Urban street four-season green vision rate estimation method based on improved local regression model

Abstract

The invention discloses a city street four-season green vision rate estimation method based on an improved local regression model, and the method comprises the steps: generating street scene sampling points according to road network data, collecting street scene images, carrying out the semantic segmentation of the street scene images through a deep learning model, and extracting a vegetation region in the street scene images. The method comprises the following steps of: calculating the green vision rate of a street scene sampling point in each season, extracting a multi-season normalized vegetation index of the sampling point through remote sensing data, establishing a regression model for the normalized vegetation index and the green vision rate of a specific season by utilizing a local regression model fused with a season weight, and estimating the missing green vision rate by utilizing the normalized vegetation index. According to the method, annual street view data are accurately collected and processed, a remote sensing technology, a deep learning technology and an image segmentation technology are combined, and an effective means is provided for scientific evaluation of urban street greening in different seasons. Meanwhile, the method provided by the invention is high in automation degree, and is suitable for practical application of large-scale urban landscaping evaluation.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing and geographic information technology, specifically to a method for estimating the seasonal green visibility rate of urban streets using a local regression model that incorporates seasonal weights. Background Technology

[0002] Urban greening generally refers to urban spaces occupied by green plants, including both natural and man-made green areas. Urban greening has significant ecological and economic value, playing a crucial role in reducing noise, lowering carbon emissions, mitigating the urban heat island effect, improving air quality, beautifying the environment, and enhancing the added value of properties. Furthermore, urban greening can promote the physical and mental health of residents, including encouraging walking, increasing physical activity, improving obesity rates, and alleviating depression. Urban greening is a vital factor in building attractive, healthy, and sustainable cities and communities.

[0003] In recent years, street view data has become a commonly used data source for monitoring urban street greening. Unlike satellite imagery, which observes the Earth from a bird's-eye view, street view images are viewed from a human-eye horizontal perspective, offering a human-centered viewpoint. Street view data is provided by commercial map providers (such as Google and Baidu). Street view images can display features such as the built environment and trees along streets. Green View Index (GVI) can be extracted from street view images, representing the proportion of perceptible green elements in the image. Current research has adopted GVI as a proxy indicator for street greening.

[0004] The invention patent application CN117036964 A (A method, system, device, and storage medium for determining the level of urban road greening) discloses a method for determining the level of urban road greening. This method involves: acquiring road network data and land cover data of the target city; setting multiple street view sampling points in the target city based on the road network data; collecting multiple street view images at each sampling point; using a land cover classification model to perform semantic segmentation on each street view image to obtain the classification results of pixels corresponding to each land cover in each street view image; and determining the level of road greening in the target city based on the green visibility rate and multi-weighted green space index of all street view sampling points, thus improving the accuracy of determining the level of urban road greening. However, it still has the following shortcomings.

[0005] 1. Ignoring temporal variations: This method calculates green view rate based on street view images at fixed times, without considering the dynamic impact of factors such as season and time period on the degree of greening. This may lead to an overestimation or underestimation of the greening level in different seasons.

[0006] 2. Insufficient response to spatiotemporal changes in greening: Because it only uses image data from a single time period, this method is difficult to dynamically and continuously reflect the seasonal changes in urban street greening.

[0007] 3. Missing street view image data: Because collecting street view image data requires a lot of manpower, material resources and financial resources, some street view sampling points may lack street view image data for specific seasons, making it impossible to calculate the corresponding GVI from the street view image data. Summary of the Invention

[0008] The purpose of this invention is to provide a method for estimating the seasonal green view rate of urban streets based on an improved local regression model. This method extracts multi-seasonal normalized vegetation index (NVR) from sampling points using remote sensing data. A local regression model incorporating seasonal weights is then used to establish a regression model between the NVR and the green view rate for a specific season. The missing green view rate is then estimated using the NVR. The completed green view rate is used to accurately assess the greening status of urban streets and can account for the impact of seasonal variations.

[0009] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0010] A method for estimating the seasonal green view rate of urban streets based on an improved local regression model includes the following steps:

[0011] S1. Collect road network data, extract road centerlines, and generate street view sampling points at preset distances;

[0012] S2. Collect street view images of each street view sampling point covering multiple time phases of the four seasons within a preset time period, and count the street view sampling points with missing street view images and their corresponding time phases;

[0013] S3. Use a deep learning model to perform semantic segmentation on the collected street view images and extract the vegetation areas in the street view images;

[0014] S4. Calculate the green visibility rate of each street view sampling point in the acquired street view image at each time phase;

[0015] S5. Calculate the green visibility rate of each street view sampling point in the acquired street view image under each season;

[0016] S6. Collect multi-temporal remote sensing image data for the same preset time period as in step S2, and calculate the normalized vegetation index for each street view sampling point in each season.

[0017] S7. By using a local regression model that incorporates seasonal weights, the green view rate of a street view sampling point in a certain season is estimated using the normalized vegetation index of the street view sampling point in each season, which cannot be calculated due to the lack of street view images.

