Urban green land high-resolution remote sensing monitoring method and system

Abstract

The invention relates to an urban green land high-resolution remote sensing monitoring method. The method comprises the following steps: data acquisition and preprocessing: synchronously acquiring point cloud data, hyperspectral data and multi-temporal remote sensing data; feature engineering: constructing a multi-dimensional feature space, and optimizing the multi-dimensional feature space through feature selection; model training: on the basis of the optimized multi-dimensional feature space, performing filling, differential normalization processing and regularization improvement on a U-net + model through pixel values of the edges of mirror images in sequence for urban green land features, and establishing the urban green land-oriented and multi-dimensional feature-based U-net + model, so that the urban green land-oriented and multi-dimensional feature-based U-net + model is obtained; adding the multi-dimensional feature data of the training sample into the U-net + model for training, predicting the spatial distribution of the urban green land after the training is completed, obtaining a prediction result of the U-net + model, and combining a mixed loss function with an adaptive training strategy; and post-processing and result generation: applying adaptive CRFs to optimize a U-net + model prediction result, generating a high-resolution urban green land distribution map, and outputting a precision evaluation report.

Description

Technical Field

[0001] This invention relates to the field of urban green space monitoring technology, and in particular to a high-resolution remote sensing monitoring method and system for urban green spaces. Background Technology

[0002] With the acceleration of urbanization, urban green spaces are playing an increasingly important role as ecological barriers, recreational spaces for residents, and climate regulators. However, traditional ground surveys and low-resolution remote sensing monitoring have limitations such as low efficiency, long cycles, insufficient accuracy, and difficulty in dynamic updates, making it difficult to meet the needs of refined urban management, ecological benefit assessment, and planning decisions. By integrating high-resolution satellite, aerial remote sensing data, and multi-source geographic information data, and combining deep learning algorithms with Geographic Information System (GIS) spatial analysis technology, a fully automated method is constructed, encompassing data acquisition, preprocessing, information extraction, accuracy verification, and dynamic monitoring. This method enables the analysis of indicators such as urban green space type, distribution range, area, shape index, vegetation coverage, and fragmentation degree. The rapid, accurate, dynamic extraction and long-term tracking of targets breaks through the technical bottlenecks of traditional monitoring methods, improves the spatiotemporal resolution, information richness and update timeliness of urban green space monitoring, provides scientific data support for urban ecological restoration, green space system planning, heat island effect mitigation and biodiversity protection, promotes the refined management and sustainable development of urban green infrastructure, and promotes the interdisciplinary integration and application of remote sensing technology, artificial intelligence and geographic information science. It forms a scalable and replicable high-resolution remote sensing monitoring technology system for urban green spaces, serving the construction needs of smart cities, ecological cities and resilient cities, and ultimately achieving the efficient protection, rational utilization and dynamic optimization of urban green space resources, and promoting the development goals of modern cities.

[0003] In existing technologies, although traditional edge padding avoids information loss, zero-value filling may introduce false edge information, especially at the boundary between green and non-green areas, resulting in jagged artifacts in the classification results. Therefore, it is particularly important to propose a high-resolution remote sensing monitoring method and system for urban green areas. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a high-resolution remote sensing monitoring method and system for urban green spaces.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A high-resolution remote sensing monitoring method for urban green spaces includes the following steps: Data acquisition and preprocessing: Simultaneously acquire point cloud data, hyperspectral data, and multi-temporal remote sensing data, and perform geometric correction, atmospheric correction, and radiometric calibration; Feature engineering: Construct a multidimensional feature space and optimize the multidimensional feature space through feature selection. The feature selection is achieved by deploying the XGBoost feature importance evaluation module to automatically filter key features and add multi-temporal features calculated by sliding window to the phenological features. Model Training: Based on the optimized multidimensional feature space and targeting urban green space features, the U-net+ model is improved by sequentially filling the pixel values ​​of the mirrored image edges, performing differential normalization processing, and regularization. A U-net+ model targeting urban green space and based on multidimensional features is established. Then, the multidimensional feature data of the training samples are added to the U-net+ model for training. After training, the spatial distribution of urban green space is predicted, and the prediction results of the U-net+ model are obtained. The mixture loss function and adaptive training strategy are combined. Post-processing and result generation: Adaptive CRFs are applied to optimize the prediction results of the U-net+ model, generate a high-resolution urban green space distribution map, and output an accuracy assessment report.

