding window transformer modules (STBs), and each STB is connected to a local perception enhancement module (LPEM).

[0133] The bottleneck module is located between the encoder and the decoder, and consists of a block merging layer and one or more sliding window transformer modules.

[0134] The decoder consists of four stages, each stage including a block expansion layer and one or more sliding window transformer modules, and each decoder stage includes a cross-layer feature fusion module (CLFFM).

[0135] The plurality of skip connections are respectively used to transmit the feature map of each stage of the encoder to the corresponding stage of the CLFFM of the decoder;

[0136] The output layer of the improved Swin-UNet model consists of a 1x1 convolutional layer and a Softmax activation function layer.

[0137] Furthermore, the structure of the Local Perception Enhancement Module (LPEM) includes:

[0138] A feature input interface is provided for receiving feature maps from the sliding window transformer module;

[0139] A multi-branch dilated convolutional unit contains multiple parallel convolutional branches, each of which employs a 3x3 depthwise separable convolution and sets different dilation rates to capture local contextual information at different scales.

[0140] The adaptive spatial feature fusion unit utilizes a spatial attention mechanism to dynamically emphasize locally important spatial regions in the feature map;

[0141] The channel splicing layer is used to splice the output of the multi-branch dilated convolutional unit with the output of the adaptive spatial feature fusion unit in the channel dimension;

[0142] A 1x1 convolutional layer is used to compress the channel dimension and fuse the concatenated features, and serve as the feature output of the LPEM;

[0143] The structure of the cross-layer feature fusion module CLFFM includes:

[0144] The encoder feature input interface is used to receive feature maps from the corresponding stage of the encoder;

[0145] The decoder upsampled feature input interface is used to receive the feature map upsampled from the previous stage of the decoder;

[0146] The channel alignment module adjusts the number of channels of the feature map from the encoder's feature input interface using a 1x1 convolutional layer to match the number of channels of the feature map from the decoder's upsampled feature input interface.

[0147] The hollow space pyramid pooling ASPP module receives the channel-aligned encoder feature map and the decoder upsampled feature map as input, and aggregates multi-scale context information through multiple parallel branches;

[0148] The channel splicing layer is used to splice the outputs of the various branches of the ASPP module along the channel dimension.

[0149] A 1x1 convolutional layer is used to perform channel dimension compression and feature fusion on the concatenated features;

[0150] The weighted summation module is used to perform a weighted summation of the features fused by the ASPP module and the upsampled feature map of the decoder.

[0151] The channel-spatial attention module receives the weighted summed features as input, generates a refined feature map, and uses it as the feature output of the CLFFM.

[0152] In this embodiment, an improved Swin-UNet model is built and trained. The improved Swin-UNet model includes an encoder, a bottleneck module, a decoder, and several skip connections. The encoder integrates a Local Perception Enhancement (LPEM) module, and the decoder integrates a Cross-Layer Feature Fusion (CLFFM) module. The specific structure is as follows: The encoder consists of four stages, each responsible for downsampling and feature extraction of the input feature map. Each stage contains a PatchMerging layer, which reduces the spatial resolution of the feature map by half and increases the number of channels. For example, merging 2x2 pixels into a patch, and its feature vector is linearly projected to become a new channel. Furthermore, each stage contains one or more SwinTransformerBlocks (STBs). The STB is the core unit of the SwinTransformer, which effectively captures local and cross-window feature dependencies through window-based multi-head self-attention and shifted-window multi-head self-attention mechanisms. In a specific implementation, each STB is followed by a Local Perception Enhancement (LPEM) module.

[0153] The LPEM architecture is designed to enhance the model's ability to capture local texture details and edge information. It includes: a feature input interface for receiving feature maps from a SwinTransformerBlock; a multi-branch dilated convolutional unit containing multiple parallel convolutional branches, each employing a 3x3 depthwise separable convolution; and a multi-branch dilated convolutional unit that effectively reduces the number of parameters and computational overhead by decomposing standard convolution into depthwise and pointwise convolutions while maintaining feature extraction capabilities. This multi-branch unit sets different dilation rates, for example, 1, 2, and 4. A convolution with a dilation rate of 1 captures immediate local information, a convolution with a dilation rate of 2 expands the receptive field to capture slightly more distant context, and a convolution with a dilation rate of 4 further expands the receptive field, thereby capturing local contextual information at different scales without increasing the number of parameters. This is particularly crucial for distinguishing features with subtle texture differences. Finally, an adaptive spatial feature fusion unit utilizes a spatial attention mechanism to dynamically emphasize locally important spatial regions in the feature map. This unit typically performs global average pooling and global max pooling on the input feature map, then concatenates the two pooling results and generates a spatial attention map through a convolutional layer and a sigmoid activation function. This attention map is then multiplied element-wise with the original feature map to enhance focus on key spatial locations. A channel concatenation layer concatenates the outputs of the multi-branch dilated convolutional unit and the adaptive spatial feature fusion unit along the channel dimension to fuse local features from different branches. Finally, a 1x1 convolutional layer performs channel dimension compression and feature fusion on the concatenated features, serving as the feature output of the LPEM to match the number of input channels in subsequent modules.

[0154] The bottleneck module is located between the encoder and the decoder. Its structure is similar to a stage of the encoder. It consists of a PatchMerging layer and one or more SwinTransformerBlocks, and is responsible for extracting the highest level of semantic features.

