A Deep Learning-Based Dynamic Monitoring Method for Urban Renewal Blue-Green Spaces

Abstract

This invention belongs to the field of data processing and discloses a deep learning-based method for dynamic monitoring of blue-green space in urban renewal. The method includes: acquiring high-resolution remote sensing images and preprocessing them to obtain preprocessed image data; determining a UNetMamba model and improving it using a ResTv2 backbone network and ECA attention mechanism to obtain an improved UNetMamba model; determining a preset dataset and training and optimizing the improved UNetMamba model based on the preset dataset to obtain a blue-green space classification model; inputting the preprocessed image data into the blue-green space classification model and outputting the blue-green space classification results; and performing post-processing and spatial transfer econometric analysis on the blue-green space classification results to obtain a blue-green space distribution map and a blue-green space transformation Sankey diagram. This invention can effectively improve the identification accuracy of blue-green space, thereby effectively revealing the evolutionary patterns of ecological space.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method for dynamic monitoring of urban renewal blue-green spaces based on deep learning. Background Technology

[0002] Against the backdrop of rapid global urbanization driving economic and social development, Blue-Green Space (BGS) encompasses water bodies (blue spaces, such as rivers and lakes) and vegetation-covered areas (green spaces, such as parks, woodlands, and wetlands). With its core ecological service functions such as regulating microclimate, conserving water resources, maintaining biodiversity, and enhancing urban resilience, it has become a core element of high-density urban development.

[0003] Traditional research methods for identifying blue-green spaces mainly rely on multi-band spectral indices and threshold segmentation, or pixel-level classification based on machine learning. However, these methods have significant limitations when facing complex urban renewal scenarios, resulting in insufficient accuracy in identifying blue-green spaces and failing to meet the quantitative analysis needs of ecological space evolution patterns during urban renewal.

[0004] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0005] The main objective of this invention is to provide a deep learning-based method for dynamic monitoring of blue-green spaces in urban renewal, aiming to solve the problem that the existing technology has insufficient accuracy in identifying blue-green spaces when facing complex urban renewal scenarios, and cannot meet the needs of quantitative analysis of the evolution law of ecological space in the process of urban renewal.

[0006] To achieve the above objectives, this invention provides a deep learning-based method for dynamic monitoring of urban renewal blue-green spaces, comprising the following steps:

[0007] Acquire high-resolution remote sensing images and preprocess the high-resolution remote sensing images to obtain preprocessed image data;

[0008] The UNetMamba model was determined, and the UNetMamba model was improved by using the ResTv2 backbone network and ECA attention mechanism to obtain the improved UNetMamba model.

[0009] A preset dataset is determined, and the improved UNetMamba model is trained and optimized based on the preset dataset to obtain a blue-green space classification model;

[0010] The preprocessed image data is input into the blue-green space classification model, and the blue-green space classification result is output.

[0011] The blue-green space classification results are post-processed and subjected to spatial transfer econometric analysis to obtain a blue-green space distribution map and a blue-green space transformation Sankey diagram.

[0012] Optionally, the deep learning-based method for dynamic monitoring of urban renewal blue-green space, wherein acquiring high-resolution remote sensing images and preprocessing the high-resolution remote sensing images to obtain preprocessed image data specifically includes:

[0013] Identify the target area and acquire high-resolution remote sensing images of the target area;

[0014] The high-resolution remote sensing image is subjected to radiometric calibration, atmospheric correction, image registration, and cropping to obtain preprocessed image data.

[0015] Optionally, the deep learning-based method for dynamic monitoring of urban renewal blue-green space, wherein determining the UNetMamba model and improving it using a ResTv2 backbone network and ECA attention mechanism to obtain an improved UNetMamba model specifically includes:

[0016] The UNetMamba model was determined, and the encoder backbone network in the UNetMamba model was replaced with the ResTv2 backbone network to obtain the improved UNetMamba model.

[0017] The ResTv2 backbone network includes multiple sequentially connected feature stages, each of which includes multiple cascaded ResTv2 modules and an ECA attention mechanism module.

[0018] Optionally, the deep learning-based method for dynamic monitoring of urban renewal blue-green space, wherein determining a preset dataset and training and optimizing the improved UNetMamba model based on the preset dataset to obtain a blue-green space classification model specifically includes:

[0019] A preset dataset is determined, and the preset dataset is reclassified and partitioned to obtain a training set and a validation set.

[0020] The improved UNetMamba model is trained using the training set to obtain an initial blue-green space classification model.

[0021] The initial blue-green space classification model is optimized using the Bayesian optimization algorithm based on the validation set to obtain the blue-green space classification model.

[0022] Optionally, the deep learning-based urban renewal blue-green space dynamic monitoring method further includes, before the step of training the improved UNetMamba model based on the training set to obtain an initial blue-green space classification model:

[0023] The training set is subjected to data augmentation processing to obtain an augmented training dataset, wherein the data augmentation processing includes geometric transformation processing, color perturbation processing, and composite augmentation processing.

[0024] Optionally, the deep learning-based method for dynamic monitoring of urban renewal blue-green space, wherein training the improved UNetMamba model using the training set to obtain an initial blue-green space classification model specifically includes:

[0025] The enhanced training dataset is input into the improved UNetMamba model, and the enhanced training dataset is processed by feature extraction and global average pooling through the ResTv2 backbone network in the improved UNetMamba model to obtain the input feature sequence and learnable weight matrix.

[0026] An attention weight matrix is ​​constructed based on the input feature sequence and the learnable weight matrix, and the attention weight matrix is ​​subjected to weighted summation and pixel recombination to obtain a multi-scale feature pyramid.

[0027] The multi-scale feature pyramid is input into the decoder network of the improved UNetMamba model. The decoder network performs spatial resolution restoration and fusion processing on the multi-scale feature pyramid, and outputs a blue-green space semantic segmentation map. The improved UNetMamba model is then trained, and an initial blue-green space classification model is obtained.

[0028] Optionally, the deep learning-based method for dynamic monitoring of urban renewal blue-green space, wherein the step of performing feature extraction and global average pooling on the enhanced training dataset through the ResTv2 backbone network in the improved UNetMamba model to obtain the input feature sequence and learnable weight matrix specifically includes:

[0029] The augmented training dataset is input into the initial feature extraction layer in the ResTv2 module, and the feature extraction processing of the augmented training dataset is performed through the initial feature extraction layer to obtain the input feature sequence.

[0030] The input feature sequence is input into the ECA attention mechanism module, which performs global average pooling and local cross-channel interaction information calculation on the input feature sequence to obtain a learnable weight matrix.