[0006] The above plan further includes: Furthermore, the point cloud data is acquired through 3D laser scanning or UAV LiDAR, containing spatial coordinates (X,Y,Z) and reflection intensity information, used to describe the 3D structure of green space. Continuous spectral information in the 400-2500nm band is collected to identify the spectral characteristics of green space vegetation. Remote sensing images of different seasons are acquired to monitor the dynamic changes of green space. Digital surface models are generated from point cloud data to assist in extracting vegetation height. Hyperspectral data is used to calculate vegetation indices. Multimodal labels are generated by combining visual interpretation. Uncertain samples are manually verified.

[0007] Furthermore, in feature engineering, the multidimensional feature space includes vegetation features, spatial features, phenological features, and contrast features. The vegetation features include Normalized Difference Vegetation Index (NDVI), Red Edge Vegetation Index (RVI), Normalized Difference Water Index (NDWI), and point cloud height histogram. The point cloud height histogram statistically analyzes the height distribution of vegetation point clouds, reflecting the three-dimensional structure of green space. The spatial features include Normalized Digital Surface Model (nDSM) and hyperspectral texture features. The phenological features include seasonal phase difference features and sliding window temporal features. The contrast features include Adaptive Contrast (AC).

[0008] Furthermore, in feature engineering, the specific steps for feature selection using the XGBoost feature importance evaluation module are as follows: Train the XGBoost model by inputting a multidimensional feature space; The contribution of each feature to the model is evaluated using feature importance scores, which are calculated using the gain during tree splitting and expressed as follows: ,in, and For the first and second derivatives of the left subtree, and Let be the first and second derivatives of the right subtree. and For regularization parameters; Select features with high importance scores and remove redundant features.

[0009] Furthermore, in model training, a U-net+ model oriented towards urban green spaces and based on multi-dimensional features is established, including the following steps; Cross-modal feature fusion: concatenating the optimized multidimensional feature space into an input tensor; Feature dynamic enhancement: Reflection padding: In all convolutional layers of the U-net encoder, the edge is padded by mirroring the pixel values ​​of the image edges to maintain boundary continuity; Differential normalization: The encoder uses BN, and the decoder uses GN; Feature fusion layer: ASPP module embedding: Apertured spatial pyramid pooling (ASPP) is introduced into the bottleneck layer of U-net, using parallel dilated convolutions with dilation rates of [3,6,12] to simultaneously capture local details and global context; Attention gates: Attention gates are inserted into skip connections to perform weighted fusion of the optimized multidimensional feature space. Hybrid loss function: A combination of cross-entropy and Dice loss is used, and a weighted summation is used to balance class probabilities and small target recognition; The adaptive training strategy automatically adjusts the learning rate based on the F1 score on the validation set. When the F1 score stagnates, the learning rate is reduced. The pre-trained model is initialized on the ImageNet dataset. During fine-tuning, the lower convolutional layers are frozen, and the high-level feature extraction modules are trained first.