[0155] The decoder also consists of four stages, each designed to progressively restore the spatial resolution of the feature map and refine the features through upsampling. Each stage contains a PatchExpanding layer, which doubles the spatial resolution of the feature map and reduces the number of channels, for example, by upsampling through pixel rearrangement or transposed convolution. Each decoder stage also includes a cross-layer feature fusion module.

[0156] The CLFFM architecture is designed to enhance the efficient fusion of multi-scale contextual information. It comprises: an encoder feature input interface for receiving feature maps from the corresponding stage of the encoder, which typically contain rich low-level spatial details; a decoder upsampled feature input interface for receiving upsampled feature maps from the previous stage of the decoder, which contain high-level semantic information; and a channel alignment module that adjusts the number of channels of the feature maps from the encoder feature input interface using a 1x1 convolutional layer to match the number of channels of the feature maps from the decoder upsampled feature input interface, ensuring dimensionality consistency in subsequent fusion. An Aperture Spatial Pyramid Pooling (ASPP) module receives channel-aligned encoder feature maps and decoder upsampled feature maps as input. It aggregates multi-scale contextual information through multiple parallel branches: a 1x1 convolutional branch to capture local information; three 3x3 dilated convolutional branches with dilation rates of 6, 12, and 18, capturing contextual information at different scales through convolutional kernels with varying dilation rates, effectively expanding the receptive field without loss of resolution; and a global average pooling branch whose output is upsampled to the original feature map size after a 1x1 convolution to incorporate global contextual information. A channel concatenation layer concatenates the outputs of the various branches of the ASPP module along the channel dimension, integrating contextual features at different scales. A 1x1 convolutional layer performs channel-dimensional compression and feature fusion on the concatenated features to generate a compact and information-rich feature representation. A weighted summation module performs a weighted summation of the features fused by the ASPP module and the decoder upsampled feature map. This weighted summation mechanism allows the model to dynamically adjust the feature contributions from the encoder (containing more details) and decoder to better balance detail and semantic information. Finally, a channel-space attention module receives the weighted summed features as input and dynamically focuses on the most informative channels and spatial regions to generate a refined feature map, which serves as the feature output of CLFFM. This attention module achieves adaptive feature enhancement by separately calculating weights in the channel and spatial dimensions of the feature map.

[0157] Several skip connections are used to transfer the feature maps from each stage of the encoder to the corresponding CLFFM stage of the decoder, effectively transferring the fine spatial details preserved in the encoder to the decoder and compensating for the loss of detail during downsampling. The output layer of the improved Swin-UNet model consists of a 1x1 convolutional layer and a Softmax activation function layer. The 1x1 convolutional layer converts the number of channels in the final output feature map of the decoder into the number of target classification categories, while the Softmax activation function layer converts these values ​​into pixel-level classification probability maps of the mountain pasture features, representing the probability that each pixel belongs to each category.

[0158] In a preferred embodiment of the present invention, an improved Swin-UNet model is constructed and trained. The training process employs the following parameters and environment settings to ensure optimal model performance and reproducibility of results. The dataset, consisting of preprocessed multi-channel two-dimensional image input data and corresponding pixel-level ground truth labels, is divided into training, validation, and test sets in a strict 8:1:1 ratio. That is, 80% is used for model training, 10% for model validation to adjust hyperparameters and prevent overfitting, and 10% for final model performance evaluation.

[0159] Before inputting the data into the model, Min-Max normalization is performed to standardize all feature values ​​to the range of 0-1. This process helps accelerate model convergence and improve training stability. The cross-entropy loss function is chosen, as it is suitable for multi-class classification tasks and effectively measures the difference between the model's predicted probability distribution and the true label distribution. The AdamW optimizer is selected, which combines Adam's adaptive learning rate and weight decay regularization effects to effectively avoid overfitting and accelerate the training process.

[0160] Further, step S5 uses the trained and improved Swin-UNet model to classify mountain pasture features from the two-dimensional image input data to obtain a high-precision pasture feature classification map, including the following sub-steps:

[0161] S51, For the pasture image dataset during the lush pasture season, the pasture feature classification map output by the improved Swin-UNet model contains six types of land features: grassland, shrubs, river-lake, cloud, bare land, and shadow.

[0162] S52, For the dry season image dataset, the pasture feature classification map output by the improved Swin-UNet model contains five types of land features: grassland, shrubs, river-lake-cloud-snow, bare land, and shadow.

[0163] After the pasture element classification map is generated, it is post-processed using morphological operations or Conditional Random Field (CRF) to eliminate the salt and pepper effect.

[0164] In this embodiment, the improved Swin-UNet model, after training, is used to classify mountain pasture features from the two-dimensional image input data to obtain a high-precision pasture feature classification map. The specific operation includes the following sub-steps:

[0165] For the pastureland image dataset, the improved Swin-UNet model outputs a pasture feature classification map containing six land cover categories. These six land cover categories are: grassland, referring to areas covered with various types of grasses; shrubs, referring to vegetation cover areas dominated by shrubs; rivers and lakes, including water bodies such as rivers, lakes, and reservoirs; clouds, referring to areas in the image covered by clouds; bare land, including bare soil, rocks, sand, and other areas without vegetation cover; and shadows, referring to shadowed areas caused by clouds, mountains, or other tall objects.