[0031] Optionally, the deep learning-based method for dynamic monitoring of urban renewal blue-green space, wherein constructing an attention weight matrix based on the input feature sequence and the learnable weight matrix, and performing weighted summation and pixel recombination on the attention weight matrix to obtain a multi-scale feature pyramid, specifically includes:

[0032] Perform matrix multiplication on the input feature sequence and the learnable weight matrix to obtain the Q vector;

[0033] The input feature sequence is arranged into a two-dimensional feature map, and the two-dimensional feature map is subjected to depthwise separable convolution to obtain a compressed feature map;

[0034] The compressed feature map is flattened and projected to obtain the K vector and V vector;

[0035] The attention weight matrix is ​​obtained based on the Q vector and the K vector, and the attention weight matrix is ​​weighted and summed with the V vector to obtain the attention output sequence.

[0036] The attention output sequence is subjected to pixel recombination processing to obtain a multi-scale feature pyramid.

[0037] Optionally, the deep learning-based method for dynamic monitoring of urban renewal blue-green space, wherein inputting the preprocessed image data into the blue-green space classification model and outputting the blue-green space classification result specifically includes:

[0038] The preprocessed image data is subjected to regular slicing to obtain sliced ​​image data;

[0039] The sliced ​​image data is input into the blue-green space classification model, and the blue-green space prediction processing of the sliced ​​image data is performed by the blue-green space classification model to obtain the blue-green space classification result.

[0040] The blue-green space classification results include blue space, green space, and construction land.

[0041] Optionally, the deep learning-based urban renewal blue-green space dynamic monitoring method, wherein the post-processing and spatial transfer econometric analysis of the blue-green space classification results to obtain a blue-green space distribution map and a blue-green space transformation Sankey diagram specifically includes:

[0042] The original spatial location of the high-resolution remote sensing image is obtained, and the blue-green spatial classification results are stitched together based on the original spatial location to obtain a blue-green spatial distribution map.

[0043] Spatial transfer econometric analysis was performed on the blue-green space classification results to obtain the blue-green space transformation Sankey diagram.

[0044] In this invention, high-resolution remote sensing images are acquired and preprocessed to obtain preprocessed image data; a UNetMamba model is determined, and the UNetMamba model is improved using a ResTv2 backbone network and an ECA attention mechanism to obtain an improved UNetMamba model; a preset dataset is determined, and the improved UNetMamba model is trained and optimized based on the preset dataset to obtain a blue-green space classification model; the preprocessed image data is input into the blue-green space classification model, and the blue-green space classification result is output; the blue-green space classification result is post-processed and subjected to spatial transfer econometric analysis to obtain a blue-green space distribution map and a blue-green space transformation Sankey diagram. This invention improves the UNetMamba model by introducing the ResTv2 backbone network and ECA attention mechanism. The improved UNetMamba model is then trained and optimized to obtain a blue-green space classification model. This model can accurately identify blue-green spaces in preprocessed high-resolution remote sensing images and construct blue-green space distribution maps and blue-green space transformation Sankey diagrams. It can accurately measure the scale and direction of mutual transformation between blue-green spaces and construction land during urban renewal, thereby effectively revealing the laws of ecological space evolution. Attached Figure Description

[0045] Figure 1 This is a flowchart of a preferred embodiment of the deep learning-based dynamic monitoring method for urban renewal blue-green space of the present invention;

[0046] Figure 2 This is a schematic diagram of the overall implementation process of a preferred embodiment of the deep learning-based dynamic monitoring method for urban renewal blue-green space of the present invention.

[0047] Figure 3 This is a schematic diagram of the improved UNetMamba model structure of a preferred embodiment of the deep learning-based dynamic monitoring method for urban renewal blue-green space of the present invention.

[0048] Figure 4 This is a schematic diagram of the Sankey diagram of blue-green space transformation in a certain city from 2014 to 2017 to 2020, representing a preferred embodiment of the deep learning-based dynamic monitoring method for urban renewal blue-green space in this invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0050] Against the backdrop of rapid global urbanization driving economic and social development, blue-green space (BGS), encompassing water bodies (blue spaces, such as rivers and lakes) and vegetation-covered areas (green spaces, such as parks, woodlands, and wetlands), has become a core element of high-density urban development due to its core ecological service functions, including regulating microclimates, conserving water resources, maintaining biodiversity, and enhancing urban resilience. Currently, many large cities have gradually shifted from large-scale outward expansion to urban renewal focused on optimizing existing resources. In this transition, the changing patterns of urban blue-green space are becoming increasingly complex. On the one hand, urban renewal may increase blue-green space through ecological restoration and park construction; on the other hand, high-intensity redevelopment and infrastructure upgrades may also lead to the encroachment, fragmentation, or even degradation of blue-green space. These dynamic changes directly relate to the ecological benefits and development level of urban renewal.

[0051] Therefore, research on the evolution of blue-green space patterns requires not only accurate identification of their spatial distribution and morphology, but also precise measurement of their mutual transfer relationships with built-up land, quantifying the loss and restoration effects of ecological space. Traditional research methods mainly rely on multi-band spectral indices (such as NDVI and NDWI) and threshold segmentation, or pixel-level classification based on machine learning (such as SVM, DT, and RF). These methods are mature and effective in large-scale homogeneous landscape mapping, but they reveal significant limitations when facing the complex scenarios of urban renewal. On the one hand, the limited bands of high-resolution images restrict improvements to spectral methods; on the other hand, traditional methods rely on expert-designed features, which have high professional thresholds and poor generalization, making it difficult to fully extract high-level semantic information, resulting in insufficient accuracy in identifying small-scale blue-green spaces. More importantly, existing methods mostly stop at static identification, lacking the ability to accurately measure the dynamic transfer process of "blue-green space-built-up land," and cannot meet the quantitative analysis needs of ecological space evolution patterns during urban renewal.

[0052] In recent years, the application of deep learning technology, especially computer vision, in the field of semantic segmentation of remote sensing images has provided a new technical path for improving the accuracy of blue-green spatial recognition. Deep learning models, represented by convolutional neural networks (CNNs), can automatically learn multi-level feature representations from raw images, significantly reducing the reliance on manual feature design. However, traditional CNN models are limited by local receptive fields and repeated downsampling operations, resulting in weak global contextual information capture capabilities and difficulty in effectively modeling semantic relationships between large-scale ground features. Although semantic segmentation models such as U-Net and DeepLab perform well in general scenarios, they still suffer from problems such as loss of spatial details and insufficient sensitivity to small target recognition when processing high-resolution urban remote sensing images, especially in supporting accurate measurement of subsequent spatial transfers.