[0010] Furthermore, in post-processing and result generation, adaptive CRFs are applied to optimize the prediction results of the U-net+ model, including the following steps: Energy function construction: The probability map output by the U-net+ model is post-processed using a Conditional Random Field (CRF). The energy function is defined as follows: ,in, This is a unary term, taken from the class probability output by the U-net+ model, and is expressed as... ,in, For pixel i, it belongs to category The probability, As paired terms, based on pixel spatial neighborhood relationships, a Gaussian kernel function is used: ,in, For tag compatibility, For pixel position, For RGB intensity, To control the spatial and spectral similarity weights; Adaptive parameter adjustment: Dynamically adjusts the hyperparameters of CRFs based on prediction uncertainty, introducing prediction confidence. Adjust the weights of paired items , , ,in, Adaptive coefficients are determined through grid search or Bayesian optimization to find the optimal ones. .

[0011] Furthermore, in post-processing and result generation, the steps for generating a high-resolution urban green space distribution map are as follows: Upsampling: The low-resolution probability map output by the U-net+ model is restored to the original remote sensing image resolution through bicubic interpolation or a deep learning super-resolution model; Multi-source data fusion: Combining point cloud nDSM and hyperspectral texture features, the final distribution map is generated through weighted fusion; Color mapping: Green areas use green tones, and non-green areas use gray tones; Layer overlay: Overlay a green space distribution map with a satellite image base map, vector road or building layers to generate an interactive map; Vectorized output: Green space vector polygons are generated using a contour extraction algorithm.

[0012] A high-resolution remote sensing monitoring system for urban green spaces, used in a high-resolution remote sensing monitoring method for urban green spaces, includes: Data Acquisition and Preprocessing Layer: Integrated Multi-Source Sensor Collaborative Acquisition System: including LiDAR point cloud scanner, hyperspectral imager, multi-temporal remote sensing satellite and UAV aerial survey platform, to perform geometric correction, atmospheric correction and radiometric calibration on the collected data, and then construct a multi-source heterogeneous dataset; Feature Engineering Layer: Constructs a multi-dimensional feature space containing five dimensions of features: spectral, spatial, textural, phenological, and contrast. Through feature selection and temporal injection, the robustness and adaptability of the feature space are improved. Model building and training layer: Green space spatial distribution is predicted by using the improved U-net+ model. The improved U-net+ model’s ability to identify small targets and its overall robustness are improved by using a hybrid loss function and adaptive training strategy. Post-processing and result generation layer: Improve the spatial continuity and boundary accuracy of prediction results through CRF post-processing, generate urban green space distribution map, and output accuracy assessment report and visualization results.

[0013] The present invention has the following beneficial effects: This invention employs reflection filling, using the pixel values ​​of mirrored image edges to fill in the gaps, thus avoiding the introduction of zero values—information irrelevant to the image content. This operation better preserves edge continuity, providing the decoder with more realistic contextual information, thereby effectively reducing jagged artifacts in the predicted boundaries and generating more natural and accurate green area boundaries. This improvement requires only modifications to the underlying network operations, resulting in low cost but a significant improvement in edge quality. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the steps of a high-resolution remote sensing monitoring method for urban green spaces proposed in this invention. Figure 2 This is a system block diagram of a high-resolution remote sensing monitoring method for urban green spaces proposed in this invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Please see Figures 1-2 As shown, this invention is a high-resolution remote sensing monitoring method for urban green spaces, comprising the following steps: Data acquisition and preprocessing: Simultaneously acquire point cloud data, hyperspectral data, and multi-temporal remote sensing data, and perform geometric correction, atmospheric correction, and radiometric calibration.

[0017] Feature engineering: Calculate features such as NDVI, nDSM, and hyperspectral index to construct a multidimensional feature space, and optimize the multidimensional feature space through feature selection. The feature selection is achieved by deploying the XGBoost feature importance assessment module to automatically filter key features and add multi-temporal features calculated by sliding window to the phenological features to enhance the model's ability to capture seasonal dynamics.

