Urban update blue-green space dynamic monitoring method based on deep learning

Abstract

The invention belongs to the field of data processing, and discloses a city update blue-green space dynamic monitoring method based on deep learning, and the method comprises the steps: obtaining a high-resolution remote sensing image, and carrying out the preprocessing of the image, and obtaining the preprocessed image data; the method comprises the following steps: determining a UNetMama model, and carrying out improvement by adopting a ResTv2 backbone network and an ECA attention mechanism to obtain an improved UNetMama model; determining a preset data set, and performing model training and model optimization on the improved UNetMama model according to the preset data set to obtain a blue-green space classification model; inputting the preprocessed image data into a blue-green space classification model, and outputting a blue-green space classification result; and performing post-processing and space transfer metering analysis on a blue-green space classification result to obtain a blue-green space distribution diagram and a blue-green space conversion mulberry-based diagram. The method can effectively improve the recognition precision of the blue-green space, and further effectively reveals the evolvement rule of the ecological space.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and particularly relates to a city renewal blue-green space dynamic monitoring method based on deep learning. BACKGROUND

[0002] Under the background of global rapid urbanization promoting economic and social development, blue-green space (BGS) covers water bodies (blue space, such as rivers, lakes, etc.) and vegetation-covered areas (green space, such as parks, woodland, wetland, etc.), and becomes a core element of high-density urban development by virtue of its core ecological service functions such as regulating microclimate, conserving water resources, maintaining biodiversity and enhancing urban resilience.

[0003] The traditional research method of blue-green space identification mainly relies on multi-band spectral index and threshold segmentation, or pixel-level classification based on machine learning, but these methods have obvious limitations in the face of complex urban renewal scenarios, resulting in insufficient identification accuracy of blue-green space and failing to meet the quantitative analysis needs of ecological space evolution rules in the process of urban renewal.

[0004] Therefore, the prior art still needs to be improved and developed. SUMMARY

[0005] The main purpose of the present application is to provide a city renewal blue-green space dynamic monitoring method based on deep learning, which aims to solve the problem of insufficient identification accuracy of blue-green space in the face of complex urban renewal scenarios in the prior art, and fails to meet the quantitative analysis needs of ecological space evolution rules in the process of urban renewal.

[0006] To achieve the above purpose, the present application provides a city renewal blue-green space dynamic monitoring method based on deep learning, which comprises the following steps: Obtain high-resolution remote sensing images, and pre-process the high-resolution remote sensing images to obtain pre-processed image data; Determine a UNetMamba model, and improve the UNetMamba model by using a ResTv2 backbone network and an ECA attention mechanism to obtain an improved UNetMamba model; Determine a preset data set, and perform model training and model optimization on the improved UNetMamba model according to the preset data set to obtain a blue-green space classification model; Input the pre-processed image data into the blue-green space classification model, and output a blue-green space classification result; Post-process and spatial transfer metrology analyze the blue-green space classification result to obtain a blue-green space distribution map and a blue-green space conversion Sankey diagram.

[0007] Optionally, the method for dynamic monitoring of urban renewal blue-green space based on deep learning, wherein the high-resolution remote sensing image is obtained, and the high-resolution remote sensing image is preprocessed to obtain preprocessed image data, specifically comprising: determining a target area and obtaining a 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.

[0008] Optionally, the method for dynamic monitoring of urban renewal blue-green space based on deep learning, 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 comprising: determining a UNetMamba model, and replacing the encoder backbone network in the UNetMamba model with a ResTv2 backbone network to obtain an improved UNetMamba model; wherein the ResTv2 backbone network comprises a plurality of feature stages connected in sequence, and each feature stage comprises a plurality of series-connected ResTv2 modules and an ECA attention mechanism module.

[0009] Optionally, the method for dynamic monitoring of urban renewal blue-green space based on deep learning, wherein a preset data set is determined, and the improved UNetMamba model is trained and optimized based on the preset data set to obtain a blue-green space classification model, specifically comprising: determining a preset data set, and performing reclassification processing and data division processing on the preset data set to obtain a training set and a validation set; training the improved UNetMamba model based on the training set to obtain an initial blue-green space classification model; performing hyperparameter optimization processing on the initial blue-green space classification model based on the validation set using a Bayesian optimization algorithm to obtain a blue-green space classification model.

[0010] Optionally, the method for dynamic monitoring of urban renewal blue-green space based on deep learning, wherein before the training of the improved UNetMamba model based on the training set to obtain an initial blue-green space classification model, it further comprises: performing data enhancement processing on the training set to obtain an enhanced training data set, wherein the data enhancement processing comprises geometric transformation processing, color disturbance processing and composite enhancement processing.

[0011] Optionally, the method for dynamic monitoring of urban renewal blue-green space based on deep learning, wherein the model training of the improved UNetMamba model according to the training set to obtain an initial blue-green space classification model specifically comprises: 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 summation 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.

[0012] Optionally, the method for dynamic monitoring of urban renewal blue-green space based on deep learning, wherein the 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 specifically comprises: 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; inputting the input feature sequence into the ECA attention mechanism module, performing global average pooling processing and local cross-channel interaction information calculation on the input feature sequence through the ECA attention mechanism module to obtain a learnable weight matrix.

[0013] Optionally, the method for dynamic monitoring of urban renewal blue-green space based on deep learning, wherein the construction of an attention weight matrix according to the input feature sequence and the learnable weight matrix, and the weighted summation processing and pixel reorganization processing on the attention weight matrix to obtain a multi-scale feature pyramid specifically comprises: 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; The compressed feature map is flattened and projected 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 the attention weight matrix is weighted and summed with the V vector to obtain an attention output sequence; The attention output sequence is subjected to pixel recombination processing to obtain a multi-scale feature pyramid.