[0053] Therefore, there is an urgent need to develop a deep learning dynamic monitoring method that can balance global perception and detail preservation while having high computational efficiency. This method should not only achieve high-precision dynamic monitoring of urban renewal blue-green spaces, but also accurately measure the transfer relationship between blue-green spaces and construction land, thereby providing reliable technical support for assessing the ecological benefits of urban renewal processes and optimizing land space protection patterns.

[0054] To address the shortcomings of existing technologies in blue-green space identification and dynamic analysis, this invention provides a deep learning-based method for dynamic monitoring of urban renewal blue-green space. This method aims to solve core problems such as the difficulty in identifying small targets and the significant loss of detail in high-resolution urban imagery, as well as the inability of traditional methods to effectively quantify the dynamic transformation relationship between blue-green space and built-up land. Using 2-meter resolution remote sensing images from the domestic Gaofen-1 (GF-1) satellite as the primary data source, and relying on the open-source WHDLD dataset (Wuhan dense labeling dataset, a densely labeled dataset used in this invention) for model training and validation, this method achieves fully automated processing from raw remote sensing image input to blue-green space distribution mapping and transformation pattern analysis through a series of steps including image preprocessing, improved semantic segmentation model construction, model training and hyperparameter optimization, inference and output, and spatial transfer quantitative analysis. The core of this invention is to construct a deep learning framework that integrates multi-scale feature extraction and channel attention mechanisms. It employs a ResTv2 encoder structure with detail preservation capabilities (a novel encoder backbone network designed in this invention) and combines it with a lightweight and efficient channel attention (ECA) module, which significantly improves the model's recognition accuracy and robustness in complex urban environments. Furthermore, it constructs a spatial transfer matrix of blue-green spaces and built-up land across multiple periods, enabling precise quantification of the ecological space loss and restoration process.

[0055] The technical effects of this invention include: 1. Significantly improving the identification accuracy of small-scale blue-green spaces, with the model's average mean Intersection over Union (mIoU) reaching over 70%; 2. Accurately measuring the scale and direction of mutual transformation between blue-green spaces and construction land during urban renewal, revealing the laws governing ecological spatial evolution; 3. The model has a lightweight structure and high computational efficiency, suitable for rapid processing and dynamic analysis of large-scale remote sensing images. This invention not only provides a technical means for high-precision identification of urban blue-green spaces, but more importantly, it provides reliable quantitative evidence and decision support for scientifically evaluating the ecological benefits of urban renewal and optimizing land space planning.

[0056] The preferred embodiment of the present invention describes a deep learning-based method for dynamic monitoring of urban renewal blue-green spaces, such as... Figure 1 As shown, the deep learning-based method for dynamic monitoring of urban renewal blue-green spaces includes the following steps:

[0057] Step S10: Acquire high-resolution remote sensing images and preprocess the high-resolution remote sensing images to obtain preprocessed image data.

[0058] like Figure 2 As shown, this invention comprises five main steps: Step 1: Preprocessing the input high-resolution remote sensing imagery, including radiometric calibration, atmospheric correction, image registration, and cropping. Step 2: Constructing an improved UNetMamba model (a U-shaped model based on a state-space model), enhancing its feature extraction capabilities by introducing a ResTv2 backbone network and ECA attention mechanism. Step 3: Training the improved UNetMamba model using the reclassified WHDLD dataset, combining data augmentation and hyperparameter optimization to improve model performance. Step 4: Generating blue-green space classification results through model inference, and obtaining the final spatial distribution map through post-processing. Step 5: Constructing a spatial transition matrix based on multi-period blue-green space distribution results to precisely measure the transformation relationship between blue-green space and built-up land. This invention achieves fully automated processing from raw remote sensing imagery to dynamic change measurement and analysis of blue-green space.

[0059] This invention takes a certain city as the research area and analyzes the distribution and changes of blue-green space in the urban renewal stage. The invention sequentially identifies and performs quantitative analysis on the blue-green space in the research area during three periods: 2014, 2017, and 2020, thereby verifying the applicability and stability of the invention in different periods.

[0060] Specifically, the target area is determined, and a high-resolution remote sensing image of the target area is acquired; the high-resolution remote sensing image is then subjected to radiometric calibration, atmospheric correction, image registration, and cropping to obtain preprocessed image data.

[0061] The preprocessing process for high-resolution remote sensing images is as follows: This invention selects 2-meter resolution multispectral remote sensing images collected by the domestic Gaofen-1 (GF-1) satellite in 2014, 2017 and 2020, covering the entire city, as the basic data source. The raw data for each period consists of multiple images. All images have undergone strict quality screening and control to ensure that the cloud cover rate of each image is less than 5%, so as to minimize the impact of cloud cover on the accuracy of ground feature identification.

[0062] This invention includes an image preprocessing module responsible for standardizing and stitching together multiple acquired raw remote sensing images (i.e., high-resolution remote sensing images in this invention) to generate standardized data products covering the entire city and meeting the input requirements of deep learning models. The specific processing flow includes the following four core steps:

[0063] 1. Radiometric Calibration: The digital quantization values ​​of each image are converted into surface reflectance to eliminate differences in the radiometric response of the sensors themselves. This is achieved by using calibration coefficients provided by the sensors and performing radiometric correction through linear transformation, thereby obtaining surface reflectance data with consistent physical meaning.

[0064] 2. Atmospheric Correction: The FLAASH (Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes, an atmospheric correction tool) model is used to perform atmospheric correction on each image after radiometric calibration to eliminate the influence of atmospheric scattering and absorption on surface reflectivity. By inputting parameters such as imaging time, geographical location, and aerosol type, atmospheric conditions are accurately estimated and correction is completed.

[0065] 3. Image Registration and Stitching: First, a registration method based on the SIFT (Scale Invariant Feature Transform) feature detector was adopted. A scene from 2020 was selected as the baseline. By extracting feature points and using a random sampling consensus algorithm to remove mismatches, sub-pixel-level accurate registration was achieved between all images from 2014 and 2017 and the baseline image, ensuring the consistency of spatial geometric positions among multiple images. The registration error was less than 0.5 pixels. Subsequently, the registered images from the same period were seamlessly stitched together to generate a complete image covering the entire study area.

[0066] 4. Image cropping: Based on the administrative division vector boundaries of a city, the city-wide images stitched together in three periods were precisely cropped to obtain 2-meter resolution remote sensing image data covering the entire city in 2014, 2017 and 2020 with a unified spatial reference.