[0018] In step S6, by specifying the spatial and temporal range of data processing and performing cloud cover filtering, multi-temporal remote sensing image data is obtained, and the normalized vegetation index (NDI) for each season is calculated. Then, a buffer zone with a certain spatial range is established for the street view sampling points, and the NDI for each season is calculated.

[0019] In step S7, for street view sampling points lacking GVI, a local regression model with seasonal weights is established using NDVI and GVI of the same season, with NDVI as the predictor variable and GVI as the target variable. The GVI is estimated using NDVI to obtain the missing GVI for each street view sampling point. For street view sampling points lacking GVI in a particular season, if there are no GVIs for other seasons, a local weighted regression model is used directly. If there are GVIs for other seasons (GVIs of 1-3 seasons), the model is combined with the GVIs from other seasons and calculated using a local regression model with seasonal weights. This model uses a Gaussian function to assign weights to sample points within the same season, assigns larger seasonal weights to samples from different seasons at the same spatial location, and fuses these two types of weights. Finally, the regression coefficients are fitted using weighted least squares to estimate the GVI, thus completing the missing GVI samples.

[0020] As a further improved technical solution of the present invention, in step S4, the green visibility rate of each street view sampling point in the acquired street view image is calculated in each time phase using the following formula.

[0021]

[0022] Where GVI is the green visibility rate of a street view sampling point at a certain time phase, i (=1,2,3,4) represents the street view images of the sampling point in four directions at that time phase, and Area Green Area is the number of pixels in a street view image from a certain direction that are identified as vegetation areas. All It represents the total number of pixels in a street view image in a certain direction.

[0023] As a further improved technical solution of the present invention, in step S5, the green visibility rate of each street view sampling point in the acquired street view image is calculated in each season using the following formula.

[0024]

[0025] Among them, GVI avg It represents the green view rate (GVI1+GVI2+...GVI) at a street view sampling point during a specific season. n It is the sum of the green visibility of the street view sampling point in each time phase during the season, and n is the number of time phases of the street view sampling point in the season.

[0026] As a further improved technical solution of the present invention, step S6 involves collecting multi-temporal remote sensing image data for the same preset time period as in step S2, and calculating the normalized vegetation index for each street view sampling point in each season, specifically including:

[0027] The specified spatial range includes all street view sampling points, and the time range is within a preset time period and covers multiple time phases of the four seasons, acquiring remote sensing images of multiple time phases within the specified spatial range;

[0028] The normalized vegetation index for each time phase within a specified spatial range is calculated using the following formula:

[0029]

[0030] Wherein, NDVI is the normalized vegetation index at a certain time within a specified spatial range, NIR is the reflectance of the near-infrared band in the remote sensing image at that time, and R is the reflectance of the red band in the remote sensing image at that time.

[0031] The normalized vegetation index for each season within a specified spatial area is calculated using the following formula:

[0032]

[0033] Among them, NDVI avg It is the average normalized vegetation index (NDVI) within a specified spatial area under a certain season, NDVI1+NDVI2+...NDVI n It is the sum of the normalized vegetation indices of the specified spatial range in each time phase under the season, and n is the number of time phases of the specified spatial range under the season;

[0034] Establish a buffer zone with a preset distance for all street view sampling points, and extract and count the NDVI values ​​within the buffer zone. avg The normalized vegetation index of each street view sampling point in each season was obtained.

[0035] As a further improved technical solution of the present invention, step S7 uses a local regression model with seasonal weights to estimate the green view rate of a street view sampling point in a certain season, which cannot be calculated due to the lack of street view images, by using the normalized vegetation index of the street view sampling point in each season. Specifically, this includes:

[0036] S71. The green view rate and normalized vegetation index of the street view sampling points in each season are divided into 4 subsets according to the season. The green view rate and normalized vegetation index of a certain street view sampling point in a certain season are taken as a sample point in the subset.

[0037] S72. Use a local regression model to fit a model to sample points in the same season;

[0038] S73. Integrate seasonal weights into the local regression model and fit the model again to estimate the green view rate of a street view sampling point in a certain season, which cannot be calculated due to the lack of street view images.

[0039] As a further improvement to the present invention, step S72 uses a local regression model to fit a model to sample points in the same season, specifically including:

[0040] A Gaussian weighting function is used to assign local weights to sample points that are in the same season but in different spatial locations as the sample points whose green visibility rate is to be estimated, and then the model is fitted by weighted least squares method.

[0041] As a further improvement to the present invention, step S73 involves fusing seasonal weights into the local regression model to refit the model and estimate the green view rate of a street view sampling point in a certain season, which cannot be calculated due to the lack of street view images. Specifically, this includes:

[0042] Sample points in other seasons that are in the same spatial location as the sample point whose green view rate is to be estimated and contain seasonal green view rate are used as auxiliary points. The seasonal weights of the auxiliary points and the sample points whose green view rate is to be estimated are calculated, and then fused with the local weights. Finally, a local regression model with fused seasonal weights is fitted. The green view rate of a certain street view sampling point in a certain season that cannot be calculated due to the lack of street view images is estimated using the local regression model with fused seasonal weights.