[0014] Optionally, the method for dynamically monitoring urban renewal blue-green space based on deep learning, wherein the preprocessed image data is input into the blue-green space classification model to output a blue-green space classification result, specifically comprising: The preprocessed image data is subjected to rule-based slicing processing 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.

[0015] Optionally, the method for dynamically monitoring urban renewal blue-green space based on deep learning, wherein the blue-green space classification result is subjected to post-processing and spatial transfer metrology analysis to obtain a blue-green space distribution map and a blue-green space conversion Sankey diagram, specifically comprising: The original spatial position of the high-resolution remote sensing image is obtained, 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.

[0016] In the present application, high-resolution remote sensing images are acquired, and the high-resolution remote sensing images are preprocessed to obtain preprocessed image data; a UNetMamba model is determined, and the UNetMamba model is improved by using a ResTv2 backbone network and an ECA attention mechanism to obtain an improved UNetMamba model; a preset data set is determined, and the improved UNetMamba model is trained and optimized according to the preset data set to obtain a blue-green space classification model; the preprocessed image data is input into the blue-green space classification model, and a blue-green space classification result is output; the blue-green space classification result is post-processed and spatial transfer metrology analyzed to obtain a blue-green space distribution map and a blue-green space conversion Sankey diagram. The UNetMamba model is improved by introducing the ResTv2 backbone network and the ECA attention mechanism, and then the improved improved UNetMamba model is trained and optimized to obtain the blue-green space classification model, so that the high-resolution remote sensing images after preprocessing can be accurately recognized by the blue-green space classification model, and the blue-green space distribution map and the blue-green space conversion Sankey diagram can be constructed, which can accurately measure the mutual conversion scale and direction between the blue-green space and the construction land in the urban renewal process, and then effectively reveal the ecological space evolution law. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is a flowchart of a preferred embodiment of the present application based on a deep learning urban renewal blue-green space dynamic monitoring method; Figure 2 is a schematic diagram of the overall implementation process of a preferred embodiment of the present application based on a deep learning urban renewal blue-green space dynamic monitoring method; Figure 3 is a schematic diagram of the improved UNetMamba model structure of a preferred embodiment of the present application based on a deep learning urban renewal blue-green space dynamic monitoring method; Figure 4 is a schematic diagram of the 2014-2017-2020 city blue-green space conversion Sankey diagram of a preferred embodiment of the present application based on a deep learning urban renewal blue-green space dynamic monitoring method. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical scheme and advantages of the present application more clear and definite, the present application will be further described in detail below with reference to the drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0019] Under the background of rapid urbanization, blue-green space (BGS) is composed of water bodies (blue space, such as rivers and lakes) and vegetation-covered areas (green space, such as parks, forests, and wetlands). With its core ecological services of regulating microclimate, conserving water resources, maintaining biodiversity, and enhancing urban resilience, BGS has become a key element of high-density urban development. Currently, many large cities have gradually shifted from large-scale expansion to urban renewal dominated by stock optimization. In this transition process, the changing pattern of urban BGS is more complex. On the one hand, urban renewal may increase BGS through ecological restoration and park construction. On the other hand, high-intensity redevelopment and infrastructure upgrading may lead to the invasion, fragmentation, and even degradation of BGS. These dynamic changes are directly related to the ecological benefits and development level of urban renewal.

[0020] Therefore, to study the evolution of BGS pattern, it is not only necessary to accurately identify its spatial distribution and morphology, but also to accurately measure its mutual transfer relationship with construction land and quantify the loss and restoration effect 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 have obvious limitations when faced with complex urban renewal scenarios. On the one hand, the limited number of high-resolution image bands limits the improvement of spectral methods. On the other hand, traditional methods rely on expert-designed features, which have high professional threshold and poor generalization, making it difficult to fully exploit high-level semantic information and resulting in insufficient recognition accuracy of small and micro BGS. More importantly, existing methods mostly stop at static recognition and lack the ability to accurately measure the dynamic transfer process of “BGS- construction land”, which cannot meet the quantitative analysis needs of ecological space evolution patterns in the urban renewal process.

[0021] In recent years, the application of deep learning technology, especially computer vision, in the field of remote sensing image semantic segmentation has provided a new technical path to improve the recognition accuracy of BGS. Deep learning models represented by convolutional neural networks (CNN) can automatically learn multi-level feature representations from raw images, significantly reducing the dependence on human-designed features. However, traditional CNN models are limited by local receptive fields and repeated downsampling operations, and have weak global context information capture ability, making it difficult to effectively model the semantic association between large-scale features. Although semantic segmentation models such as U-Net and DeepLab perform well in general scenarios, they still have problems such as loss of spatial details, insufficient sensitivity to small and micro target recognition, and inability to support subsequent spatial transfer accurate measurement when processing high-resolution urban remote sensing images.

[0022] Therefore, it is urgent to develop a deep learning dynamic monitoring method that can balance global perception and detail preservation and has high computational efficiency, not only to realize high-precision dynamic monitoring of urban renewal blue-green space, but also to accurately measure the transfer relationship between blue-green space and construction land, so as to provide reliable technical support for evaluating the ecological benefits of urban renewal process and optimizing the land space protection pattern.

[0023] In view of the deficiencies of the prior art in blue-green space identification and dynamic analysis, the present application provides a deep learning-based dynamic monitoring method for urban renewal blue-green space, aiming to solve the core problems of small and micro target identification difficulty, serious detail loss in high-resolution urban images, and the inability of traditional methods to effectively quantify the dynamic conversion relationship between blue-green space and construction land. The method takes domestic GF-1 satellite 2-meter resolution remote sensing images as the basic data source, relies on the open source WHDLD dataset (Wuhan dense labeling dataset, a kind of dense labeling dataset set by the present application) for model training and verification, and realizes the full-process automation processing from the input of original remote sensing images to the distribution mapping and conversion rule analysis of blue-green space through a series of steps such as image preprocessing, improved semantic segmentation model construction, model training and hyperparameter optimization, reasoning and output, and spatial transfer measurement analysis. The core scheme of the present application is to construct a deep learning framework that integrates multi-scale feature extraction and channel attention mechanism, adopts a ResTv2 encoder structure (a new encoder backbone network designed by the present application) with detail preservation capability, and combines a lightweight and efficient channel attention (ECA, Efficient Channel Attention) module, which can significantly improve the recognition accuracy and robustness of the model in complex urban environment. On this basis, a spatial transfer matrix of multi-period blue-green space and construction land is further constructed to realize accurate quantification of the ecological space loss and restoration process.