[0067] Step S20: Determine the UNetMamba model, and improve the UNetMamba model using the ResTv2 backbone network and ECA attention mechanism to obtain the improved UNetMamba model.

[0068] The improved UNetMamba model described in this invention inherits the encoder-decoder framework and skip connection structure of the classic UNet architecture to ensure the effective fusion and transmission of multi-scale feature information. Addressing the core challenges of small target recognition and the difficulty in balancing global contextual information capture and detail preservation in blue-green space identification of high-resolution remote sensing images, this invention makes key improvements to the encoder backbone network and feature enhancement mechanism of the original UNetMamba model, aiming to comprehensively improve the model's perceptual accuracy and robustness.

[0069] Specifically, a UNetMamba model is determined, and the encoder backbone network in the UNetMamba model is replaced by a ResTv2 backbone network to obtain an improved UNetMamba model; wherein, the ResTv2 backbone network includes multiple sequentially connected feature stages, and each feature stage includes multiple cascaded ResTv2 modules and an ECA attention mechanism module.

[0070] The improved model structure is as follows Figure 3 As shown, the original ResT structure (the original encoder backbone network in the UNetMamba model) is replaced by ResTv2.

[0071] One of the core innovations of this invention lies in the significant optimization of the encoder backbone network. Although the ResT structure used in the original UNetMamba model can improve model performance, it still has obvious shortcomings when processing high-resolution remote sensing images. The shortcomings are reflected in the fact that its multi-head interaction module leads to high computational complexity and limited inference speed. In addition, the original UNetMamba model needs to be strengthened in its ability to preserve image detail information during downsampling.

[0072] To address the aforementioned issues, this invention introduces ResTv2 as a novel encoder backbone network to replace the original ResT structure. This encoder undertakes the crucial task of layer-by-layer feature extraction and abstraction from the input remote sensing image. Specifically, the input tensor (i.e., the training set) first enters the first stage of the ResTv2 backbone network. This network comprises four sequentially connected feature stages. Each stage progressively downsamples the input feature map and increases the number of channels, ultimately outputting a multi-scale feature pyramid, thereby effectively capturing semantic information at different levels.

[0073] Furthermore, each feature stage in the ResTv2 backbone network contains multiple cascaded ResTv2 modules, the core of which is an efficient multi-head self-attention mechanism (i.e., EMSA v2).

[0074] Step S30: Determine the preset dataset, and train and optimize the improved UNetMamba model based on the preset dataset to obtain the blue-green space classification model.

[0075] Specifically, a preset dataset is determined, and the preset dataset is subjected to reclassification and data partitioning to obtain a training set and a validation set; the training set is subjected to data augmentation to obtain an augmented training dataset, wherein the data augmentation includes geometric transformation, color perturbation, and composite augmentation.

[0076] For the training and optimization process of the model, this invention uses the reclassified WHDLD dataset to train and optimize the improved UNetMamba model, specifically including the following steps:

[0077] 1. Training Data Preparation: This invention uses the open-source WHDLD dataset as the basic data for model training and validation. To align with the blue-green space recognition task, the original annotation system of the dataset is reclassified. The reclassification process includes: classifying "water bodies" as blue space, "vegetation" as green space, and merging "buildings," "roads," "sidewalks," and "bare soil" into construction land, thereby constructing a three-category semantic segmentation system for this invention. Furthermore, to ensure the objectivity and statistical validity of the model evaluation, the reclassified dataset is randomly divided into a training set (3705 samples in total) and a validation set (1235 samples in total) at a ratio of approximately 3:1.

[0078] 2. Data Augmentation: To improve the model's generalization ability and robustness in the face of complex and varied remote sensing imagery scenes, data augmentation strategies were implemented during training. These strategies mainly include the following three categories: Geometric Transformation: including random horizontal flipping, vertical flipping, and rotation within a range of ±45°; Color Perturbation: randomly adjusting brightness and contrast in the RGB space (adjustment range ±30%), and randomly perturbing hue, saturation, and brightness in the HSV space; Composite Augmentation: using Mosaic augmentation technology, randomly cropping and stitching together four training images to simulate complex scenes and enhance the model's contextual awareness of multi-scale targets.

[0079] The enhanced training dataset is input into the improved UNetMamba model, and the enhanced training dataset is input into the initial feature extraction layer in the ResTv2 module. The enhanced training dataset is processed by the initial feature extraction layer to obtain the input feature sequence.

[0080] The input feature sequence is input into the ECA attention mechanism module, which performs global average pooling and local cross-channel interaction information calculation on the input feature sequence to obtain a learnable weight matrix.

[0081] The ECA attention mechanism embedding process is as follows: To enhance the model's ability to focus on key feature channels in complex scenes, this invention embeds an Efficient Channel Attention (ECA) module at the beginning of each stage of the encoder path. For example... Figure 3 As shown, for Stage 1 (first level), the ECA module is embedded after the initial feature extraction layer (Stem) to recalibrate the channel weights of the output primary feature map. For Stages 2 (second level) to 4 (fourth level), after receiving the downsampled feature map, each Stage processes it through the ECA module before inputting it into the subsequent feature extraction units within that Stage. This embedding strategy enables the model to autonomously emphasize information-rich channels and suppress redundant or noisy channels before entering each feature extraction stage, thereby guiding the network to focus on the features most relevant to blue-green space recognition.

[0082] The ECA module operates as follows: 1. First, the input feature map is globally averaged in the spatial dimension to obtain the channel description vector; 2. Then, local cross-channel interaction information is captured through one-dimensional convolution (the kernel size k is adaptively determined by the number of channels C); 3. Finally, channel weights are generated through the Sigmoid activation function, and the generated weight vector (B, C, 1, 1) is fused with the original input feature map (B, C, H, W) through channel-by-channel multiplication. That is, the feature map of each channel is multiplied by its corresponding weight value (where B is Batch Size, which means the number of data samples input into the model for parallel computation at one time; C is Channels, which means the dimension of the feature map. For the input image, it is usually the color channel; for the intermediate feature map, it represents the number of different types of features; H is Height, which means the number of pixels or feature points in the vertical direction of the feature map; W is Width, which means the number of pixels or feature points in the horizontal direction of the feature map). This invention significantly enhances the model's ability to perceive and select channel-dimensional features by embedding a lightweight ECA module beforehand, without significantly increasing the model's computational complexity. This effectively improves the model's sensitivity to small blue-green space targets and its robustness in complex contexts.