[0166] For the dry season image dataset, the improved Swin-UNet model outputs a pasture feature classification map containing five land cover categories. These five categories are: grassland, which is mostly withered or low-lying winter grass; shrubs, whose spectral characteristics differ from those during the wet season; river-lake-cloud-snow, as clouds and snow have a more pronounced influence on spectral characteristics during the dry season and can cause some spectral confusion with water bodies, so these three categories are combined; bare land; and shadows. After generation, the pasture feature classification map undergoes post-processing using morphological operations or conditional random fields to eliminate salt-and-pepper effects in the classification results, i.e., discrete isolated pixels or small patches, and to enhance the spatial continuity of land cover. Morphological opening operations, through erosion followed by dilation, can eliminate small bright spots and fine connections while smoothing object edges; morphological closing operations, through dilation followed by erosion, can fill small holes and breaks, making object boundaries more complete. The kernel size for these two operations can be adjusted according to the scale of the actual classification error, typically using a 3x3 or 5x5 square kernel. Conditional random fields (CRFs) optimize pixel-level classification results by introducing spatial relationships between pixels, making it more likely that adjacent pixels have the same label, thereby generating smoother and more continuous classification boundaries.

[0167] Further, step S6, based on the pasture element classification map, performs a dynamic assessment of the grass-livestock balance to determine the grass-livestock balance status level of the mountain pasture, including the following sub-steps:

[0168] S61, Calculate the area of ​​grassland elements based on the pasture element classification map;

[0169] S62, Based on the grassland area and field survey data, calculate the theoretical carrying capacity R and the actual carrying capacity A of the mountain pasture;

[0170] S63, Calculate the grass-livestock balance index BGLI of the mountain pasture based on the theoretical carrying capacity R and the actual carrying capacity A;

[0171] S64, Determine the grass-livestock balance status level of the mountain pasture based on the value of the grass-livestock balance index (BGLI);

[0172] The theoretical carrying capacity R is calculated as follows: R=(Y*E) / (I*D), where R is the reasonable carrying capacity of the grassland during the grazing period, Y is the total grass yield of the grassland, E is the grassland utilization rate, I is the daily feed intake of sheep, and D is the grazing duration.

[0173] The formula for calculating the grass-livestock balance index (BGLI) is: BGLI = (AR) / R;

[0174] The grass-livestock balance status is divided into five levels:

[0175] Extreme overloading, when the BGLI is greater than 150%; severe overloading, when the BGLI is greater than 80% and less than or equal to 150%; overloading, when the BGLI is greater than 20% and less than or equal to 80%; grass-livestock balance, when the BGLI is greater than -20% and less than or equal to 20%; underloading, when the BGLI is less than or equal to -20%.

[0176] In this embodiment, a dynamic assessment of the grass-livestock balance is conducted based on the pasture element classification map to determine the grass-livestock balance status level of the mountain pasture. The specific sub-steps are as follows:

[0177] Based on the pasture feature classification map, the area of ​​grassland features is accurately calculated. By counting all pixels marked as grassland in the classification map and multiplying by the actual ground area of ​​a single pixel (for example, for an image with a resolution of 0.8 meters, each pixel represents 0.64 square meters), the total grassland area can be obtained, expressed in hectares or square kilometers.

[0178] Based on grassland area and field survey data, the theoretical carrying capacity (R) and actual carrying capacity (A) of mountain pastures are calculated. The theoretical carrying capacity (R) refers to the reasonable carrying capacity of the grassland during the grazing season, expressed in standard sheep units (SU). The calculation formula is: R = (Y*E) / (I*D). Where Y is the grassland yield (kg), which during the wet season is the grassland area obtained from the classification results multiplied by the dry hay weight per unit area (kg / m²). For example, the dry hay weight of alpine meadows can be set at 200-300 g / m². During the dry season, the yield is set as the remaining hay from the wet season, generally estimated at 20%-30% of the wet season yield to reflect the available forage resources in winter. E represents grassland utilization rate (%), which is based on the recommended utilization rates for different grassland types in the national agricultural industry standard NY / T635-2015. For example, it can be set at 0.3-0.4 for degraded grassland and 0.5-0.6 for high-quality grassland. I represents daily feed intake per sheep unit, usually set at 1.8 kg of standard hay / (SU·d). The conversion relationship for standard sheep units for different livestock types is as follows: 1 adult cow equals 6 standard sheep units, 1 horse equals 5 standard sheep units, and 1 goat or sheep equals 1 standard sheep unit. D represents grazing duration, which is set at 5 months (150 days) during the abundant grass season and 7 months (210 days) during the dry grass season. Actual carrying capacity A refers to the number of standardized livestock actually grazing on a certain area of ​​grassland within a certain utilization period. This data is usually obtained through on-site livestock surveys, herder questionnaires, or statistical data from livestock management departments, and is also converted to standard sheep units.

[0179] Based on the theoretical carrying capacity R and the actual carrying capacity A, the grassland-livestock balance index (BGLI) for mountain pastures is calculated. The formula for BGLI is: BGLI = (AR) / R * 100. This index directly reflects the degree of deviation between the actual carrying capacity and the theoretical carrying capacity. Section S6.4 determines the grassland-livestock balance status level of mountain pastures based on the BGLI value. The grassland-livestock balance status level is divided into five levels: extremely overloaded (BGLI > 150%); severely overloaded (BGLI > 80% and <= 150%); overloaded (BGLI > 20% and <= 80%); balanced (BGLI > -20% and <= 20%); and underloaded (BGLI <= -20%). This refined classification helps pastoral management departments make targeted decisions.