[0018] Model Training: Based on the optimized multidimensional feature space and targeting urban green space features, the U-net+ model is improved by sequentially filling the pixel values ​​of the mirrored image edges, performing differential normalization processing, and regularization. A U-net+ model targeting urban green space and based on multidimensional features is established. Then, the multidimensional feature data of the training samples are added to the U-net+ model for training. After training, the spatial distribution of urban green space is predicted, and the prediction results of the U-net+ model are obtained. The mixture loss function and adaptive training strategy are combined. Post-processing and result generation: Adaptive CRFs are applied to optimize the prediction results of the U-net+ model, generate a high-resolution urban green space distribution map, and output an accuracy assessment report.

[0019] In one embodiment, the point cloud data is acquired through 3D laser scanning or UAV LiDAR, containing spatial coordinates (X, Y, Z) and reflectance intensity information, used to describe the 3D structure of green space. The hyperspectral data: continuous spectral information in the 400-2500nm band is collected, with a resolution typically reaching 5-10nm, used to identify the spectral characteristics of green space vegetation. The multi-temporal data: remote sensing images of different seasons are acquired to monitor the dynamic changes of green space. A digital surface model (DSM) is generated from the point cloud data to assist in extracting vegetation height. The hyperspectral data is used to calculate 10+ vegetation indices such as NDVI and NDWI. Combined with visual interpretation, multimodal labels are generated (the height, spectrum, and texture attributes are preserved when converting vector files to raster). Uncertain samples (such as the boundary between shadows and green space) are manually reviewed to reduce the subjectivity of visual interpretation and improve sample consistency.

[0020] In one embodiment, in feature engineering, the multidimensional feature space includes vegetation features, spatial features, phenological features, and contrast features. The vegetation features include Normalized Difference Vegetation Index (NDVI), Red Edge Vegetation Index (RVI), Normalized Difference Water Index (NDWI), and a point cloud height histogram. The point cloud height histogram statistically analyzes the height distribution of vegetation point clouds, reflecting the three-dimensional structure of green space. The spatial features include Normalized Digital Surface Model (nDSM) and hyperspectral texture features. The phenological features include seasonal phase difference features and sliding window temporal features. The contrast features include Adaptive Contrast (AC).

[0021] In one embodiment, the specific steps for feature selection using the XGBoost feature importance evaluation module in feature engineering are as follows: Train the XGBoost model by inputting a multidimensional feature space; The contribution of each feature to the model is evaluated using feature importance scores, which are calculated using the gain during tree splitting and expressed as follows: ,in, and For the first and second derivatives of the left subtree, and Let be the first and second derivatives of the right subtree. and For regularization parameters; Select features with high importance scores and remove redundant features.

[0022] In one embodiment, during model training, a U-net+ model oriented towards urban green spaces and based on multidimensional features is established, including the following steps; Cross-modal feature fusion: The optimized multi-dimensional feature space is concatenated into an input tensor (57-channel input). Feature dynamic enhancement: Reflection padding: In all convolutional layers of the U-net encoder, reflection padding maintains boundary continuity by mirroring the pixel values ​​of the image edges. For the input feature map x (dimensions N×C×H×W), reflection padding symmetrically replicates pixel values ​​at the boundaries. For example, when padding=1, one pixel is used to pad the left / right / top / bottom, and the padding value is a mirror image of the boundary pixel (e.g., the left boundary pixel). The left side is filled with (mirror image) Differential normalization: The encoder uses BN (Batch Normalization) (using Conv2D→BN→ReLU structure to enhance the stability of feature abstraction), and the decoder uses GN (Group Normalization, number of groups = 4) (using Conv2D→ReLU→Group Normalization (GN) structure, GN normalizes by grouping by channel (e.g., 4 groups) to adapt to small batch processing). Ablation experiments show that GN improves the small target recognition rate by 12% in small batch (batch=4). Feature fusion layer: ASPP module embedding: Apertured spatial pyramid pooling (ASPP) is introduced into the bottleneck layer of U-net, using parallel dilated convolutions with dilation rates of [3,6,12] to simultaneously capture local details and global context; Attention gate: An attention gate is inserted into the skip connections to perform weighted fusion on the optimized multidimensional feature space (dynamically adjusting the weights of different features (such as phenological features and spatial features) to adapt to the monitoring needs of different seasons, and automatically learning to suppress irrelevant background features (such as building shadows) and enhance the salience of green space targets (such as isolated trees)). Hybrid loss function: A combination of cross-entropy (CE) and Dice loss is used, and the class probability and small target recognition are balanced by weighted summation; The adaptive training strategy automatically adjusts the learning rate based on the F1 score on the validation set (such as the ReduceLROnPlateau strategy). When the F1 score stagnates, the learning rate is reduced (factor=0.1, patience=10). The pre-trained model is initialized on the ImageNet dataset. During fine-tuning, the lower convolutional layers (such as the first 10 layers) are frozen, and the higher-level feature extraction modules (such as the bottleneck layer) are trained first.