[0083] Matrix multiplication is performed on the input feature sequence and the learnable weight matrix to obtain the Q vector; the input feature sequence is arranged into a two-dimensional feature map, and the two-dimensional feature map is subjected to depthwise separable convolution to obtain a compressed feature map; the compressed feature map is flattened and projected to obtain the K vector and the V vector; an attention weight matrix is ​​obtained based on the Q vector and the K vector, and the attention weight matrix and the V vector are weighted and summed to obtain an attention output sequence; the attention output sequence is pixel reorganized to obtain a multi-scale feature pyramid.

[0084] The specific processing procedure of the ResTv2 backbone network is as follows:

[0085] 1. First, perform matrix multiplication on the input feature sequence and the learnable weight matrix to generate the Query(Q) vector, and then rearrange the input feature sequence into a two-dimensional feature map with height H and width W.

[0086] 2. Perform depthwise separable convolution on the two-dimensional feature map. First, use depthwise convolution to perform spatial convolution on each input channel to capture spatial features. Then, use pointwise convolution (1×1 convolution) to integrate the output of each channel to achieve spatial downsampling and information compression.

[0087] 3. After flattening the compressed feature map (i.e. the compressed feature map in this invention) back into a sequence, project it to obtain the Key (K) and Value (V) vectors.

[0088] 4. Calculate the dot product of the transpose of Q and K, divide by the scaling factor (which is generally the square root of the dimension of the Key vector), normalize by the Softmax function to obtain the attention weight matrix, and then sum it with the V vector to obtain the attention output sequence.

[0089] 5. Finally, by performing pixel recombination on the attention output sequence, the low-resolution, high-channel-count feature map (B, C×r) is obtained. 2 The features (B, C, H×r, W×r) are rearranged into a high-resolution, low-channel-number feature map (B, C, H×r, W×r), where r is the upsampling factor, in order to recover detailed information.

[0090] 6. The attention output sequence is fused with the original input features by element-wise addition using a shortcut connection to complete the forward computation of this module.

[0091] By leveraging the hourglass structure formed by the aforementioned "convolutional downsampling-pixel recombinant upsampling", the ResTv2 encoder can not only efficiently capture global contextual information in remote sensing images, but also significantly improve the ability to preserve local details of ground features, laying a key technical foundation for the subsequent accurate identification of small blue-green spaces (such as pocket parks and ecological ponds).

[0092] The multi-scale feature pyramid is input into the decoder network of the improved UNetMamba model. The decoder network performs spatial resolution restoration and fusion processing on the multi-scale feature pyramid, and outputs a blue-green space semantic segmentation map. The improved UNetMamba model is then trained, and an initial blue-green space classification model is obtained.

[0093] The decoder path receives the deepest features from the encoder output (i.e., the multi-scale feature pyramid in this invention) and gradually restores the spatial resolution through a series of upsampling convolutional layers. After each upsampling, the decoder receives feature maps from the corresponding stage of the encoder path through skip connections and fuses them to supplement the details that may have been lost during the upsampling process, ultimately reconstructing a high-resolution, high-precision blue-green spatial semantic segmentation map.

[0094] In summary, the improved UNetMamba model constructed in this invention is not a simple patch to the original UNetMamba model, but rather a novel, synergistic architecture formed through two core improvements: the introduction of a ResTv2 encoder and the embedding of a multi-layered ECA attention mechanism. This architecture fully leverages the advantages of ResTv2 in global modeling and detail preservation, while utilizing the ECA module to achieve adaptive optimization of feature channels. Ultimately, this significantly improves the model's accuracy and robustness in blue-green space recognition in complex urban environments.

[0095] The initial blue-green space classification model is optimized using the Bayesian optimization algorithm based on the validation set to obtain the blue-green space classification model.

[0096] The specific process of hyperparameter optimization is as follows: The core evaluation metric for model performance is set as the mean Intersection over Union (mIoU) on the validation set. To obtain the optimal model performance, this invention adopts the Optuna automated hyperparameter search framework based on the Bayesian optimization algorithm. This framework models the black-box functional relationship between the validation set mIoU and hyperparameters by constructing a probabilistic surrogate model (usually a Gaussian process model), and uses a sampling function (such as the desired improvement in EI) to balance exploration and utilization, thereby guiding the efficient and targeted search of hyperparameter combinations. The hyperparameter optimization space includes the learning rate (lr), weight decay coefficient, training batch size, and validation batch size. First, through 30 rounds of prior exploration experiments (i.e., randomly sampling initial hyperparameter combinations within a wide range of values), the effective value range of each hyperparameter is initially determined, excluding regions with significantly low performance, providing a high-quality starting point for subsequent fine-grained searches. Based on this, 200 rounds of iterative fine-grained searches are conducted. In each iteration, Optuna updates the surrogate model based on the existing experimental results and recommends the next most promising hyperparameter combination by maximizing the acquisition function. This combination is then used for model training and validation, obtaining its mIoU value and feeding it back to the optimizer to gradually approach the global optimum. Ultimately, the optimal hyperparameter combination that maximizes the mIoU on the validation set was determined: learning rate lr = 3.3 × 10⁻⁶. -5 Weight decay wd = 2.11 × 10 -4 The training batch size was 32, and the validation batch size was 16. Based on this parameter combination, the model achieved an mIoU of 84.25% on the validation set, which is about 3.7 percentage points higher than the baseline model. This optimal model (i.e., the blue-green space classification model in this invention) was saved and used for subsequent inference and blue-green space recognition applications.

[0097] Step S40: Input the preprocessed image data into the blue-green space classification model and output the blue-green space classification result.

[0098] The specific processing flow of the improved and trained blue-green space classification model is as follows: The preprocessed remote sensing image (i.e., the preprocessed image data in this invention) is input into the model. First, the improved encoder performs multi-level feature extraction. Before performing deep feature transformation at each stage of the encoder path, the input features are recalibrated in the channel dimension using the ECA module to enhance key features. Subsequently, the ResTv2 module captures long-range contextual dependencies using its efficient multi-head self-attention mechanism and preserves detailed information through its unique hourglass structure. The encoder ultimately outputs a set of multi-scale feature pyramids with rich global semantics and local details. Afterward, the decoder in the blue-green space classification model processes the data and outputs the blue-green space classification result.

[0099] Specifically, the preprocessed image data is subjected to regular slicing to obtain sliced ​​image data; the sliced ​​image data is input into the blue-green space classification model, and the blue-green space classification model is used to perform blue-green space prediction processing on the sliced ​​image data to obtain blue-green space classification results; wherein, the blue-green space classification results include blue space, green space and construction land.