[0024] The technical effects of the present application include: 1. Significantly improve the recognition accuracy of small and micro blue-green space, the model average intersection over union (mIoU, mean Intersection over Union) reaches more than 70%; 2. Accurately measure the mutual conversion scale and direction between blue-green space and construction land in the urban renewal process, and reveal the ecological space evolution law; 3. The model structure is lightweight and has high computational efficiency, which is suitable for rapid processing and dynamic analysis of large-scale remote sensing images. The present application not only provides a technical means for high-precision identification of urban blue-green space, but more importantly, provides reliable quantitative basis and decision support for scientific evaluation of the ecological benefits of urban renewal and optimization of land space planning.

[0025] The deep learning-based dynamic monitoring method for urban renewal blue-green space of the preferred embodiment of the present application, as Figure 1As shown, the deep learning-based urban renewal blue-green space dynamic monitoring method comprises the following steps: Step S10, obtaining high-resolution remote sensing images, and preprocessing the high-resolution remote sensing images to obtain preprocessed image data.

[0026] As shown Figure 2 As shown, the present application comprises 5 steps: the first step: pre-processing the input high-resolution remote sensing images, the pre-processing includes radiation calibration, atmospheric correction, image registration and cropping. The second step: constructing an improved UNetMamba model (U-shaped model based on state space model), introducing ResTv2 backbone network and ECA attention mechanism to enhance the feature extraction capability of UNetMamba model. The third step: using the reclassified WHDLD dataset to train the improved UNetMamba model, combining data enhancement and hyperparameter optimization to improve the model performance. The fourth step: generating blue-green space classification results through model inference, and obtaining the final spatial distribution map through post-processing. The fifth step: based on the multi-period blue-green space distribution results, constructing a spatial transfer matrix to finely measure the conversion relationship between blue-green space and construction land. The present application realizes the full automation processing from the original remote sensing image to the blue-green space dynamic change measurement analysis.

[0027] The present application takes a city as the research area, analyzes the blue-green space distribution and change of the area in the urban renewal stage. The present application sequentially identifies and transfer measurement analyzes the blue-green space of the research area in 2014, 2017 and 2020, thereby verifying the applicability and stability of the present application in different periods.

[0028] Specifically, a target area is determined, and high-resolution remote sensing images of the target area are obtained; the high-resolution remote sensing images are subjected to radiation calibration processing, atmospheric correction processing, image registration processing and cropping processing to obtain preprocessed image data.

[0029] The preprocessing process of high-resolution remote sensing images is as follows: the present application selects 2-meter resolution multispectral remote sensing images collected by domestic GF-1 satellite covering the whole city in 2014, 2017 and 2020 as the basic data source, wherein the original data of each period is composed of multiple images, all images have undergone strict quality screening and control, and the cloud coverage rate of single image is less than 5%, so as to minimize the influence of cloud cover on feature recognition accuracy.

[0030] The image preprocessing module is arranged in the application, is responsible for standardizing and splicing the obtained multi-scene original remote sensing image (i.e. the high-resolution remote sensing image in the application), generating standardized data products covering the whole city and meeting the input requirements of the deep learning model, and the specific processing procedure includes the following four core steps: 1. Radiometric calibration: converting the digital quantization value of each image into surface reflectivity, eliminating the radiation response difference of the sensor itself. Among them, the calibration coefficient provided by the sensor is used to realize the radiation correction through linear transformation, so as to obtain surface reflectivity data with consistent physical meaning.

[0031] 2. Atmospheric correction: using the FLAASH (Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes, an atmospheric calibration tool) model to correct the atmosphere of each image after radiometric calibration, so as to eliminate the influence of atmospheric scattering and absorption on the surface reflectivity. By inputting imaging time, geographic location and aerosol type and other parameters, the atmospheric condition is accurately estimated and corrected.

[0032] 3. Image registration and splicing: first, the registration method based on SIFT (Scale Invariant Feature Transform, scale invariant feature transform) feature detector is used, one image in 2020 is selected as the reference, the feature points are extracted and the random sample consensus algorithm is used to remove the mismatched points, the sub-pixel level accurate registration of all images in 2014 and 2017 with the reference image is realized respectively, the consistency of the spatial geometric position of the multi-period images is ensured, and the registration error is less than 0.5 pixels. Then, the multi-scene images of the same period and after registration are seamlessly spliced to generate a complete image covering the whole study area.

[0033] 4. Image cropping: based on the vector boundary of the administrative division of a city, the city-wide images of the three periods spliced respectively are accurately cropped, and finally the 2-meter resolution remote sensing image data covering the whole city and having unified spatial reference in 2014, 2017 and 2020 are obtained.

[0034] Step S20, determine the UNetMamba model, and improve the UNetMamba model by using the ResTv2 backbone network and the ECA attention mechanism to obtain an improved UNetMamba model.

[0035] The improved UNetMamba model inherits the encoder-decoder basic framework and skip connection structure of the classic UNet architecture as a whole to ensure effective fusion and transmission of multi-scale feature information. In view of the core challenges such as small target recognition difficulty, global context information capture and detail preservation difficulty in the blue-green space recognition of high-resolution remote sensing images, the encoder backbone network and feature enhancement mechanism of the original UNetMamba model are improved, aiming to comprehensively improve the perception accuracy and robustness of the model.