[0043] The seasonal weights of auxiliary points and sample points for estimating green visibility are calculated using the following formula:

[0044] w season (p)=λexp(-dist season (t,t0))

[0045] Among them, w season (p) represents the seasonal weights of the auxiliary points and the sample points for estimating the green visibility rate, λ is a hyperparameter used to amplify the seasonal weights, and dist season (t,t0) is the seasonal distance between the auxiliary point and the green visibility sample point to be estimated;

[0046] The regression coefficients are calculated using the following formula:

[0047]

[0048] in, Here, X is the regression coefficient, W is the design matrix, and y is the dependent variable vector;

[0049] Finally, the green view rate of the sample points to be estimated is calculated using the following formula:

[0050]

[0051] in, Let X be the green view rate of the sample points to be estimated, and let X be the design matrix. is the regression coefficient.

[0052] As a further improvement of the present invention, after collecting road network data in step S1, the method also includes road screening and simplification, then extracting the road centerline and generating street view sampling points at a preset distance.

[0053] As a further improvement of the present invention, step S2 includes stitching the street view images into a panoramic image and then acquiring the image.

[0054] As a further improvement of the present invention, the deep learning model in step S3 is the deep learning semantic segmentation model Mask2Former pre-trained on the Cityscapes dataset.

[0055] Compared with the prior art, the technical advantages of the present invention are as follows:

[0056] Introducing multi-temporal data: This invention constructs a multi-temporal dataset by collecting street view images of target city streets at different times. It considers not only the green view rate at a single moment but also seasonal variations throughout the year. This method overcomes the limitations of fixed-temporal calculations of green view rate in existing technologies, providing more dynamic and comprehensive greening assessment results.

[0057] Spatiotemporal dynamic monitoring capability: This invention can dynamically monitor the greening status of urban streets and reflect the spatiotemporal change trend of greening. It not only provides data support for the assessment of the current status of urban greening, but also provides more continuous and accurate information for long-term urban planning and greening management.

[0058] Completing Missing Green Visibility (GVI) Samples: Due to the significant human, material, and financial resources required for street view image data acquisition, some street view sampling points may lack street view image data for specific seasons. Given the high dynamic monitoring capabilities of remote sensing imagery, all sampling points typically possess multi-temporal remote sensing imagery. Both the Normalized Difference Vegetation Index (NDVI) extracted from remote sensing imagery and GVI can reflect vegetation growth, and previous studies have shown a strong correlation between the two. For these samples lacking GVI, a seasonally weighted local regression model is constructed to estimate the missing GVI using NDVI. The completed GVI can more accurately assess the greening status of urban streets. Attached Figure Description

[0059] Figure 1 This is a flowchart illustrating a method for estimating the year-round green visibility rate of urban streets based on an improved local regression model, according to a specific embodiment of the present invention.

[0060] Figure 2 This is a schematic diagram of the principle of constructing a local regression model with fused seasonal weights using auxiliary points in Embodiment 1 of the present invention;

[0061] Figure 3 This is the local regression model in Embodiment 1 of the present invention;

[0062] Figure 4 This is the local regression model with fused seasonal weights in Embodiment 1 of the present invention. Detailed Implementation

[0063] The present invention will now be described in detail with reference to the specific embodiments shown in the accompanying drawings. However, these embodiments do not limit the present invention, and any modifications to the methods or functions made by those skilled in the art based on these embodiments are included within the scope of protection of the present invention.

[0064] In this invention, GVI (Green View Index) represents the green view rate, and NDVI (Normalized Difference Vegetation Index) represents the normalized vegetation index.

[0065] Please see Figure 1 A method for estimating the seasonal green view rate of urban streets based on an improved local regression model includes the following steps:

[0066] S1. Collect road network data, extract road centerlines, and generate street view sampling points at preset distances;

[0067] S2. Collect street view images of each street view sampling point covering multiple time phases of the four seasons within a preset time period, and count the street view sampling points with missing street view images and their corresponding time phases;

[0068] S3. Use a deep learning model to perform semantic segmentation on the collected street view images and extract the vegetation areas in the street view images;

[0069] S4. Calculate the green visibility rate of each street view sampling point in the acquired street view image at each time phase;

[0070] S5. Calculate the green visibility rate of each street view sampling point in the acquired street view image under each season;

[0071] S6. Collect multi-temporal remote sensing image data for the same preset time period as in step S2, and calculate the normalized vegetation index for each street view sampling point in each season.

[0072] S7. By using a local regression model that incorporates seasonal weights, the green view rate of a street view sampling point in a certain season is estimated using the normalized vegetation index of the street view sampling point in each season, which cannot be calculated due to the lack of street view images.

[0073] It should be noted that "phase" in this invention is defined as the growth cycle of a plant, and there may be multiple phases in a season.