[0100] The specific process for outputting the blue-green space classification results is as follows: This invention applies the optimal hyperparameter model obtained from training (i.e., the blue-green space classification model in this invention) to high-resolution remote sensing image data of a certain city in 2014, 2017, and 2020, to achieve automated identification and product generation of blue-green spaces in multiple periods. Specifically, it includes the following steps:

[0101] 1. Input image slices: The preprocessed remote sensing images of the entire city in three periods are sliced ​​into regular slices of 256×256 pixels each, resulting in 9447 valid image slices as model input. The slice size is consistent with the sample size used in the model training stage to ensure that the input data format matches the model structure.

[0102] 2. Model Inference: Input the image slices from the three periods batch by batch with the loaded optimal parameters (i.e., learning rate lr = 3.3 × 10⁻⁶). -5 Weight decay wd = 2.11 × 10 -4 Forward inference is performed in the blue-green space classification model, and the model output is a pixel-level semantic segmentation result (i.e., the blue-green space classification result in this invention). Each pixel is assigned a category label, corresponding to blue space (water body), green space (vegetation) or construction land respectively.

[0103] Step S50: Post-process and perform spatial transfer econometric analysis on the blue-green space classification results to obtain a blue-green space distribution map and a blue-green space transformation Sankey diagram.

[0104] Specifically, the original spatial location of the high-resolution remote sensing image is obtained, and the blue-green spatial classification results are stitched together based on the original spatial location to obtain a blue-green spatial distribution map; spatial transfer measurement analysis is performed on the blue-green spatial classification results to obtain a blue-green spatial transformation Sankey diagram.

[0105] Semantic recognition slices from 2014, 2017, and 2020 were stitched together based on their original spatial locations to reconstruct a blue-green spatial distribution map covering the entire city. During the stitching process, based on the geographic coordinates of each pixel, the slices were precisely aligned to their corresponding geographic regions, and unified values ​​were assigned to the three categories of labels: blue space, green space, and construction land, resulting in a seamless city-wide classification map. This achievement preserves complete geographic coordinate information and can be directly used for spatial analysis, mapping output, and multi-temporal variation studies.

[0106] The accuracy verification process for the blue-green space classification results is as follows: To objectively evaluate the reliability of the identification results across multiple periods, this invention employs stratified random sampling to statistically verify the identification results for each period. From the 9477 slices in each period, 845 samples located at administrative boundaries that might lead to incomplete identification are removed, leaving 8602 valid units in each period as the verification population. Based on the Cochran sample size calculation formula, at a 95% confidence level (Z=1.96) and a 5% tolerance, the minimum sample size is determined to be 400 (sampling ratio of 4.65%) to ensure the representativeness of the estimation results.

[0107] The validation samples were manually annotated by a professional team to form ground truth reference data, and quantitative analysis was performed using three widely used semantic segmentation evaluation metrics. The validation results are shown in Table 1 below:

[0108] The semantic segmentation evaluation metrics include: 1. Pixel Accuracy (PA): the proportion of pixels that are correctly classified overall; 2. Mean Class Accuracy (MCA): the arithmetic mean of the accuracy of each class, used to evaluate the model's classification balance; 3. Mean Intersection over Union (mIoU): the mean of the intersection over union ratio between the predicted result and the ground truth, which comprehensively reflects the consistency between the class and the boundary.

[0109] Table 1: Verification Results of Blue-Green Space Recognition Accuracy (%)

[0110]

[0111] The results show that the method proposed in this invention maintains high recognition accuracy and stability at different times, and all three indicators perform well, verifying that the method has strong generalization ability and practicality.

[0112] Furthermore, this invention employs spatial transfer econometric analysis: to accurately quantify the dynamic transformation relationship between blue-green spaces and construction land during urban renewal, and to identify hotspots of spatial change, this invention systematically conducts the following analytical steps based on multi-period classification results:

[0113] 1. Construction of the transfer matrix: Based on the previous classification results, blue-green space and corresponding construction land transfer matrices were constructed for two periods: 2014-2017 and 2017-2020. This matrix was achieved through cross-classification analysis, systematically statistically analyzing the area and direction of mutual transformation between different spaces, providing fundamental data support for quantifying the spatial transformation patterns in the urban renewal process.

[0114] 2. Spatial change coding: In order to standardize the change detection process and realize pixel-by-pixel change recognition, the digital coding rules of each spatial category are first clearly defined: construction land corresponds to category coding (1), blue space corresponds to category coding (2), and green space corresponds to category coding (3).

[0115] Based on this coding system, pixel-by-pixel change detection and recoding are performed on the classification results of the two periods: the category labels of the same pixel in the spatial distribution maps of the two periods are combined to form a change type identifier, and a change detection result map is automatically generated. In this map, the value of each pixel is uniformly represented by a two-digit code, with the tens digit representing the land type code of the previous period and the units digit representing the land type code of the next period. For example, the code "21" indicates that the blue space (2) has changed to construction land (1), and the code "32" indicates that the green space (3) has changed to blue space (2). Through this coding and detection method, the system generates a result map containing 9 possible change types, providing structured input for subsequent quantitative analysis and hotspot identification.

[0116] 3. Focusing on Two Types of Changes: Based on the change detection results, the analysis focuses on two key transformation processes: First, the transformation of blue-green space into construction land, that is, extracting pixels with change type codes "21" (blue space → construction land) and "31" (green space → construction land) to identify the specific location and scale of ecological space encroachment; Second, the transformation of construction land into blue-green space, that is, extracting pixels with change type codes "12" (construction land → blue space) and "13" (construction land → green space) to assess the spatial distribution and effectiveness of ecological restoration and construction.

[0117] 4. Visualization and Output: Generate a Sankey diagram of blue-green space transformation in a certain city from 2014 to 2017 to 2020 (e.g., Figure 4As shown in the figure, the flow paths and volume between different spaces at different times are clearly displayed. Simultaneously, based on the change detection results, representative transformation types (such as "green space → construction land") are selected to generate change hotspot distribution maps, intuitively presenting the spatial clustering characteristics and distribution patterns of typical transformation processes. Combining the quantitative statistics of the transfer matrix with the spatial distribution of typical change hotspots, the driving mechanism of blue-green space evolution in urban renewal can be further analyzed, providing multi-dimensional data support for assessing the laws of ecological space evolution.

[0118] In summary, this invention achieves high-precision, automated identification and detailed quantitative analysis of urban blue-green spaces through efficient image preprocessing, improved deep learning model structure, scientific training methods and inference processes, and accurate spatial transfer measurement analysis, providing reliable technical tools and data support for urban ecological planning and management.

[0119] This invention addresses the core technical challenges of identifying small targets, weak global contextual information capture capabilities, and severe detail loss in blue-green space recognition in high-resolution urban remote sensing images. It proposes the following key innovations and protection schemes.