[0180] According to a second embodiment of the present invention, referring to Figure 4 This invention claims protection for a mountain pasture element extraction system, comprising:

[0181] The remote sensing image preprocessing module is used to acquire and preprocess remote sensing image data to generate a multi-temporal, high-precision remote sensing image dataset with uniform radiometric and geometric characteristics.

[0182] The spectral-spatial feature extraction module is used to extract spectral features, topographic features, texture features, and regional aggregation features from the remote sensing image dataset output by the remote sensing image preprocessing module, so as to construct a comprehensive spectral-spatial joint feature dataset.

[0183] The feature 2D conversion module is used to convert the spectral-spatial joint feature dataset output by the spectral-spatial feature extraction module from a one-dimensional sequence to a two-dimensional image, generating two-dimensional image input data with multi-channel features;

[0184] An improved Swin-UNet model is used to receive two-dimensional image input data output by the feature 2D transformation module and output a mountain pasture feature classification map. The improved Swin-UNet model includes an encoder, a bottleneck module, a decoder and several skip connections. The encoder integrates a local perception enhancement module and the decoder integrates a cross-layer feature fusion module.

[0185] The grass-livestock balance dynamic assessment module is used to perform a grass-livestock balance dynamic assessment based on the pasture element classification map output by the improved Swin-UNet model, so as to determine the grass-livestock balance status level of the mountain pasture.

[0186] The aforementioned mountain pasture element extraction system is used to execute the aforementioned mountain pasture element extraction method.

[0187] This embodiment selects a typical alpine pastoral area on the Qinghai-Tibet Plateau in China, covering a geographical area of ​​approximately 200 square kilometers. This area has complex terrain and diverse landforms, including alpine meadows, shrublands, exposed rocks, seasonal rivers and lakes, and scattered farmland and settlements. This embodiment aims to verify the classification accuracy and grassland-livestock balance assessment capabilities of the mountain pasture element extraction method of this invention in practical applications.

[0188] Remote sensing imagery data from Gaofen-2 (GF-2) and Gaofen-7 (GF-7) satellites were acquired between 2020 and 2022, covering the grassy growth period (July-August) and the dry season (November-March of the following year). The GF-2 imagery has a panchromatic resolution of 0.8 meters and a multispectral resolution of 3.2 meters, while the GF-7 imagery has a panchromatic resolution of 0.65 meters and a multispectral resolution of 2.6 meters. DEM data were obtained using ASTERGDEMV3 with a resolution of 30 meters. All raw images underwent radiometric and atmospheric corrections (using the ENVIFLAASH module) and were geometrically registered to the WGS84 / UTM coordinate system with an accuracy better than 0.5 pixels. Spatial resolution was improved by fusing the GF-2 / GF-7 multispectral and panchromatic bands using the Gram-Schmidt algorithm. Image stitching and histogram matching ensured data consistency. The final result is a high-resolution remote sensing image dataset of 20 scenes covering both the grassy growth and dry seasons.

[0189] For each pixel, the following features are extracted from the preprocessed image data:

[0190] Spectral characteristics: reflectance of the four original bands of GF-2 / GF-7 (blue, green, red, and near-infrared); soil modified vegetation index (SAVI), with L value set to 0.5; PCA analysis of the original four bands to extract the first principal component PCA1.

[0191] Topographic features: Slope values ​​calculated based on ASTERGDEM data.

[0192] Texture features: Calculate the gray-level co-occurrence matrix (GLCM) for the red band grayscale image of GF-2 / GF-7, taking the average distance of 1 pixel and the average direction as the mean of the GLCM in all 8 directions.

[0193] Region aggregation features: The GF-2 fused image was segmented at multiple scales, with a segmentation scale of 100, a shape factor of 0.1, and a compactness factor of 0.5. Then, the mean PCA1 value within each segmented object was calculated.

[0194] The above 7 features, 4 original bands, SAVI, PCA1, slope, GLCM mean, and multi-scale PCA1 mean are extracted for each pixel to form a 7-dimensional one-dimensional feature vector.

[0195] The 7-dimensional and 1-dimensional feature vectors of each pixel are used as input, and GAF, MTF, and RP transformations are performed respectively.

[0196] GAF transformation: Normalizes the 7-dimensional feature vector to the interval [-1,1] and maps it to polar coordinate angles to generate a 7x7 two-dimensional matrix.

[0197] MTF Transform: The numerical range of the 7-dimensional feature vector is divided into 10 equally wide intervals as states. The first-order transition probabilities between each state are calculated, generating a 10x10 two-dimensional matrix. To match the output dimension of GAF / RP, it is resampled to 7x7 using bilinear interpolation.

[0198] RP Transformation: Calculate the Euclidean distance between points in the 7-dimensional feature vector, set the distance threshold to 0.15 times the standard deviation of the feature values, and generate a 7x7 symmetric matrix.

[0199] These three 7x7 two-dimensional feature matrices are concatenated along the channel dimension to generate a 7x7x3 multi-channel two-dimensional image input data.