[0023] Input the optimized feature set (57 channels) into the U-net+ model; A green space probability map is generated using an encoder-decoder structure; Optimize network weights using a hybrid loss function; Output the binarized green space segmentation result (threshold=0.5).

[0024] In one embodiment, during post-processing and result generation, adaptive CRFs are applied to optimize the prediction results of the U-net+ model, including the following steps: Energy function construction: Conditional Random Field (CRF) is used to post-process the probability map output by the U-net+ model to optimize boundary details. The energy function is defined as follows: ,in, This is a unary term, taken from the class probability (e.g., green area / non-green area) output by the U-net+ model, and is represented as... ,in, For pixel i, it belongs to category The probability, As paired terms, based on pixel spatial neighborhood relationships, a Gaussian kernel function is used: ,in, For label compatibility (such as the Potts model). For pixel position, For RGB intensity, To control the spatial and spectral similarity weights; Adaptive parameter adjustment: Dynamically adjusts the hyperparameters of CRFs based on prediction uncertainty, introducing prediction confidence. Adjust the weights of paired items , , ,in, An adaptive coefficient (e.g., 0.5) is used to enhance spatial smoothness constraints in high-confidence regions and preserve details in low-confidence regions. The optimal solution is determined through grid search or Bayesian optimization. .

[0025] In one embodiment, the steps for generating a high-resolution urban green space distribution map during post-processing and result generation are as follows: Upsampling: The low-resolution probability map output by the U-net+ model is restored to the original remote sensing image resolution through bicubic interpolation or a deep learning super-resolution model; Multi-source data fusion: Combining point cloud nDSM and hyperspectral texture features, the final distribution map is generated through weighted fusion; Color mapping: Green areas use green tones (e.g., RGB=[0,255,0]), and non-green areas use gray tones (e.g., RGB=[128,128,128]). Layer overlay: Overlay a green space distribution map with a satellite image base map, vector road or building layers to generate an interactive map; Vectorized output: Green space vector polygons are generated using contour extraction algorithms (such as Canny edge detection), which can be used directly in GIS software.

[0026] A high-resolution remote sensing monitoring system for urban green spaces, used in a high-resolution remote sensing monitoring method for urban green spaces, is characterized by comprising: Data Acquisition and Preprocessing Layer: Integrating a multi-source sensor collaborative acquisition system, including a LiDAR point cloud scanner (acquiring 3D point cloud data), a hyperspectral imager (acquiring continuous spectrum from 400-2500nm), multi-temporal remote sensing satellites (such as WorldView-3 and Sentinel-2), and an UAV aerial survey platform (acquiring RGB imagery and winter phenological imagery). The collected data undergoes geometric correction, atmospheric correction, and radiometric calibration to construct a multi-source heterogeneous dataset, ensuring a unified benchmark for point clouds, hyperspectral data, and remote sensing images in terms of spatial resolution (e.g., 0.5m), temporal dimension (multi-seasonal / multi-temporal), and spectral dimension (hyperspectral bands). Feature Engineering Layer: Constructs a multi-dimensional feature space containing five dimensions of features: spectral, spatial, textural, phenological, and contrast, to achieve in-depth mining and dynamic optimization of feature information. Through feature selection and temporal injection, it enhances the robustness and adaptability of the feature space and provides high-information input for model training. Model building and training layer: The improved U-net+ model is used to predict the spatial distribution of green space, improve edge accuracy and detail preservation, enhance the ability to identify multi-scale targets, and improve the ability to identify small targets and the overall robustness of the improved U-net+ model by using a hybrid loss function and adaptive training strategy. Post-processing and result generation layer: Improve the spatial continuity and boundary accuracy of prediction results through CRF post-processing, generate urban green space distribution map, ensure the reliability and generalization ability of results through multi-source collaborative verification, and output accuracy assessment report and visualization results.