[0120] The key innovations of this invention are as follows:

[0121] 1. Improved Deep Learning Model Architecture: Addressing the challenges of small target identification, weak global context information capture, and severe detail loss in blue-green space recognition of high-resolution remote sensing imagery, this invention constructs an improved UNetMamba model. While retaining the basic UNet encoder-decoder framework, this model significantly enhances feature extraction capabilities and recognition accuracy by introducing a ResTv2 encoder backbone network and embedding a lightweight, efficient channel attention (ECA) module. The ResTv2 encoder, through its unique "convolutional downsampling-pixel recombinant upsampling" hourglass structure, effectively preserves ground feature details while capturing global context information; while the ECA module, through adaptive channel weight adjustment, enhances the focusing ability on key features and reduces interference from redundant information.

[0122] 2. A Refined Spatial Transfer Measurement and Analysis Method Based on Multi-Period Identification Results: This invention proposes a complete dynamic transfer measurement and analysis process for blue-green spaces. By constructing a spatial transfer matrix and a change coding system, it achieves the entire process from pixel-level change detection to statistical analysis of type transformation relationships. Based on the classification results of two consecutive periods, this method automatically identifies and quantifies the bidirectional transformation between blue-green spaces and built-up land by defining category codes and change type identifiers. It then uses Sankey diagrams and spatial distribution heatmaps for visualization, providing data support for the analysis of ecological spatial evolution mechanisms in urban renewal processes.

[0123] The purpose of this invention is to protect a deep learning-based method for dynamic monitoring of blue-green space in urban renewal. Its core is to construct a deep learning framework that integrates multi-scale feature enhancement and channel attention mechanisms. It aims to solve problems such as the difficulty in identifying small targets in blue-green space in high-resolution urban images, severe loss of details, and the inability of traditional methods to effectively quantify the dynamic transformation relationship between them and construction land.

[0124] This invention presents a deep learning-based method for dynamic monitoring of urban renewal blue-green space, comprising the following complete technical process: 1. First, high-resolution remote sensing images are acquired and preprocessed, including radiometric calibration, atmospheric correction, image registration, stitching, and cropping, to generate standardized input data with a unified spatial reference; 2. Subsequently, an improved UNetMamba semantic segmentation model is constructed. This model is based on an encoder-decoder structure, with the encoder employing a ResTv2 backbone network to achieve multi-scale feature extraction and detail preservation, and embedding an efficient channel attention (ECA) module at the beginning of each encoding stage to enhance key channel features and suppress noise; 3. Subsequently, a deep learning-based method is used for dynamic monitoring of urban renewal blue-green space. The model was trained on the reclassified WHDLD dataset. The reclassification system includes three categories: blue space, green space, and construction land. Data augmentation strategies and hyperparameter optimization methods were introduced during training to improve the model's generalization ability and recognition accuracy. 4. The trained model was used to perform inference on preprocessed images from multiple periods to achieve pixel-level semantic segmentation. After stitching and post-processing, blue-green space distribution maps of different periods were generated. 5. Finally, based on the results of blue-green space distribution from multiple periods, a spatial transition matrix was constructed. The mutual transformation relationship between blue-green space and construction land was identified through change type encoding. The transformation process was quantitatively analyzed and visualized using Sankey diagrams and spatial distribution heatmaps.

[0125] Based on this embodiment, there are also various possible design changes and modifications, all of which fall within the scope of protection of this invention, including but not limited to:

[0126] 1. Model Architecture Replacement and Variation: The encoder backbone network of the improved UNetMamba model can adopt other network structures with strong global context modeling capabilities and detail preservation mechanisms, such as replacing it with advanced visual backbone networks like SwinTransformer, PVTv2 (Pyramid Vision Transformer v2), or EfficientNetv2. The efficient channel attention (ECA) module can be replaced with an SE (Squeeze-and-Excitation) module to enhance channel dependency modeling, or CBAM (Convolutional Block Attention Module) can be used to achieve collaborative optimization of channel and spatial attention. The decoder part of the UNet architecture can be optimized, for example, by introducing structures such as Feature Pyramid Network (FPN) or PANet (Aggregation Network for Instance Segmentation) to enhance multi-scale feature fusion capabilities, or by using sub-pixel convolutions instead of transposed convolutions for upsampling to reduce checkerboard effects.

[0127] 2. Expansion of Data Processing and Enhancement Strategies: The atmospheric correction model in the image preprocessing stage is not limited to FLAASH; other mainstream atmospheric correction models such as 6S and DOS can also be used. The data enhancement strategies can be further expanded. In addition to geometric transformations, color perturbations, and Mosaic, advanced regularization techniques such as CutOut, CutMix, and MixUp can be introduced, or enhancement methods targeting remote sensing image characteristics can be added, such as simulating different cloud and fog coverages and sensor noise. The training data source is not limited to the WHDLD dataset; it can also be compatible with other open-source or self-built remote sensing image semantic segmentation datasets, and supports custom land cover category systems for reclassification based on specific application scenarios.

[0128] 3. Adjustment of Training and Optimization Strategies: The hyperparameter optimization algorithm is not limited to Optuna; it can also be implemented using GridSearch, Random Search, Hyperband, or other Bayesian optimization toolkits (such as Scikit-optimize). The loss function in model training is not limited to standard cross-entropy loss; it can introduce loss functions suitable for semantic segmentation tasks, such as Dice Loss and Focal Loss, or combinations thereof, to better handle class imbalance problems or optimize specific evaluation metrics. The optimizer is not limited to AdamW or SGD; it can also use adaptive optimization algorithms such as Adam, NAdam, and RAdam, combined with various learning rate scheduling strategies (such as Cosine Annealing, Step Decay, OneCycleLR, etc.).

[0129] 4. Expansion of Application Scenarios and Data Sources: The method is not limited to processing GF-1 satellite data, but can also be applied to other high-resolution remote sensing image data from China and abroad, such as image data of different resolutions from GF-2, ZY-3, Sentinel-2, Landsat series, and WorldView. The method is not limited to identifying urban blue-green spaces; after appropriate adjustments and training, it can be extended to other remote sensing image semantic segmentation tasks, such as crop classification, building extraction, road extraction, and land use / cover classification.