[0074] Furthermore, in step S4, the green visibility rate of each street view sampling point in the acquired street view image at each time phase is calculated using the following formula 1;

[0075]

[0076] Where GVI is the green visibility rate of a street view sampling point at a certain time phase, i (=1,2,3,4) represents the street view images of the sampling point in four directions at that time phase, and Area Green Area is the number of pixels in a street view image from a certain direction that are identified as vegetation areas. All It represents the total number of pixels in a street view image in a certain direction.

[0077] The green visibility ratio (GVI) of a street view sampling point at a given time phase can be obtained by calculating the percentage of the sum of the number of pixels identified as vegetation areas in the four street view images collected at a given time phase, relative to the sum of the total number of pixels in the four street view images. Repeating this calculation formula iteratively yields the GVI for each street view sampling point at each time phase.

[0078] Furthermore, in step S5, the green visibility rate of each street view sampling point in the acquired street view image under each season is calculated using the following formula 2;

[0079]

[0080] Among them, GVI avg It represents the green view rate (GVI1+GVI2+...GVI) at a street view sampling point during a specific season. n It is the sum of the green visibility of the street view sampling point in each time phase during the season, and n is the number of time phases of the street view sampling point in the season.

[0081] For example, street view sampling point 1 has 6 sampling phases in spring, with corresponding green view rates GVI1=50%, GVI2=45%, GVI3=45%, GVI4=50%, GVI5=46%, and GVI6=46%. Therefore, the green view rate GVI of street view sampling point 1 in spring is... spring = (50% + 45% + 45% + 50% + 46% + 46%) / 6 = 47%.

[0082] Furthermore, in step S6, multi-temporal remote sensing image data for the same preset time period as in step S2 are collected, and the normalized vegetation index for each street view sampling point in each season is calculated, specifically including:

[0083] The specified spatial range includes all street view sampling points, and the time range is within a preset time period and covers multiple time phases of the four seasons, acquiring remote sensing images of multiple time phases within the specified spatial range;

[0084] The normalized vegetation index for each time phase within a specified spatial range is calculated using the following formula 3:

[0085]

[0086] Wherein, NDVI is the normalized vegetation index at a certain time within a specified spatial range, NIR is the reflectance of the near-infrared band in the remote sensing image at that time, and R is the reflectance of the red band in the remote sensing image at that time.

[0087] The normalized vegetation index for each season within a specified spatial range is calculated using the following formula 4:

[0088]

[0089] Among them, NDVI avg It is the average normalized vegetation index (NDVI) within a specified spatial area under a certain season, NDVI1+NDVI2+...NDVI n It is the sum of the normalized vegetation indices of the specified spatial range in each time phase under the season, and n is the number of time phases of the specified spatial range under the season;

[0090] Establish a buffer zone with a preset distance for all street view sampling points, and extract and count the NDVI values ​​within the buffer zone. avg The normalized vegetation index of each street view sampling point in each season was obtained.

[0091] Furthermore, in step S7, a local regression model incorporating seasonal weights is used to estimate the green view rate of a street view sampling point in a given season, which cannot be calculated due to the lack of street view images. This specifically includes:

[0092] S71. The green view rate and normalized vegetation index of the street view sampling points in each season are divided into 4 subsets according to the season. The green view rate and normalized vegetation index of a certain street view sampling point in a certain season are taken as a sample point in the subset.

[0093] S72. Use a local regression model to fit a model to sample points in the same season;

[0094] S73. Integrate seasonal weights into the local regression model and fit the model again to estimate the green view rate of a street view sampling point in a certain season, which cannot be calculated due to the lack of street view images.

[0095] Furthermore, step S72 uses a local regression model to fit a model to the sample points of the same season, specifically including:

[0096] A Gaussian weighting function is used to assign local weights to sample points that are in the same season but in different spatial locations as the sample points whose green visibility rate is to be estimated, and then the model is fitted by weighted least squares method.

[0097] Furthermore, in step S73, seasonal weights are incorporated into the local regression model to refit the model and estimate the green view rate of a street view sampling point in a certain season, which cannot be calculated due to the lack of street view images. Specifically, this includes:

[0098] Sample points in other seasons that are in the same spatial location as the sample point whose green view rate is to be estimated and contain seasonal green view rate are used as auxiliary points. The seasonal weights of the auxiliary points and the sample points whose green view rate is to be estimated are calculated, and then fused with the local weights. Finally, a local regression model with fused seasonal weights is fitted. The green view rate of a certain street view sampling point in a certain season that cannot be calculated due to the lack of street view images is estimated using the local regression model with fused seasonal weights.

[0099] The seasonal weights of the auxiliary points and the sample points for estimating the green visibility rate are calculated using the following formula 6:

[0100] w season (p)=λexp(-dist season (t,t0)) (6)

[0101] Among them, w season (p) represents the seasonal weights of the auxiliary points and the sample points for estimating the green visibility rate, λ is a hyperparameter used to amplify the seasonal weights, and dist season (t,t0) is the seasonal distance between the auxiliary point and the green visibility sample point to be estimated;

[0102] The regression coefficients are calculated using the following formula 9:

[0103]

[0104] in, Here, X is the regression coefficient, W is the design matrix, and y is the dependent variable vector;

[0105] Finally, the green view rate of the sample points to be estimated is calculated using the following formula 10:

[0106]

[0107] in, Let X be the green view rate of the sample points to be estimated, and let X be the design matrix. is the regression coefficient.