[0027] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring urban green space by high-resolution remote sensing, characterized in that, The method comprises the following steps: Data acquisition and preprocessing: synchronously acquire point cloud data, hyperspectral data, and multi-temporal remote sensing data, and perform geometric correction, atmospheric correction, and radiation calibration; Feature engineering: construct a multi-dimensional feature space, and optimize the multi-dimensional feature space through feature selection, which automatically screens key features by deploying an XGBoost feature importance evaluation module, and adds multi-temporal features calculated by a sliding window to the phenological features; Model training: based on the optimized multi-dimensional feature space, and facing the features of urban green land, sequentially fill, differentially normalize, and regularize the U-net+ model by using the pixel values of the edges of mirror images, establish the U-net+ model facing the urban green land and based on the multi-dimensional features, add the multi-dimensional feature data of the training samples to the U-net+ model for training, and after the training is completed, predict the spatial distribution of the urban green land, obtain the prediction result of the U-net+ model, and combine a hybrid loss function and an adaptive training strategy; Post-processing and result generation: apply adaptive CRFs to optimize the prediction result of the U-net+ model, generate a high-resolution urban green land distribution map, and output an accuracy evaluation report. 2.The urban green space high-resolution remote sensing monitoring method according to claim 1, characterized in that, The point cloud data is acquired by three-dimensional laser scanning or unmanned aerial vehicle LiDAR, contains spatial coordinates (X, Y, Z) and reflection intensity information, is used to describe the three-dimensional structure of the green land, acquires continuous spectral information in the 400-2500 nm waveband range, is used to identify the spectral features of the green land vegetation, acquires remote sensing images of different seasons, is used to monitor the dynamic changes of the green land, generates a digital surface model by using the point cloud data to assist in extracting the vegetation height, calculates the vegetation index by using the hyperspectral data, and generates multi-modal labels by combining visual interpretation, and manually reviews the uncertain samples. 3.The urban green space high-resolution remote sensing monitoring method according to claim 1, characterized in that, In the feature engineering, the multi-dimensional feature space comprises vegetation features, spatial features, phenological features, and contrast features, the vegetation features comprise a normalized difference vegetation index NDVI, a red edge vegetation index RVI, a normalized difference water index NDWI, and a point cloud height histogram, the point cloud height histogram statistically reflects the height distribution of the vegetation point cloud and reflects the three-dimensional structure of the green land, the spatial features comprise a normalized digital surface model nDSM and a hyperspectral texture feature, the phenological features comprise a seasonal phase difference feature and a sliding window time series feature, and the contrast feature comprises an adaptive contrast AC. 4.The urban green space high-resolution remote sensing monitoring method according to claim 3, characterized in that, In the feature engineering, the specific steps of performing feature selection by using the XGBoost feature importance evaluation module are as follows: Train an XGBoost model, and input the multi-dimensional feature space; Each feature's contribution to the model is assessed by a feature importance score, which is computed by the gain at tree split and is denoted as wherein, with the first and second derivatives for the left child, with the first and second derivatives for the right child, with is a regularization parameter; Screen the features with high importance scores, and remove redundant features. 5.The urban green space high-resolution remote sensing monitoring method according to claim 1, characterized in that, In the model training, the U-net+ model facing the urban green land and based on the multi-dimensional features comprises the following steps: Cross-modal feature fusion: splice the optimized multi-dimensional feature space into an input tensor; Feature dynamic enhancement: Reflection padding: fill the pixel values of the edges of mirror images in all convolution layers of the U-net encoder to maintain boundary continuity; Differential normalization: the encoder adopts BN, and the decoder adopts GN; Feature fusion layer: ASPP module embedding: Introducing the spatial pyramid pooling ASPP in the bottleneck layer of U-net, using parallel dilated convolution with dilation rates [3, 6, 12] to capture local details and global context simultaneously; Attention gate: Inserting an attention gate in the skip connection to weight and fuse the optimized multi-dimensional feature space; Hybrid loss function: Using a combination of cross-entropy and Dice loss to balance class probability and small target recognition through weighted summation; Adaptive training strategy based on validation set F1 score to automatically adjust the learning rate, reduce the learning rate when F1 score stagnates, pre-train the model on ImageNet dataset, freeze the bottom convolutional layer during fine-tuning, and prioritize training high-level feature extraction modules. 6.The urban green space high-resolution remote sensing monitoring method according to claim 5, characterized in that, In post-processing and result generation, adaptive CRFs are applied to optimize the U-net+ model prediction results, including the following steps: Energy function construction: the conditional random field (CRF) is used to post-process the probability map output by the U-net+ model, and the energy function is defined as: wherein, is a unary term, taken from the class probability output by the U-net+ model, denoted as wherein, is the probability of pixel i belonging to class . is a pairwise term, based on the spatial neighborhood relationship of pixels, using a Gaussian kernel function: wherein, is the label compatibility, is the pixel position, is the RGB intensity, is the control space and spectral similarity weight; Adaptive parameter tuning: dynamically tune CRFs hyperparameters according to prediction uncertainty, introduce prediction confidence , adjust pair-wise term weights , , where, Adaptive coefficients, determine optimal by grid search or Bayesian optimization .