[0036] Specifically, the UNetMamba model is determined, and the ResTv2 backbone network is used to replace the encoder backbone network in the UNetMamba model to obtain the improved UNetMamba model; wherein the ResTv2 backbone network includes a plurality of feature stages connected in turn, and each feature stage includes a plurality of ResTv2 modules connected in series and an ECA attention mechanism module.

[0037] The improved model structure is as shown in Figure 3 The original ResT structure in the UNetMamba model is replaced by the ResTv2.

[0038] One of the core innovations of the present application is the important optimization of the encoder backbone network. Although the ResT structure used in the original UNetMamba model can improve the model performance, it still has obvious shortcomings when processing high-resolution remote sensing images. The multi-head interaction module leads to high computational complexity and limited inference speed. In addition, the original UNetMamba model needs to strengthen the ability to maintain image detail information during the downsampling process.

[0039] To solve the above problems, the ResTv2 is introduced as a new encoder backbone network to replace the original ResT structure. This encoder undertakes the key task of layer-by-layer feature extraction and abstraction of input remote sensing images. Specifically, the input tensor (i.e. training set) first enters the first stage of the ResTv2 backbone network. This network includes four feature stages (Stage) connected in turn. Each stage gradually downsamples the input feature map and increases the number of channels. Finally, a set of multi-scale feature pyramids (Feature Pyramid) is output, effectively capturing semantic information at different levels.

[0040] Further, each feature stage in the ResTv2 backbone network includes a plurality of ResTv2 modules connected in series, and the core is the internal efficient multi-head self-attention mechanism (i.e. EMSAv2).

[0041] Step S30, determine a preset data set, and perform model training and model optimization on the improved UNetMamba model according to the preset data set, to obtain a blue-green space classification model.

[0042] Specifically, a preset data set is determined, and the preset data set is subjected to reclassification processing and data division processing to obtain a training set and a validation set; the training set is subjected to data enhancement processing to obtain an enhanced training data set, wherein the data enhancement processing includes geometric transformation processing, color disturbance processing, and composite enhancement processing.

[0043] For the training and optimization process of the model, the improved UNetMamba model is trained and optimized by using the reclassified WHDLD data set, and the process specifically includes the following steps: 1. Training data preparation: the open-source WHDLD data set is used as the basic data for model training and verification. To meet the goal of blue-green space recognition, the original annotation system of the data set is reclassified, and the reclassification process includes: classifying "water body" as blue space, "vegetation" as green space, and "building", "road", "sidewalk" and "bare soil" as construction land, thereby constructing a three-class semantic segmentation system for the present application. Further, to ensure the objectivity and statistical effectiveness of model evaluation, the reclassified data set is randomly divided into a training set (a total of 3705 samples) and a validation set (a total of 1235 samples) at a ratio of about 3:1.

[0044] 2. Data enhancement: to improve the generalization ability and robustness of the model in the face of complex and variable remote sensing image scenes, data enhancement strategies (Data Augmentation) are implemented during training, mainly including the following three categories: geometric transformation: including random horizontal flipping, vertical flipping and rotation within ±45°; color disturbance: randomly adjusting brightness and contrast (adjustment range ±30%) in RGB space, and randomly disturbing hue, saturation and brightness in HSV space; composite enhancement: using Mosaic augmentation (Mosaic augmentation algorithm) technology, randomly cropping and splicing four training images to simulate complex scenes and enhance the model's context perception ability for multi-scale targets.

[0045] The enhanced training data set is input into the improved UNetMamba model, and the enhanced training data set is input into the initial feature extraction layer in the ResTv2 module, and the initial feature extraction layer is used to perform feature extraction processing on the enhanced training data set to obtain an input feature sequence.

[0046] The input feature sequence is input to the ECA attention mechanism module, and global average pooling processing and local cross-channel interaction information calculation are performed on the input feature sequence by the ECA attention mechanism module to obtain a learnable weight matrix.

[0047] The ECA attention mechanism embedding process is as follows: in order to strengthen the focusing ability of the model on key feature channels in a complex scene, an efficient channel attention (ECA) module is embedded at the beginning of each stage of the encoder path. As shown in Figure 3 the ECA module is embedded after the initial feature extraction layer (Stem) for Stage 1 (the first stage), and the channel weight of the output primary feature map is recalibrated; for Stage 2 (the second stage) to Stage 4 (the fourth stage), each stage receives a down-sampled feature map, which is then processed by the ECA module before being input to the subsequent feature extraction unit in the 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 features most relevant to blue-green space recognition.

[0048] The operation process of the ECA module is as follows: 1. First, the input feature map is globally averaged in the spatial dimension to obtain a channel description vector; 2. Then, local cross-channel interaction information is captured through one-dimensional convolution (the convolution kernel size k is adaptively determined by the number of channels C); 3. Finally, the channel weight is 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, i.e., each channel of the feature map is multiplied by its corresponding weight value (where B is Batch Size, which means the number of data samples input to the model for parallel computation at one time; C is Channels, which means the dimension of the feature map. For an input image, it is usually the color channel; for an 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). By embedding the lightweight ECA module in front, the model significantly enhances its perception and selection ability for channel dimension features without significantly increasing the computational complexity of the model, effectively improving the sensitivity to small and micro blue-green space targets and their robustness in complex backgrounds.

[0049] The input feature sequence and the learnable weight matrix are subjected to matrix multiplication operation to obtain a Q vector; the input feature sequence is arranged into a two-dimensional feature map, and the two-dimensional feature map is subjected to deep separable convolution processing to obtain a compressed feature map; the compressed feature map is subjected to flattening processing and projection processing 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 the attention weight matrix is subjected to weighted summation processing with the V vector to obtain an attention output sequence; the attention output sequence is subjected to pixel recombination processing to obtain a multi-scale feature pyramid.