[0108] Furthermore, after collecting road network data in step S1, the process also includes road filtering and simplification, extracting road centerlines, and generating street view sampling points at preset distances.

[0109] Furthermore, step S2 includes stitching the street view images into a panoramic image before acquiring the data.

[0110] Furthermore, the deep learning model in step S3 is Mask2Former, a deep learning semantic segmentation model pre-trained on the Cityscapes dataset.

[0111] This invention provides an effective means for the scientific evaluation of urban street greening in different seasons by accurately collecting and processing street view data throughout the year, combining remote sensing technology, deep learning technology, and image segmentation technology. Furthermore, the system provided by this invention has a high degree of automation and is suitable for large-scale urban greening assessment applications.

[0112] Example 1

[0113] This embodiment uses the greening assessment of Nanjing City as an example to describe the detailed process of implementing the present invention.

[0114] In this embodiment, the street view images are sourced from Baidu's static image API, the road network data is sourced from the OSM website, and the remote sensing image data is sourced from Sentinel-2 data. The specific implementation steps are as follows:

[0115] 1. Data preparation and preprocessing

[0116] OSM road network data was downloaded from the Open Street Map website, and the data source was filtered by attribute; primary, secondary, tertiary, residential, living street, and trunk were selected as the main data sources.

[0117] The filtered data sources are buffered and rasterized, and then converted into a vectorized simplified road network. Multi-lane roads are merged into single lanes, ensuring that all roads are single-line and broken at intersections.

[0118] In ArcGIS (which is the existing geographic information system software), use the point generation tool along the route to generate street view sampling points every 20 meters along the road.

[0119] 2. Capture multi-temporal street view images for each street view sampling point.

[0120] Using a Python script, the latitude and longitude of all street view sampling points were crawled, and the Baidu coordinate conversion API was called to convert the wgs84ll coordinate system (GPS coordinates) to the bd09ll coordinate system (Baidu coordinates).

[0121] Using Baidu Maps' "Time Machine" feature, the URL parameters sent via standard HTTP requests are defined through the Baidu Maps API. The converted coordinate system is input to obtain the panoid (panoramic image ID) of the most recent time point at the street view sampling point. The panoid can uniquely determine the timeline of the street view image at the street view sampling point. Based on the timeline, the panoid and related information of all time points between 2017 and 2022 (preset time period) of the street view image are obtained.

[0122] Using web crawling technology to access Baidu's static image API, the panoid was used with parameters such as the required crawling time and the set collection direction. The images were collected from the street direction by rotating clockwise by 0°, 90°, 180° and 270° (i.e., collecting images in the four directions of north, east, south and west). The street view images collected in the four directions were stitched together to obtain a panoramic image with the human-centered viewpoint. During crawling, the street view image was obtained for each street view sampling point based on its latitude, longitude and road orientation angle. The main street view image acquisition parameter settings are shown in Table 1.

[0123] Table 1

[0124]

[0125] 3. Vegetation region extraction based on deep learning

[0126] The Mask2Former deep learning semantic segmentation model, pre-trained on the Cityscapes dataset (a large-scale dataset for semantic understanding and object detection, specifically designed for urban scene analysis), performs well in urban scenes (such as the Cityscapes dataset) and is suitable for semantic segmentation of street view images. The Mask2Former model is loaded into the PyTorch environment to segment the acquired street view images, identifying various objects such as trees, sky, roads, and buildings.

[0127] The vegetation component obtained by semantic segmentation is used for subsequent calculations. For multi-temporal street view images, the green visibility rate (GVI) of each street view sampling point in each temporal phase is calculated using the above formula 1.

[0128] After obtaining the green view rate (GVI) of each street view sampling point in each time phase, since the preset sampling period (2017-2022) for street view image data is relatively long, there may be multiple street view images in a season. Therefore, the average green view rate (GVI) for each season is calculated using the mean method according to Formula 2 above. avg .

[0129] Taking Nanjing as an example, due to the long winter and summer seasons and the short spring and autumn seasons, spring is defined as April-May, summer as June-September, autumn as October-November, and winter as December-March of the following year. For instance, for a certain street view sampling point, two street view images were collected in April 2018 and May 2020, and the corresponding green view rate was calculated. Since both images belong to spring, the green view rate for that point in spring is the average of the green view rates of these two images. This method is used to calculate the green view rate for all street view sampling points in all four seasons. However, because street view image acquisition consumes a lot of resources, street view images for some seasons are missing, resulting in some street view sampling points lacking green view rates for a particular season.

[0130] This invention innovatively utilizes a local regression model that incorporates seasonal weights, using NVDI as the independent variable and GVI as the dependent variable, to estimate GVI through NDDI. The model integrates samples from other sampling points within the same season and samples from the same sampling point in different seasons. Different weights are assigned to these two types of samples using Gaussian weighting and seasonal weighting functions, respectively. A hyperparameter is set for the seasonal weight to amplify the weight of samples from the same sampling point. Finally, regression coefficients are fitted using weighted least squares to estimate the GVI for sampling points lacking it.