[0200] Construction and training of the improved Swin-UNet model:

[0201] Based on the PyTorch framework, the improved Swin-UNet model of this invention is constructed.

[0202] Encoder: Each SwinTransformerBlock is followed by an LPEM. The multi-branch dilated convolutional units of the LPEM use 3x3 depth-separable convolutions with dilation rates set to 1, 2, and 4.

[0203] Decoder: Each stage embeds CLFFM, and the ASPP module of CLFFM uses 1x1 convolution, three 3x3 dilated convolutions with dilation rates of 6, 12, and 18, and a global average pooling branch.

[0204] The training parameters are configured as follows:

[0205] Dataset partitioning: Training set: Validation set: Test set = 8:1:1.

[0206] Data normalization: Min-Max normalization to 0-1.

[0207] Loss function: Cross-entropy loss function.

[0208] Optimizer: AdamW, initial learning rate 1e-5, beta1=0.9, beta2=0.999, epsilon=1e-8, weight decay 0.01.

[0209] Learning rate strategy: cosine annealing with hot restart, initial restart period of 20 epochs, period growth factor of 2, minimum learning rate of 1e-6.

[0210] Batch size: 150.

[0211] Hardware environment: NVIDIA RTX 3090 GPU.

[0212] The model was trained on both the lush grass season dataset and the dry grass season dataset. The lush grass season dataset was labeled with six categories: grassland, shrubs, river-lake, cloud, bare land, and shade. The dry grass season dataset was labeled with five categories: grassland, shrubs, river-lake-cloud-snow, bare land, and shade.

[0213] After training, the improved Swin-UNet model was used to classify the test set images. Morphological opening and closing operations were performed on the output classification images for post-processing, using a 3x3 square kernel and one iteration to eliminate small patch noise.

[0214] On the test set, the overall precision of the classification results during the lush grass stage reached 92.14%, with an average intersection-union ratio of 85.32% and a Kappa coefficient of 0.89. The OA of the classification results during the dry grass stage was 89.76%, the mIoU was 82.11%, and the Kappa coefficient was 0.86. Notably, the F1-scores for key land cover types such as grassland, shrubs, and bare land all exceeded 0.88.

[0215] Accurately calculate grassland area based on classification map.

[0216] The lush pasture season is from July to September: the extracted grassland area is 120.5 square kilometers. According to field surveys, the average dry hay weight of the alpine meadow is 250 g / m², and the grassland utilization rate is taken as 0.5. The average daily feed intake per sheep is 1.8 kg, and the grazing period is 5 months (150 days).

[0217] Theoretical carrying capacity R_{lush pasture period} = (120.5*10^6m^2*0.25kg / m^2*0.5) / (1.8kg / SU*150d) = 55787SU.

[0218] Dry season (October to April of the following year): Grassland area is 115.2 square kilometers (considering some seasonal variations). Grass yield during the dry season is estimated at 25% of the remaining yield during the wet season, i.e., 0.25 * 0.25 kg / m² = 0.0625 kg / m². Grassland utilization rate is taken as 0.35. Grazing duration is 7 months (210 days).

[0219] Theoretical carrying capacity R_{dry season} = (115.2 * 10^6 m^2 * 0.0625 kg / m^2 * 0.35) / (1.8 kg / SU * 210 d) = 6614 SU.

[0220] Field surveys and statistics show that the actual carrying capacity of the pastoral area during the wet season is 72,500 SU, and the actual carrying capacity during the dry season is 8,500 SU.

[0221] Calculate the grass-livestock balance index:

[0222] BGLI during the lush grass season: (72500-55787) / 55787*100=29.95%.

[0223] BGLI during the dry season: (8500-6614) / 6614 = approximately 28.52%.

[0224] According to the BGLI classification, both the abundant grass period and the dry grass period are classified as "overgrazing" (greater than 20% and less than or equal to 80%). This result is highly consistent with the long-term monitoring and expert evaluation conclusions of pastoral management departments, indicating that this method can accurately reflect the dynamic status of the grass-livestock balance in pastoral areas.

[0225] To further demonstrate the superiority of the technical solution of this invention, this comparative example uses a method based on the standard Swin-UNet model, without integrating LPEM and CLFFM, and combining traditional feature input for mountain pasture feature extraction and grassland-livestock balance assessment. This method uses only the original four-band data and SAVI as spectral features, and directly uses these spectral features as input to the standard Swin-UNet, without performing 1D-2D feature transformation or extracting terrain, texture, and region aggregation features. Other data preprocessing steps, training parameters, test areas, and evaluation metrics are consistent with the example.

[0226] Only the original four-band data and SAVI, totaling five spectral features, were extracted and directly stitched together to form a multi-channel input. Taking a remote sensing image acquired on September 2, 2023, as an example, the original four spectral bands and the fused spectral-spatial features were input into the classification model for land cover identification. The resulting overall classification results for the study area are as follows: Figure 5 As shown, Figure 5 (a) represents the classification results of the original four-band input; Figure 5 (b) Represents the classification result of the spectral-spatial feature input; local classification details are as follows: Figure 6 As shown, Figure 6 (a) and 6(d) are local original images in true color; Figure 6 (b) and 6(e) are the classification results of the original four-band input in the local area; Figure 6 (c) Figure 6 (f) shows the classification results based on local spectral-spatial feature input. Model performance is quantitatively evaluated using accuracy metrics.