[0050] The specific processing procedure of the ResTv2 backbone network is as follows: 1. First, the input feature sequence and the learnable weight matrix are subjected to matrix multiplication operation to generate a Query (Q) vector, and the input feature sequence is rearranged into a two-dimensional feature map form with a height H and a width W.

[0051] 2. Deep separable convolution operation is performed on the two-dimensional feature map, that is, first, spatial convolution is performed on each input channel separately using deep convolution to capture spatial features, and then point-by-point convolution (1x1 convolution) is used to integrate the outputs of each channel to realize spatial dimension downsampling and information compression.

[0052] 3. After the compressed feature map (i.e., the compressed feature map in the present application) is flattened into a sequence, Key (K) and Value (V) vectors are obtained by projection.

[0053] 4. The dot product of the transpose of Q and K is calculated, divided by a scaling factor (the scaling factor is generally the square root of the dimension of the Key vector), and then normalized by a Softmax function to obtain an attention weight matrix, which is then subjected to weighted summation with the V vector to obtain an attention output sequence.

[0054] 5. Finally, pixel recombination operation is performed on the attention output sequence, that is, the low-resolution high-channel feature map (B, CxrH, W) is rearranged into a high-resolution low-channel feature map (B, C, Hxr, Wxr), where r is the up-sampling ratio, to recover the detail information. 2

[0055] 6. The attention output sequence and the original input feature are fused by element-level addition through a shortcut connection to complete the forward calculation of the module.

[0056] With the hourglass structure composed of "convolution downsampling-pixel recombination up-sampling", the ResTv2 encoder can not only efficiently capture the global context information in remote sensing images, but also significantly improve the ability to preserve local detail features of ground objects, especially laying a key technical foundation for subsequent accurate identification of small and micro blue-green spaces (such as pocket parks, ecological ponds, etc.).​

[0057] The multi-scale feature pyramid is input to a decoder network of the improved UNetMamba model, the multi-scale feature pyramid is subjected to spatial resolution recovery processing and fusion processing through the decoder network, and a blue-green space semantic segmentation map is output, and the improved UNetMamba model is trained to obtain an initial blue-green space classification model.

[0058] The decoder receives the deepest level features output by the encoder (i.e., the multi-scale feature pyramid in the present application) and gradually recovers the spatial resolution through a series of up-sampling convolution layers. After each up-sampling, the decoder receives the feature map from the corresponding stage of the encoder path through a jump connection and fuses it to supplement the details that may be lost in the up-sampling process, and finally gradually reconstructs a high-resolution, high-precision blue-green space semantic segmentation map.

[0059] In summary, the improved UNetMamba model constructed in the present application is not a simple repair of the original UNetMamba model, but through the introduction of the ResTv2 encoder and the embedded multi-level ECA attention mechanism, a new architecture with synergistic effect is formed. This architecture fully utilizes the advantages of ResTv2 in global modeling and detail preservation, and at the same time, with the help of the ECA module, it realizes the adaptive optimization of the feature channels, and finally the recognition accuracy and robustness of the model in the complex urban environment are significantly improved.

[0060] The Bayesian optimization algorithm is used to perform hyperparameter optimization processing on the initial blue-green space classification model according to the validation set, and a blue-green space classification model is obtained.

[0061] The specific processing process of the hyperparameter optimization is as follows: the core evaluation index of model performance is set as the mean Intersection over Union (mIoU) on the validation set, in order to obtain the optimal model performance, the Optuna automatic hyperparameter search framework based on the Bayesian optimization algorithm is adopted. The framework models the black box function relationship between the validation set mIoU and the hyperparameters by constructing a probability agent model (usually a Gaussian process model), and balances exploration and utilization by means of a collection function (such as expected improvement EI) to guide efficient and directional search of hyperparameter combinations. The hyperparameter optimization space includes learning rate (lr), weight decay, training batch size and validation batch size. First, through 30 rounds of prior exploration experiments (that is, randomly sampling initial hyperparameter combinations in a large range of values), the effective value range of each hyperparameter is preliminarily determined, and the area with significantly low performance is excluded, providing a high-quality starting point for subsequent fine search. On this basis, 200 rounds of iterative fine search are carried out. In each iteration, Optuna updates the agent model based on the existing experimental results, and recommends the next most potential hyperparameter combination by maximizing the collection function, and then trains and validates the combination, obtains its mIoU value and feeds it back to the optimizer, gradually approaching the global optimal solution. Finally, the optimal hyperparameter combination that maximizes the validation set mIoU is determined: learning rate lr = 3.3x10 -5 , weight decay wd = 2.11x10 -4 , training batch size = 32, and validation batch size = 16. Based on this parameter combination, the model achieves an mIoU of 84.25% on the validation set, which is about 3.7 percentage points higher than the baseline model. The optimal model (i.e., the blue-green space classification model in the present application) is saved and used for subsequent inference and blue-green space recognition applications.

[0062] Step S40, input the preprocessed image data into the blue-green space classification model, and output the blue-green space classification result.

[0063] The specific processing procedure 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 the present application) is input into the model, first multi-level feature extraction is performed via the improved encoder, and before each stage of the encoder path performs depth feature transformation, the input feature is first rescaled in the channel dimension by the ECA module to strengthen the key features. Subsequently, the ResTv2 module captures long-range context dependencies by virtue of its efficient multi-head self-attention mechanism and preserves detailed information through its unique hourglass structure, and the encoder finally outputs a multi-scale feature pyramid with rich global semantics and local details. Then, the decoder in the blue-green space classification model is processed, and the blue-green space classification result is output.

[0064] Specifically, the preprocessed image data is subjected to rule-based slicing to obtain sliced image data; the sliced image data is input into the blue-green space classification model, and the sliced image data is subjected to blue-green space prediction processing by the blue-green space classification model to obtain a blue-green space classification result; wherein the blue-green space classification result includes a blue space, a green space and a construction land.