[0131] 4. Collect multi-temporal remote sensing image data within the same preset time period (2017-2022) and calculate the seasonal NDVI of street view sampling points.

[0132] Using the Google Earth Engine (GEE) platform, the spatial extent of the remote sensing imagery was specified as Nanjing City, and the time range was 2017-2022. Cloud cover filtering was applied to remove Sentinel-2 images with cloud cover >20%, thus obtaining multi-temporal remote sensing imagery data. Similar to the calculation of the average GVI mentioned above, the calculation of the average NDVI for each temporal phase first extracts the NDVI of all remote sensing images within that temporal phase, and then takes the average value.

[0133] Finally, the multi-temporal NDVI data was exported from the GEE platform.

[0134] The Normalized Difference Vegetation Index (NDVI) of Nanjing City at different time phases is calculated using Formula 3 above. The average NDVI of Nanjing City in each season is also calculated. avg Calculated using Formula 4 above.

[0135] Finally, in ArcGIS Pro, a 30-meter buffer was established for all street view sampling points for each time phase. Since the Sentinel-2 image and its extracted multi-temporal NDVI resolution is 10 meters, the 30-meter buffer of the street view sampling points was first converted into raster data using the partition statistics tool in ArcGIS Pro, and then the average NDVI value within the buffer was calculated to obtain the seasonal NDVI of the street view sampling points.

[0136] 5. The missing GVI is estimated using NDVI through a local regression model that incorporates seasonal weights.

[0137] For street view sampling points lacking GVI, a local regression model with seasonal weights is established using NDVI and GVI from the same time phase, with NDVI as the predictor variable and GVI as the target variable. The GVI is estimated using NDVI through this model to obtain the seasonal GVI of each street view sampling point.

[0138] We utilize multi-seasonal NDVI and GVI data from street view sampling points, dividing the dataset into four subsets according to season for separate estimation. Each subset contains NDVI and GVI data for the corresponding season. For each subset, we estimate GVI using NDVI. Simultaneously, we construct a SeasonWeight Local Regression (SWLR) model by combining the sample points from other seasons with the sample points whose GVI is to be estimated, thereby improving the prediction accuracy of GVI.

[0139] like Figure 2 As shown, the dataset is divided into four sub-datasets: spring, summer, autumn, and winter. The local regression model with seasonal weights consists of two steps: the first step is to directly fit the model to the sample points of the same season (the sample points are data in the dataset, specifically the green view rate and normalized vegetation index data of a certain street view sampling point in a certain season); the second step is to integrate seasonal weights into the local regression model and fit the model again to improve accuracy.

[0140] Step 1: As Figure 2 As shown, assuming point p0 is the sample point for which the GVI is to be estimated, a local regression model is fitted to the points with NDVI values ​​around point p0 to predict the GVI of point p0, as follows. Figure 3 As shown ( Figure 3 The meaning of midpoint and Figure 4 (Same). Since p0 is a sample point in spring, for street view sampling points at different locations in the same season (spring), the Gaussian weight function is used to assign weights to the sample points (points p1, p2, p3 and p4 are sample points with GVI within the bandwidth range and have weights greater than zero, while other sample points with GVI outside the bandwidth range have weights of zero). The Gaussian weight function is shown in formula (5):

[0141]

[0142] Where x is the NDVI value of sample point p, x0 is the NDVI value of sample point p0 to be estimated, σ is the bandwidth of the Gaussian weighting function, and w local (p) represents the local weight of sample point p. After obtaining the weights of the sample points, the model is fitted using the weighted least squares method. The bandwidth σ is obtained by dividing the subset dataset into training and validation sets, and evaluating the model's performance on the validation set using the goodness-of-fit R². Finally, for the four seasons of spring, summer, autumn, and winter, the bandwidth σ is set to 0.5, 0.8, 1.2, and 0.6, respectively.

[0143] Step 2: For the sample point p0 where GVI needs to be estimated, although point p0 does not have GVI, sample points at the same spatial location (street view sampling point) in other seasons corresponding to point p0 may have GVI. For example... Figure 2 As shown, the sample points at the same location corresponding to point p0 in other seasons (summer, autumn, and winter) are p5, p6, and p7. Among them, p5 and p7 are sample points containing GVI and are called auxiliary seasonal sample points (or simply auxiliary points). When the number of auxiliary seasonal sample points is 1, 2, or 3, the seasonal weights of the auxiliary points and the point to be predicted are calculated, and then fused with the local weights. Finally, a local regression model with fused seasonal weights is fitted, as shown below. Figure 4 As shown.

[0144] The seasonal weights between auxiliary points and the sample points to be estimated are calculated as shown in Formula 6:

[0145] w season (p)=λexp(-dist season (t,t0)) (6)

[0146] Among them, w season (p) represents the seasonal weight between the auxiliary point and the sample point of green visibility to be estimated, λ is the hyperparameter used to amplify the seasonal weight, t is the season of the auxiliary point p, t0 is the season of the sample point p0 to be estimated, and dist season (t, t0) is the seasonal distance between the auxiliary point and the green visibility sample point to be estimated. This distance is calculated using the Pearson distance with all sample points containing four seasons, as shown in Formula 7.