[0130] In summary, this invention provides a deep learning-based method for dynamic monitoring of urban renewal blue-green space. The method includes: acquiring high-resolution remote sensing images and preprocessing the high-resolution remote sensing images to obtain preprocessed image data; determining a UNetMamba model and improving the UNetMamba model using a ResTv2 backbone network and an ECA attention mechanism to obtain an improved UNetMamba model; determining a preset dataset and training and optimizing the improved UNetMamba model based on the preset dataset to obtain a blue-green space classification model; inputting the preprocessed image data into the blue-green space classification model and outputting the blue-green space classification result; and performing post-processing and spatial transfer econometric analysis on the blue-green space classification result to obtain a blue-green space distribution map and a blue-green space transformation Sankey diagram. This invention improves the UNetMamba model by introducing the ResTv2 backbone network and ECA attention mechanism. The improved UNetMamba model is then trained and optimized to obtain a blue-green space classification model. This model can accurately identify blue-green spaces in preprocessed high-resolution remote sensing images and construct blue-green space distribution maps and blue-green space transformation Sankey diagrams. It can accurately measure the scale and direction of mutual transformation between blue-green spaces and construction land during urban renewal, thereby effectively revealing the laws of ecological space evolution.

[0131] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal that includes that element.

[0132] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1.A deep learning-based dynamic monitoring method for urban renewal blue-green space, characterized in that, The deep learning-based urban renewal blue-green space dynamic monitoring method comprises: acquiring high-resolution remote sensing images, and preprocessing the high-resolution remote sensing images to obtain preprocessed image data; determining a UNetMamba model, and improving the UNetMamba model by using a ResTv2 backbone network and an ECA attention mechanism to obtain an improved UNetMamba model; determining a preset data set, and training and optimizing the improved UNetMamba model according to the preset data set to obtain a blue-green space classification model; inputting the preprocessed image data into the blue-green space classification model to output a blue-green space classification result; post-processing and spatial transfer metrology analysis of the blue-green space classification result to obtain a blue-green space distribution map and a blue-green space conversion Sankey diagram; determining a preset data set, and training and optimizing the improved UNetMamba model according to the preset data set to obtain a blue-green space classification model, specifically comprising: determining a preset data set, and reclassifying and dividing the data set to obtain a training set and a validation set; training the improved UNetMamba model according to the training set to obtain an initial blue-green space classification model; using a Bayesian optimization algorithm to optimize the initial blue-green space classification model according to the validation set to obtain a blue-green space classification model; training the improved UNetMamba model according to the training set to obtain an initial blue-green space classification model, specifically comprising: inputting the enhanced training data set into the improved UNetMamba model, performing feature extraction processing and global average pooling processing on the enhanced training data set through the ResTv2 backbone network in the improved UNetMamba model to obtain an input feature sequence and a learnable weight matrix; constructing an attention weight matrix according to the input feature sequence and the learnable weight matrix, and performing weighted sum processing and pixel reorganization processing on the attention weight matrix to obtain a multi-scale feature pyramid; inputting the multi-scale feature pyramid into the decoder network of the improved UNetMamba model, performing spatial resolution recovery processing and fusion processing on the multi-scale feature pyramid through the decoder network to output a blue-green space semantic segmentation map, and the improved UNetMamba model training is completed to obtain an initial blue-green space classification model; inputting the enhanced training data set into the initial feature extraction layer in the ResTv2 module, performing feature extraction processing on the enhanced training data set through the initial feature extraction layer to obtain an input feature sequence; ​ Input the input feature sequence into an ECA attention mechanism module, perform global average pooling processing and local cross-channel interaction information calculation on the input feature sequence through the ECA attention mechanism module, and obtain a learnable weight matrix; The attention weight matrix is constructed according to the input feature sequence and the learnable weight matrix, and weighted summation processing and pixel reorganization processing are performed on the attention weight matrix to obtain a multi-scale feature pyramid, specifically including: Performing matrix multiplication operation on the input feature sequence and the learnable weight matrix to obtain a Q vector; Arranging the input feature sequence into a two-dimensional feature map, and performing depth separable convolution processing on the two-dimensional feature map to obtain a compressed feature map; Performing flattening processing and projection processing on the compressed feature map to obtain a K vector and a V vector; An attention weight matrix is obtained according to the Q vector and the K vector, and weighted summation processing is performed on the attention weight matrix and the V vector to obtain an attention output sequence; Performing pixel reorganization processing on the attention output sequence to obtain a multi-scale feature pyramid. 2.The deep learning-based urban renewal blue-green space dynamic monitoring method according to claim 1, wherein, The high-resolution remote sensing image is obtained, and the high-resolution remote sensing image is preprocessed to obtain preprocessed image data, specifically including: Determine the target area, and obtain the high-resolution remote sensing image of the target area; Performing radiation calibration processing, atmospheric correction processing, image registration processing and cropping processing on the high-resolution remote sensing image to obtain preprocessed image data. 3.The deep learning-based urban renewal blue-green space dynamic monitoring method according to claim 1, wherein, The UNetMamba model is determined, and the ResTv2 backbone network and the ECA attention mechanism are used to improve the UNetMamba model to obtain an improved UNetMamba model, specifically including: The UNetMamba model is determined, and the ResTv2 backbone network is used to replace the encoder backbone network in the UNetMamba model to obtain an improved UNetMamba model. The ResTv2 backbone network includes a plurality of feature stages connected in sequence, and each feature stage includes a plurality of series-connected ResTv2 modules and an ECA attention mechanism module. 4.The deep learning-based urban renewal blue-green space dynamic monitoring method of claim 1, wherein, The improved UNetMamba model is trained according to the training set to obtain an initial blue-green space classification model, and the previous also includes: Perform data enhancement processing on the training set to obtain an enhanced training data set, wherein the data enhancement processing includes geometric transformation processing, color disturbance processing and composite enhancement processing. 5.The deep learning-based urban renewal blue-green space dynamic monitoring method according to claim 1, wherein, The preprocessed image data is input into the blue-green space classification model, and a blue-green space classification result is output, specifically including: Performing rule slicing processing on the preprocessed image data to obtain sliced image data; The sliced image data is input into the blue-green space classification model, and the blue-green space classification model is used to perform blue-green space prediction processing on the sliced image data to obtain a blue-green space classification result; The blue-green space classification result includes blue space, green space and construction land. 6.The deep learning-based urban renewal blue-green space dynamic monitoring method according to claim 1, wherein, The blue-green space classification result is post-processed and subjected to spatial transfer metrology analysis to obtain a blue-green space distribution map and a blue-green space conversion Sankey diagram, and specifically includes the following steps: An original spatial position of the high-resolution remote sensing image is acquired, and the blue-green space classification result is spliced according to the original spatial position to obtain a blue-green space distribution map; The blue-green space classification result is subjected to spatial transfer metrology analysis to obtain a blue-green space conversion Sankey diagram.

Images