[0065] The specific process of outputting the blue-green space classification result is as follows: the present application applies the optimal hyperparameter model (i.e., the blue-green space classification model in the present application) obtained by training to the high-resolution remote sensing image data of a certain city in 2014, 2017 and 2020, realizes the automatic identification and product generation of multi-period blue-green space, and specifically includes the following steps: 1. Input image slices: the preprocessed remote sensing images of the whole city in the three periods are respectively subjected to rule-based slicing according to the size of 256x256 pixels, and 9447 effective image slices are obtained as model inputs, the slicing size is consistent with the sample size used in the model training stage, and the input data format is matched with the model structure.

[0066] 2. Model inference: the image slices of the three periods are input into the blue-green space classification model with the optimal parameters (i.e., the learning rate lr = 3.3x10 -5 , the weight decay wd = 2.11x10 -4 ) loaded in batches to perform forward inference, and the model output is a pixel-level semantic segmentation result (i.e., the blue-green space classification result in the present application), each pixel is assigned a class label, which corresponds to a blue space (water body), a green space (vegetation) or a construction land.

[0067] Step S50, post-processing and spatial transfer metrology analysis are performed on the blue-green space classification result to obtain a blue-green space distribution map and a blue-green space conversion Sankey diagram.

[0068] Specifically, the original spatial position of the high-resolution remote sensing image is acquired, and the blue-green space classification result is stitched 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 Sang diagram.

[0069] The semantic recognition slices of 2014, 2017 and 2020 are stitched according to their original spatial positions to reconstruct a blue-green space distribution map covering the entire city. In the stitching process, based on the geographic coordinates of each pixel, the slices are accurately aligned to the corresponding geographic area, and the blue space, green space and construction land are uniformly valued to form a seamless city-wide classification result map. The result retains complete geographic coordinate information and can be directly used for spatial analysis, mapping output and multi-temporal change research.

[0070] The accuracy verification process of the blue-green space classification result output is as follows: in order to objectively evaluate the reliability of the recognition results of multiple periods, the present application uses stratified random sampling method to statistically verify the recognition results of each period. 845 samples located in the administrative boundary which may cause incomplete identification are removed from 9477 slices of each period, and 8602 valid units are reserved as the verification population. According to the Cochran sample size calculation formula, under the condition of 95% confidence level (Z=1.96) and 5% allowable error, the minimum sample size is determined to be 400 (sampling ratio is 4.65%) to ensure the representativeness of the estimation result.

[0071] The verification samples are manually annotated by a professional team to form the true value reference data, and three widely used semantic segmentation evaluation indexes are used for quantitative analysis, and the verification results are shown in Table 1: Among them, the semantic segmentation evaluation indexes include: 1. Pixel accuracy (PA, Pixel Accuracy): the proportion of overall correct pixels; 2. Mean class accuracy (MCA, Mean Class Accuracy): the arithmetic mean of the accuracy of each class, used to evaluate the balance of model classification; 3. Mean intersection over union (mIoU): the mean of the intersection over union between the predicted result and the true value, which comprehensively reflects the consistency of the class and the boundary.

[0072] Table 1: Blue-green space recognition accuracy verification results (%)

[0073] The results show that the method proposed by the present application maintains high recognition accuracy and stability in different periods, and the three indexes are good, which verifies that the method has strong generalization ability and practicality.

[0074] Further, the present application adopts spatial transfer measurement analysis: in order to accurately quantify the dynamic conversion relationship between blue-green space and construction land in the urban renewal process, and identify the hot spot area of spatial change, the present application is based on the classification results of multiple periods, and the following analysis steps are carried out: 1. Transition matrix construction: taking the classification results of the previous period as the reference, the transition matrices of blue-green space and corresponding construction land in the periods of 2014-2017 and 2017-2020 are constructed respectively. The matrix is realized through cross-classification analysis, and the mutual conversion area and direction between different spaces are systematically counted, which provides basic data support for quantifying the spatial conversion rule in the urban renewal process.

[0075] 2. Spatial change coding: in order to standardize the change detection process and realize pixel-by-pixel change identification, first of all, the digital coding rules of each space category are defined: construction land corresponding category code (1), blue space corresponding category code (2), and green space corresponding category code (3).

[0076] Based on the coding system, the pixel-by-pixel change detection and recoding processing are carried out on the classification results of two periods: the combination of the category labels of the same pixel in the spatial distribution maps of the previous and next two periods forms the change type identification, and the change detection result map is automatically generated. In the map, the value of each pixel is uniformly represented by a two-digit code, the ten digits represent the previous period land class code, and the individual digits represent the next period land class code. For example, the code "21" indicates the conversion of blue space (2) to construction land (1), and the code "32" indicates the conversion of green space (3) to blue space (2). Through this coding and detection method, the result map containing 9 possible change types is generated, which provides structured input for subsequent quantitative analysis and hotspot identification.

[0077] 3. Focus on two types of changes: based on the change detection results, the analysis focuses on two key conversion processes: one is the conversion of blue-green space to construction land, that is, the pixels with change type code "21" (blue space→construction land) and "31" (green space→construction land) are extracted, which are used to identify the specific location and scale of ecological space occupation; the other is the conversion of construction land to blue-green space, that is, the pixels with change type code "12" (construction land→blue space) and "13" (construction land→green space) are extracted, which are used to evaluate the spatial distribution and effectiveness of ecological restoration and construction.

[0078] 4. Visualization and output of results: generate the blue-green space conversion Sankey diagram of a city in 2014-2017-2020 (such as Figure 4The flow path and flow size between spaces in different periods are clearly shown. Meanwhile, based on the change detection result, a representative conversion type (such as “green space → construction land”) is selected to generate a change hotspot distribution map, so that the spatial aggregation characteristics and distribution rules of the typical conversion process are intuitively presented. Combined with the quantitative statistics of the transfer matrix and the spatial distribution of the typical change hotspots, the driving mechanism of the blue-green space evolution in urban renewal can be further analyzed, and multi-dimensional data support is provided for the evaluation of the evolution rules of ecological space.