[0147] dist season (t,t0)=1-Pearson(t,t0) (7)

[0148] Wherein, Pearson(t,t0) is the Pearson correlation coefficient of all sample points containing GVI between season t and t0, calculated by formula 8.

[0149]

[0150] Where, x i With y i The GVI of the same street view sampling point at seasons t and t0.

[0151] In Formula 6, exp+(-dist) season (t,t0)) represents the seasonal weights before amplification, and λ is a hyperparameter used to amplify the seasonal weights, which can take values ​​of 10, 100, 103, 104, etc. The values ​​of λ are obtained by evaluating the effect on the validation set. The values ​​of λ for spring, summer, autumn, and winter are 104, 105, 106, and 104, respectively. season (p) represents the seasonal weights of the final auxiliary sample points.

[0152] For the sampling point p0 to be predicted, merge (merging refers to merging two vectors, for example, vectors A = [1,2], B = [3,4], A and B are merged into C = [1,2,3,4])w local (p) and w season (p) represents the weight w of all sample points (all sample points refer to all sample points in the current season and auxiliary seasonal points). i (p) (i = 1, ..., n), where n is the number of samples, and a weight matrix W is constructed, where W is a diagonal matrix and its diagonal elements are w i (p), where X is the design matrix and y is the dependent variable vector. The regression coefficients are solved using the weighted least squares method according to Formula 8.

[0153] The design matrix X is a matrix containing the NDVI values ​​of all sample points and auxiliary seasonal points for that season (because it has only a single feature, it is actually a vector, n*1). The dependent vector y is a vector containing the GVI values ​​of all sample points and auxiliary seasonal points for that season. T stands for transpose matrix.

[0154]

[0155] Finally, the predicted value of GVI is calculated using Formula 9.

[0156]

[0157] Using the above method, the GVI of street view sampling points is predicted for each of the four seasons, and the GVI of all street view sampling points for the four seasons is obtained.

[0158] Compared with the prior art, the technical advantages of the present invention are as follows:

[0159] Introducing multi-temporal data: This invention constructs a multi-temporal dataset by collecting street view images of target city streets at different times. It considers not only the green view rate at a single moment but also seasonal variations throughout the year. This method overcomes the limitations of fixed-temporal calculations of green view rate in existing technologies, providing more dynamic and comprehensive greening assessment results.

[0160] Spatiotemporal dynamic monitoring capability: This invention can dynamically monitor the greening status of urban streets and reflect the spatiotemporal change trend of greening. It not only provides data support for the assessment of the current status of urban greening, but also provides more continuous and accurate information for long-term urban planning and greening management.

[0161] Completing Missing Green Visibility (GVI) Samples: Due to the significant human, material, and financial resources required for street view image data acquisition, some street view sampling points may lack street view image data for specific seasons. Given the high dynamic monitoring capabilities of remote sensing imagery, all sampling points typically possess multi-temporal remote sensing imagery. Both the Normalized Difference Vegetation Index (NDVI) extracted from remote sensing imagery and GVI can reflect vegetation growth, and previous studies have shown a strong correlation between the two. For these samples lacking GVI, a seasonally weighted local regression model is constructed to estimate the missing GVI using NDVI. The completed GVI can more accurately assess the greening status of urban streets.

[0162] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for estimating the seasonal green view rate of urban streets based on an improved local regression model, characterized in that, Includes the following steps: S1. Collect road network data, extract road centerlines, and generate street view sampling points at preset distances; S2. Collect street view images of each street view sampling point covering multiple time phases of the four seasons within a preset time period, and count the street view sampling points with missing street view images and their corresponding time phases; S3. Use a deep learning model to perform semantic segmentation on the collected street view images and extract the vegetation areas in the street view images; S4. Calculate the green visibility rate of each street view sampling point in the acquired street view image at each time phase; S5. Calculate the green visibility rate of each street view sampling point in the acquired street view image under each season; S6. Collect multi-temporal remote sensing image data for the same preset time period as in step S2, and calculate the normalized vegetation index for each street view sampling point in each season. S7. By using a local regression model that incorporates seasonal weights, the green view rate of a street view sampling point in a certain season is estimated using the normalized vegetation index of the street view sampling point in each season, which cannot be calculated due to the lack of street view images.

2. The method for estimating the seasonal green view rate of urban streets based on an improved local regression model according to claim 1, characterized in that, In step S4, the green visibility rate of each street view sampling point in the acquired street view image at each time phase is calculated using the following formula; Where GVI is the green visibility rate of a street view sampling point at a certain time phase, i (=1,2,3,4) represents the street view images of the sampling point in four directions at that time phase, and Area Green Area is the number of pixels in a street view image from a certain direction that are identified as vegetation areas. All It represents the total number of pixels in a street view image in a certain direction.