[0227] A standard Swin-UNet model was constructed, whose encoder and decoder do not include the LPEM and CLFFM modules described in this invention. Other model architectures, skip connections, and training parameters remain consistent with the Swin-UNet basic architecture in the embodiments. The model was trained on both the abundant grass season dataset and the dry grass season dataset, with the same classification labels as in the embodiments.

[0228] After training, the standard Swin-UNet model was used to classify the test set images. Morphological post-processing was also performed.

[0229] On the test set, the overall accuracy (OA) of the classification results during the lush grass stage was 85.21%, the mean intersection-over-union ratio (mIoU) was 76.58%, and the Kappa coefficient was 0.80. During the dry grass stage, the OA was 83.15%, the mIoU was 73.92%, and the Kappa coefficient was 0.76. The F1-scores for key land cover types such as grassland, shrubs, and bare land were generally below 0.80, and the classification results showed more salt-and-pepper effect and blurred boundaries.

[0230] Calculate grassland area based on the standard Swin-UNet classification map.

[0231] Grassland lush period (July-September): The extracted grassland area was 112.8 square kilometers. Other parameters were the same as in the previous example.

[0232] Theoretical carrying capacity R_{lush pasture period} = (112.8*10^6m^2*0.25kg / m^2*0.5) / (1.8kg / SU*150d) = 52222SU.

[0233] Dry season (October to April of the following year): grassland area was 105.1 square kilometers. Other parameters were the same as in the previous example.

[0234] Theoretical carrying capacity R_{dry season} = (105.1*10^6m^2*0.0625kg / m^2*0.35) / (1.8kg / SU*210d) = 6031SU.

[0235] The actual carrying capacity was the same as in the example: 72,500 SU during the abundant grass period and 8,500 SU during the dry grass period.

[0236] Calculate the grass-livestock balance index:

[0237] BGLI during the lush grass season: (72500-52222) / 52222*100=38.84%.

[0238] BGLI during the dry season: (8500-6031) / 6031*=40.95%.

[0239] According to the BGLI classification, both the abundant pasture period and the dry pasture period are classified as "overgrazing". However, compared with the example, the BGLI value is higher, which means that the standard Swin-UNet, due to insufficient classification accuracy, underestimates the grassland area, resulting in a lower theoretical carrying capacity, and thus exaggerates the degree of overgrazing, which may lead to inaccurate pasture management decisions.

[0240] The mountain pasture feature extraction method of this invention significantly outperforms the comparative method in all classification accuracy metrics. The overall accuracy during the lush pasture period is improved by 6.93 percentage points, and the average intersection-union ratio (IUR) is improved by 8.74 percentage points; the overall accuracy during the dry pasture period is improved by 6.61 percentage points, and the IUR is improved by 8.19 percentage points. The Kappa coefficient and F1-scores for each major land cover type also show similar significant advantages, indicating that the method of this invention can more accurately and precisely identify and distinguish land cover types in complex mountain pasture environments. This improvement in classification accuracy is attributed to the effective fusion of multi-dimensional spectral-spatial joint features, particularly through the innovative 1D-to-2D conversion, which provides richer and more structured input information for the deep learning model. Furthermore, the improved Swin-UNet model's enhanced perception of local details via LPEM and deep fusion of multi-scale features via CLFFM effectively address the shortcomings of traditional methods in handling fine textures and complex land cover boundaries, reducing misclassification and enhancing the spatial continuity of the classification results.

[0241] Regarding grassland-livestock balance assessment, although both methods indicate that pastoral areas are in an "overgrazing" state, the BGLI value calculated by the method of this invention is closer to the actual situation assessed by experts. Its BGLI during the abundant grass season and the dry grass season are 29.95% and 28.52%, respectively, which are more reasonable compared to the comparative method's 38.84% and 40.95%. The comparative method, due to its lower classification accuracy, may have underestimated grassland area, thus underestimating the theoretical carrying capacity and consequently overestimating the calculated BGLI, exaggerating the degree of overgrazing. This further highlights the crucial supporting role of high-precision pasture element classification in the scientific assessment of grassland-livestock balance. The more accurate classification results provided by the method of this invention can provide pastoral managers with more reliable and refined decision-making basis, which has important practical significance for maintaining grassland ecological balance and achieving sustainable livestock development.

[0242] In summary, this invention effectively overcomes the key problems of insufficient local detail perception and poor multi-scale context fusion in the fine classification of remote sensing images of complex mountain pastures by employing an innovative multi-dimensional spectral-spatial joint feature input strategy, advanced 1D-to-2D feature conversion technology, and a deeply optimized Swin-UNet model structure. The proposed method significantly improves classification accuracy, reduces misclassification, and enhances spatial continuity and robustness in pixel-level classification tasks of high-resolution remote sensing images. Finally, combined with high-precision pasture element extraction results, the refined grassland-livestock balance dynamic assessment model constructed in this invention provides solid technical support for the scientific management of grasslands and grassland-livestock balance monitoring in fragile ecological areas such as the Qinghai-Tibet Plateau.

[0243] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0244] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made based on the description and drawings of this application, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

[0245] The specific embodiments of the invention have been described in detail above, but they are only examples, and this application is not limited to the specific embodiments described above. For those skilled in the art, any equivalent modifications or substitutions to the invention are also within the scope of this application. Therefore, all equivalent changes, modifications, and improvements made without departing from the spirit and principles of this application should be covered within the scope of this application.