[0079] In summary, the present application realizes high-precision, automatic identification and fine measurement analysis of urban blue-green space through efficient image preprocessing, improved deep learning model structure, scientific training method and inference process, and accurate spatial transfer measurement analysis, and provides a reliable technical tool and data support for urban ecological planning and management.

[0080] The present application proposes the following innovative key points and protection schemes to solve the core technical problems of small and micro target recognition difficulty, weak global context information capture ability and serious detail loss in blue-green space recognition in high-resolution urban remote sensing images.

[0081] The innovative key points of the present application are as follows: 1. Improved deep learning model architecture: The present application constructs an improved UNetMamba model to solve the problems of small target recognition difficulty, weak global context information capture ability and serious detail loss in blue-green space recognition in high-resolution remote sensing images. The model retains the basic framework of UNet encoder-decoder, and introduces ResTv2 encoder backbone network and embedded lightweight efficient channel attention (ECA) module, which significantly improves the feature extraction ability and recognition accuracy of the model. The ResTv2 encoder has a unique “convolution downsampling-pixel reorganization upsampling” hourglass structure, which effectively maintains the details of ground objects while capturing global context information. The ECA module adjusts the adaptive channel weight to enhance the focusing ability on key features and reduce redundant information interference.

[0082] 2. Fine spatial transfer measurement analysis method based on multi-period recognition results: The present application proposes a complete dynamic transfer measurement analysis process of blue-green space, which realizes the whole process from pixel-level change detection to type conversion association by constructing a spatial transfer matrix and a change coding system. Based on the classification results of the previous and subsequent periods, the method automatically identifies and quantifies the bidirectional conversion between blue-green space and construction land by defining class coding and change type identification, and visualizes the results by means of Sankey diagram and spatial distribution hotspot map, which provides data support for the analysis of ecological space evolution mechanism in urban renewal process.

[0083] The application aims to protect a deep learning-based urban renewal blue-green space dynamic monitoring method, the core of which is to construct a deep learning framework integrating multi-scale feature enhancement and channel attention mechanism, aiming to solve the problems of small and micro target recognition difficulty in high-resolution urban images, serious detail loss, and the inability of traditional methods to effectively quantify the dynamic conversion relationship between blue-green space and construction land.

[0084] The deep learning-based urban renewal blue-green space dynamic monitoring method provided by the application comprises the following complete technical processes: 1. First, high-resolution remote sensing images are acquired and preprocessed, including radiation calibration, atmospheric correction, image registration, splicing and cutting, to generate standardized input data with unified spatial reference; 2. Then, an improved UNetMamba semantic segmentation model is constructed, which is based on an encoder-decoder structure, the encoder uses a ResTv2 backbone network to realize multi-scale feature extraction and detail preservation, and an efficient channel attention (ECA) module is embedded at the beginning of each encoding stage to enhance key channel features and suppress noise; 3. Then, the model is trained using the reclassified WHDLD dataset, the reclassification system includes three categories of blue space, green space and construction land, and data enhancement strategies and hyperparameter optimization methods are introduced during the training process to improve the model's generalization ability and recognition accuracy; 4. The trained model is used to infer the preprocessed images of multiple periods to realize pixel-level semantic segmentation, and the blue-green space distribution maps of different periods are generated after splicing and post-processing; 5. Finally, based on the multi-period blue-green space distribution results, a spatial transfer matrix is constructed to identify the mutual conversion relationship between blue-green space and construction land through change type coding, and the conversion process is quantitatively analyzed and visually expressed by means of a trophic diagram and a spatial distribution hotspot map.

[0085] On the basis of this embodiment, there are various possible design directions and deformation schemes, all of which belong to the protection scope of the application, including but not limited to: 1. Alternative and modification of model architecture: The encoder backbone network of the improved UNetMamba model can adopt other network structures with strong global context modeling capability and detail preservation mechanism, such as replacing it with SwinTransformer, PVTv2 (Pyramid Vision Transformer v2), or EfficientNetv2, etc. advanced visual backbone network. The ECA (Efficient Channel Attention) module can be replaced with an SE (Squeeze-and-Excitation) module to enhance channel dependency modeling, or CBAM (Convolutional Block Attention Module) to realize channel and spatial attention collaborative optimization. The decoder part of the UNet architecture can be optimized, such as introducing FPN (Feature Pyramid Network), PANet (Aggregation Network for Instance Segmentation), etc. structure to enhance multi-scale feature fusion capability, or using sub-pixel convolution instead of transposed convolution for up-sampling to reduce the chessboard effect.

[0086] 2. Extension of data processing and enhancement strategy: The atmospheric correction model in the image preprocessing link can not be limited to FLAASH, but also can use other mainstream atmospheric correction models such as 6S, DOS, etc. The data enhancement strategy can be further expanded, in addition to geometric transformation, color disturbance, and Mosaic, advanced regularization techniques such as CutOut, CutMix, MixUp, or enhancement methods specific to remote sensing image characteristics, such as simulating different cloud and fog coverage, sensor noise, etc. The training data source can not be limited to the WHDLD dataset, but also compatible with other open source or self-built remote sensing image semantic segmentation datasets, and support custom reclassification of the feature class system according to specific application scenarios.

[0087] 3. Adjustment of training and optimization strategy: The hyperparameter optimization algorithm can not be limited to Optuna, but also can use GridSearch, Random Search, Hyperband, or other Bayesian optimization tool libraries (such as Scikit-optimize) to realize. The loss function in the model training can not be limited to the standard cross-entropy loss, but can introduce Dice Loss, Focal Loss, etc. loss function suitable for semantic segmentation tasks or their combination to better handle class imbalance problems or optimize specific evaluation indicators. The optimizer can not be limited to AdamW or SGD, but also can choose adaptive optimization algorithms such as Adam, NAdam, RAdam, etc. and cooperate with various learning rate scheduling strategies (such as Cosine Annealing, Step Decay, OneCycleLR, etc.).