3. The method for estimating the seasonal green view rate of urban streets based on an improved local regression model according to claim 2, characterized in that, In step S5, the green visibility rate of each street view sampling point in the acquired street view image under each season is calculated using the following formula; Among them, GVI avg It represents the green view rate (GVI1+GVI2+...GVI) at a street view sampling point during a specific season. n It is the sum of the green visibility of the street view sampling point in each time phase during the season, and n is the number of time phases of the street view sampling point in the season.

4. The method for estimating the seasonal green view rate of urban streets based on an improved local regression model according to claim 1, characterized in that, Step S6 involves collecting multi-temporal remote sensing image data for the same preset time period as in step S2, and calculating the normalized vegetation index for each street view sampling point in each season. Specifically, this includes: The specified spatial range includes all street view sampling points, and the time range is within a preset time period and covers multiple time phases of the four seasons, acquiring remote sensing images of multiple time phases within the specified spatial range; The normalized vegetation index for each time phase within a specified spatial range is calculated using the following formula: Wherein, NDVI is the normalized vegetation index at a certain time within a specified spatial range, NIR is the reflectance of the near-infrared band in the remote sensing image at that time, and R is the reflectance of the red band in the remote sensing image at that time. The normalized vegetation index for each season within a specified spatial area is calculated using the following formula: Among them, NDVI avg It is the average normalized vegetation index (NDVI) within a specified spatial area under a certain season, NDVI1+NDVI2+...NDVI n It is the sum of the normalized vegetation indices of the specified spatial range in each time phase under the season, and n is the number of time phases of the specified spatial range under the season; Establish a buffer zone with a preset distance for all street view sampling points, and extract and count the NDVI values ​​within the buffer zone. avg The normalized vegetation index of each street view sampling point in each season was obtained.

5. The method for estimating the seasonal green view rate of urban streets based on an improved local regression model according to claim 4, characterized in that, In step S7, a local regression model incorporating seasonal weights is used to estimate the green view rate of a street view sampling point in a given season, which cannot be calculated due to a lack of street view images. This includes: S71. The green view rate and normalized vegetation index of the street view sampling points in each season are divided into 4 subsets according to the season. The green view rate and normalized vegetation index of a certain street view sampling point in a certain season are taken as a sample point in the subset. S72. Use a local regression model to fit a model to sample points in the same season; S73. Integrate seasonal weights into the local regression model and fit the model again to estimate the green view rate of a street view sampling point in a certain season, which cannot be calculated due to the lack of street view images.

6. The method for estimating the seasonal green view rate of urban streets based on an improved local regression model according to claim 5, characterized in that, Step S72 involves fitting a model to sample points from the same season using a local regression model, specifically including: A Gaussian weighting function is used to assign local weights to sample points that are in the same season but in different spatial locations as the sample points whose green visibility rate is to be estimated, and then the model is fitted by weighted least squares method.

7. The method for estimating the seasonal green view rate of urban streets based on an improved local regression model according to claim 6, characterized in that, Step S73 involves incorporating seasonal weights into the local regression model to refit the model and estimating the green view rate of a street view sampling point in a given season, which cannot be calculated due to the lack of street view images. Specifically, this includes: Sample points in other seasons that are in the same spatial location as the sample point whose green view rate is to be estimated and contain seasonal green view rate are used as auxiliary points. The seasonal weights of the auxiliary points and the sample points whose green view rate is to be estimated are calculated, and then fused with the local weights. Finally, a local regression model with fused seasonal weights is fitted. The green view rate of a certain street view sampling point in a certain season that cannot be calculated due to the lack of street view images is estimated using the local regression model with fused seasonal weights. The seasonal weights of auxiliary points and sample points for estimating green visibility are calculated using the following formula: w season (p)=λexp(-dist season (t,t0)) Among them, w season (p) represents the seasonal weights of the auxiliary points and the sample points for estimating the green visibility rate, λ is a hyperparameter used to amplify the seasonal weights, and dist season (t,t0) is the seasonal distance between the auxiliary point and the green visibility sample point to be estimated; The regression coefficients are calculated using the following formula: in, Here, X is the regression coefficient, W is the design matrix, and y is the dependent variable vector; Finally, the green view rate of the sample points to be estimated is calculated using the following formula: in, Let X be the green view rate of the sample points to be estimated, and let X be the design matrix. is the regression coefficient.

8. The method for estimating the seasonal green view rate of urban streets based on an improved local regression model according to claim 1, characterized in that, After collecting road network data in step S1, the process also includes road filtering and simplification, extracting road centerlines and generating street view sampling points at preset distances.

9. The method for estimating the seasonal green view rate of urban streets based on an improved local regression model according to claim 1, characterized in that, Step S2 includes stitching the street view images into a panoramic image before acquiring the data.

10. A method for estimating the seasonal green view rate of urban streets based on an improved local regression model according to claim 1, characterized in that, The deep learning model in step S3 is Mask2Former, a deep learning semantic segmentation model pre-trained on the Cityscapes dataset.

Images