7. The urban green space high-resolution remote sensing monitoring method according to claim 6, characterized in that, In post-processing and result generation, the generation steps of high-resolution urban green space distribution map: Upsampling processing: Recover the low-resolution probability map output by the U-net+ model to the original remote sensing image resolution through bicubic interpolation or deep learning super-resolution model; Multi-source data fusion: Combine point cloud nDSM and hyperspectral texture features to generate the final distribution map through weighted fusion; Color mapping: Green areas are colored in green, and non-green areas are colored in gray; Layer superposition: Superimpose the green space distribution map with satellite image base map, vector road or building layer to generate an interactive map; Vector output: Generate green space vector polygons through contour extraction algorithm. 8.The urban green space high-resolution remote sensing monitoring system used in the method of claim 1, characterized in that, Including: Data acquisition and preprocessing layer: Integrate multi-source sensor cooperative acquisition system: including LiDAR point cloud scanner, hyperspectral imager, multi-temporal remote sensing satellite and unmanned aerial vehicle surveying and mapping platform, perform geometric correction, atmospheric correction and radiometric calibration on the collected data, and then build a multi-source heterogeneous data set; Feature engineering layer: Build a multi-dimensional feature space containing five-dimensional features of spectrum, space, texture, phenology and contrast, improve the robustness and adaptability of the feature space through feature selection and time series injection; Model construction and training layer: Use the improved U-net+ model to predict green space distribution, use hybrid loss function and adaptive training strategy to improve the recognition ability of small targets and overall robustness of the improved U-net+ model; Post-processing and result generation layer: Improve the spatial continuity and boundary accuracy of the prediction results through CRFs post-processing, generate urban green space distribution map, output precision evaluation report and visualized results.

Images