Claims

1. A method for extracting elements from mountain pastures, characterized in that, Includes the following steps: S1: Acquire remote sensing image data and preprocess it to generate a multi-temporal, high-precision remote sensing image dataset with uniform radiometric and geometric characteristics. S2, Based on the preprocessed remote sensing image dataset, extract spectral features, terrain features, texture features and regional aggregation features to construct a comprehensive spectral-spatial joint feature dataset; S3, perform a one-dimensional sequence to two-dimensional image conversion on the spectral-spatial joint feature dataset to generate two-dimensional image input data with multi-channel features; S4. Construct and train an improved Swin-UNet model. The improved Swin-UNet model includes an encoder, a bottleneck module, a decoder, and several skip connections. The encoder integrates a local perception enhancement module, and the decoder integrates a cross-layer feature fusion module. S5. The improved Swin-UNet model, after training, is used to classify the mountain pasture elements in the two-dimensional image input data to obtain a high-precision pasture element classification map. S6. Based on the pasture element classification map, conduct a dynamic assessment of the grass-livestock balance to determine the grass-livestock balance status level of the mountain pasture. The region aggregation feature is the mean of the first principal component of the multi-scale segmented region; The structure of the Local Perception Enhancement Module (LPEM) includes: A feature input interface is provided for receiving feature maps from the sliding window transformer module; A multi-branch dilated convolutional unit contains multiple parallel convolutional branches, each of which employs a 3x3 depthwise separable convolution and sets different dilation rates to capture local contextual information at different scales. The adaptive spatial feature fusion unit utilizes a spatial attention mechanism to dynamically emphasize locally important spatial regions in the feature map; The channel splicing layer is used to splice the output of the multi-branch dilated convolutional unit with the output of the adaptive spatial feature fusion unit in the channel dimension; A 1x1 convolutional layer is used to compress the channel dimension and fuse the concatenated features, and serve as the feature output of the LPEM; The structure of the cross-layer feature fusion module CLFFM includes: The encoder feature input interface is used to receive feature maps from the corresponding stage of the encoder; The decoder upsampled feature input interface is used to receive the feature map upsampled from the previous stage of the decoder; The channel alignment module adjusts the number of channels of the feature map from the encoder's feature input interface using a 1x1 convolutional layer to match the number of channels of the feature map from the decoder's upsampled feature input interface. The hollow space pyramid pooling ASPP module receives the channel-aligned encoder feature map and the decoder upsampled feature map as input, and aggregates multi-scale context information through multiple parallel branches; The channel splicing layer is used to splice the outputs of the various branches of the ASPP module along the channel dimension. A 1x1 convolutional layer is used to perform channel dimension compression and feature fusion on the concatenated features; The weighted summation module is used to perform a weighted summation of the features fused by the ASPP module and the upsampled feature map of the decoder. The channel-spatial attention module receives the weighted summed features as input, generates a refined feature map, and uses it as the feature output of the CLFFM. The S4 constructs and trains an improved Swin-UNet model, as follows: The encoder consists of four stages, each stage including a block merging layer and one or more sliding window transformer modules (STBs), and each STB is connected to a local perception enhancement module (LPEM). The bottleneck module is located between the encoder and the decoder, and consists of a block merging layer and one or more sliding window transformer modules. The decoder consists of four stages, each stage including a block expansion layer and one or more sliding window transformer modules, and each decoder stage includes a cross-layer feature fusion module (CLFFM). The plurality of skip connections are respectively used to transmit the feature map of each stage of the encoder to the corresponding stage of the CLFFM of the decoder; The output layer of the improved Swin-UNet model consists of a 1x1 convolutional layer and a Softmax activation function layer.

2. The method according to claim 1, characterized in that, S1 includes the following sub-steps: S11, acquire high-resolution remote sensing images, the image data including images of the dry grass period and images of the lush grass period, wherein the high-resolution remote sensing images are images from the Gaofen-2 satellite and images from the Gaofen-7 satellite. S12, perform standardized preprocessing on the high-resolution remote sensing image, the standardized preprocessing includes radiometric correction, atmospheric correction, geometric correction, image fusion, image mosaicking, histogram matching and spatial alignment. S13, the preprocessed remote sensing images are divided into a grassy season image dataset and a grassy dry season image dataset according to the acquisition time. The grassy season is set to May to September each year, and the grassy dry season is set to October to April of the following year.

3. The method according to claim 1, characterized in that, S2 extracts spectral features, topographic features, texture features, and regional aggregation features based on the preprocessed remote sensing image dataset to construct a comprehensive spectral-spatial joint feature dataset; The spectral features include raw four-band data, soil-modified vegetation index (SAVI), and first principal component (PCA1). The raw four-band data includes blue light, green light, red light, and near-infrared light. The terrain feature is slope; The texture feature is the mean of the gray-level co-occurrence matrix (GLCM).