[0088] 4. Application scenarios and data source expansion: The method can not be limited to processing GF-1 satellite data, and can also be applied to other domestic and foreign high-resolution remote sensing image data, such as GF-2, ZY-3, Sentinel-2, Landsat series, WorldView and image data of different resolutions. The method can not be limited to identifying urban blue-green space, and can be extended to other remote sensing image semantic segmentation tasks, such as crop classification, building extraction, road extraction, land use / cover classification, etc. after appropriate adjustment and training.

[0089] In summary, the present application provides a kind of urban renewal blue-green space dynamic monitoring method based on deep learning, the method comprises: obtaining high-resolution remote sensing image, and the high-resolution remote sensing image is preprocessed, and the preprocessed image data is obtained;Determine UNetMamba model, and the ResTv2 backbone network and ECA attention mechanism are used to improve the UNetMamba model, and the improved UNetMamba model is obtained;Determine the preset data set, and according to the preset data set, the improved UNetMamba model is trained and optimized, and the blue-green space classification model is obtained;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 spatial transfer measurement analysis, and the blue-green space distribution map and blue-green space conversion Sanji diagram are obtained.The UNetMamba model is improved by introducing ResTv2 backbone network and ECA attention mechanism, and then the improved UNetMamba model is trained and optimized, so as to obtain the blue-green space classification model, which can accurately identify the blue-green space of the preprocessed high-resolution remote sensing image through the blue-green space classification model, and construct the blue-green space distribution map and the blue-green space conversion Sanji diagram, which can accurately measure the mutual conversion scale and direction between blue-green space and construction land in the process of urban renewal, and further effectively reveal the ecological space evolution law.

[0090] It should be noted that in this paper, the term "includes", "contains" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or terminal including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or terminal. Without more limitation, the element defined by the sentence "includes a" does not exclude the presence of another identical element in the process, method, article or terminal including the element.

[0091] It should be understood that the application is not limited to the examples described above, which can be modified or adapted in various ways by those skilled in the art without thereby departing from the scope of the present application, as defined in the appended claims.

Claims

1. A method for dynamic monitoring of urban renewal blue-green space based on deep learning, characterized in that, The deep learning-based method for dynamic monitoring of urban renewal blue-green spaces includes: Acquire high-resolution remote sensing images and preprocess the high-resolution remote sensing images to obtain preprocessed image data; 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. 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 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.

2. The method for dynamic monitoring of urban renewal blue-green space based on deep learning according to claim 1, characterized in that, The acquisition of high-resolution remote sensing images and the preprocessing of the high-resolution remote sensing images to obtain preprocessed image data specifically include: Identify the target area and acquire high-resolution remote sensing images of the target area; The high-resolution remote sensing image is subjected to radiometric calibration, atmospheric correction, image registration, and cropping to obtain preprocessed image data.

3. The method for dynamic monitoring of urban renewal blue-green space based on deep learning according to claim 1, characterized in that, The process of determining the UNetMamba model and improving it using a ResTv2 backbone network and ECA attention mechanism to obtain an improved UNetMamba model specifically includes: 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. The ResTv2 backbone network includes multiple sequentially connected feature stages, each of which includes multiple cascaded ResTv2 modules and an ECA attention mechanism module.

4. The method for dynamic monitoring of urban renewal blue-green space based on deep learning according to claim 3, characterized in that, The process of 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: A preset dataset is determined, and the preset dataset is reclassified and partitioned to obtain a training set and a validation set. The improved UNetMamba model is trained using the training set to obtain an initial blue-green space classification model. 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.

5. The method for dynamic monitoring of urban renewal blue-green space based on deep learning according to claim 4, characterized in that, The step of training the improved UNetMamba model based on the training set to obtain the initial blue-green space classification model also includes: 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.

6. The method for dynamic monitoring of urban renewal blue-green space based on deep learning according to claim 5, characterized in that, The step of training the improved UNetMamba model based on the training set to obtain the initial blue-green space classification model specifically includes: 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. 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. 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.

7. The method for dynamic monitoring of urban renewal blue-green space based on deep learning according to claim 6, characterized in that, The step of performing feature extraction and global average pooling on the enhanced training dataset using the ResTv2 backbone network in the improved UNetMamba model to obtain the input feature sequence and learnable weight matrix specifically includes: 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. 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.

8. The method for dynamic monitoring of urban renewal blue-green space based on deep learning according to claim 6, characterized in that, The step of constructing an attention weight matrix based on the input feature sequence and the learnable weight matrix, and then performing weighted summation and pixel recombination on the attention weight matrix to obtain a multi-scale feature pyramid, specifically includes: Perform matrix multiplication 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 V vector; 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. The attention output sequence is subjected to pixel recombination processing to obtain a multi-scale feature pyramid.

9. The method for dynamic monitoring of urban renewal blue-green space based on deep learning according to claim 1, characterized in that, The step of inputting the preprocessed image data into the blue-green space classification model and outputting the blue-green space classification result specifically includes: 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 prediction processing of the sliced ​​image data is performed by the blue-green space classification model to obtain the blue-green space classification result. The blue-green space classification results include blue space, green space, and construction land.

10. The method for dynamic monitoring of urban renewal blue-green space based on deep learning according to claim 1, characterized in that, The post-processing and spatial transfer econometric analysis of the blue-green space classification results yields a blue-green space distribution map and a blue-green space transformation Sankey diagram, specifically including: 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 econometric analysis was performed on the blue-green space classification results to obtain the blue-green space transformation Sankey diagram.

Images