4. The method according to claim 3, characterized in that, Also includes: The soil-regulated vegetation index (SAVI) is calculated as follows: SAVI = (NIR - Red) / (NIR + Red + L) * (1 + L), where NIR is the near-infrared reflectance, Red is the red light reflectance, and L is the soil regulation coefficient, which ranges from 0 to 1. The first principal component, PCA1, was obtained by performing principal component analysis on the original four-band data. The slope is obtained by calculation from the digital elevation model (DEM) data; The mean value of the gray-level co-occurrence matrix (GLCM) is obtained by calculating the gray-level image of the remote sensing image. The mean value of the GLCM is the sum of the products of all element values ​​in the GLCM matrix and their corresponding frequencies. The mean value of the first principal component of the multi-scale segmented region is obtained by calculating the mean value of the first principal component PCA1 within each segmented object after performing multi-scale segmentation processing on the remote sensing image.

5. The method according to claim 1, characterized in that, Step S3 converts the spectral-spatial joint feature dataset from a one-dimensional sequence to a two-dimensional image, generating two-dimensional image input data with multi-channel features, including the following sub-steps: S31, for each pixel in the spectral-spatial joint feature dataset, its spectral feature sequence and spatial feature sequence are concatenated to form a one-dimensional feature vector; S32, the one-dimensional feature vector is taken as input and converted into three two-dimensional feature matrices through Gram angle field transformation, Markov transition field transformation and recursive graph transformation respectively. S33, the three two-dimensional feature matrices are spliced ​​together in the channel dimension to generate a two-dimensional image input data with multi-channel features; The Gram angle field transformation maps the elements in the one-dimensional feature vector to a polar coordinate system and calculates the cosine values ​​of the sum or difference between different time points, generating a symmetric matrix representing time correlation. The Markov transition field transformation is based on a first-order Markov chain, which calculates the transition probabilities between states in the one-dimensional feature vector to generate a two-dimensional matrix representing the state transition mode. The recursive graph transformation generates a symmetric matrix representing the dynamic behavior of a time series by calculating the Euclidean distance between points in the one-dimensional feature vector and determining whether state recursion has occurred based on a preset threshold.

6. The method according to claim 1, characterized in that, Step S5 uses the trained and improved Swin-UNet model to classify mountain pasture features from the 2D image input data to obtain a high-precision pasture feature classification map, including the following sub-steps: S51, For the pasture image dataset during the lush pasture season, the pasture feature classification map output by the improved Swin-UNet model contains six types of land features: grassland, shrubs, river-lake, cloud, bare land, and shadow. S52, For the dry season image dataset, the pasture feature classification map output by the improved Swin-UNet model contains five types of land features: grassland, shrubs, river-lake-cloud-snow, bare land, and shadow. After the pasture element classification map is generated, it is post-processed using morphological operations or Conditional Random Field (CRF) to eliminate the salt and pepper effect.

7. The method according to claim 1, characterized in that, Step S6, based on the pasture element classification map, performs a dynamic assessment of the grass-livestock balance to determine the grass-livestock balance status level of the mountain pasture, including the following sub-steps: S61, Calculate the area of ​​grassland elements based on the pasture element classification map; S62, Based on the grassland area and field survey data, calculate the theoretical carrying capacity R and the actual carrying capacity A of the mountain pasture; S63, Calculate the grass-livestock balance index BGLI of the mountain pasture based on the theoretical carrying capacity R and the actual carrying capacity A; S64, Determine the grass-livestock balance status level of the mountain pasture based on the value of the grass-livestock balance index (BGLI); The theoretical carrying capacity R is calculated as follows: R=(Y*E) / (I*D), where R is the reasonable carrying capacity of the grassland during the grazing period, Y is the total grass yield of the grassland, E is the grassland utilization rate, I is the daily feed intake of sheep, and D is the grazing duration. The formula for calculating the grass-livestock balance index (BGLI) is: BGLI = (AR) / R; The grass-livestock balance status is divided into five levels: Extreme overloading, when the BGLI is greater than 150%; severe overloading, when the BGLI is greater than 80% and less than or equal to 150%; overloading, when the BGLI is greater than 20% and less than or equal to 80%; grass-livestock balance, when the BGLI is greater than -20% and less than or equal to 20%; underloading, when the BGLI is less than or equal to -20%.

8. A system for extracting elements from mountain pastures, comprising: The remote sensing image preprocessing module is used to acquire and preprocess remote sensing image data to generate a multi-temporal, high-precision remote sensing image dataset with uniform radiometric and geometric characteristics. The spectral-spatial feature extraction module is used to extract spectral features, topographic features, texture features, and regional aggregation features from the remote sensing image dataset output by the remote sensing image preprocessing module, so as to construct a comprehensive spectral-spatial joint feature dataset. The feature 2D conversion module is used to convert the spectral-spatial joint feature dataset output by the spectral-spatial feature extraction module from a one-dimensional sequence to a two-dimensional image, generating two-dimensional image input data with multi-channel features; An improved Swin-UNet model is used to receive two-dimensional image input data output by the feature 2D transformation module and output a mountain pasture feature classification map. The improved Swin-UNet model includes an encoder, a bottleneck module, a decoder and several skip connections. The encoder integrates a local perception enhancement module and the decoder integrates a cross-layer feature fusion module. The grass-livestock balance dynamic assessment module is used to perform a grass-livestock balance dynamic assessment based on the pasture element classification map output by the improved Swin-UNet model, so as to determine the grass-livestock balance status level of the mountain pasture. The mountain pasture element extraction system is used to perform a mountain pasture element extraction method as described in any one of claims 1-7.

Images