Urban green land fine classification method based on high-resolution remote sensing image

Abstract

The invention discloses an urban green land fine classification method based on a high-resolution remote sensing image. The method comprises the following steps: S1, obtaining a VHR image and monthly Sentinel-2 remote sensing time sequence data; s2, obtaining a greenbelt boundary sample and a greenbelt type sample according to the VHR image and monthly Sentinel-2 remote sensing time sequence data; s3, performing land parcel semantic segmentation on the greenbelt boundary sample to obtain urban greenbelt land parcels, and performing land parcel super-pixel segmentation to obtain super-pixel land parcels so as to obtain greenbelt classification land parcels; s4, performing plot time sequence feature extraction on the green land type samples to obtain time sequence feature pixels; s5, performing morphological feature calculation on the classified land parcels of the green land to obtain a land parcel morphological feature subset, and performing feature engineering extraction on the time sequence feature pixels to obtain a land parcel time sequence element feature subset; and S6, carrying out urban green land fine category classification by adopting a classifier to obtain a classification result. According to the method, fine classification can be accurately carried out on the urban green land, and technical support is provided for urban green land management and ecological protection in the future.

Description

Technical Field

[0001] The present invention relates to the technical field of urban green space identification, and in particular to a method for finely classifying urban green space based on high-resolution remote sensing images. Background Art

[0002] As an important part of the urban ecosystem, urban green space is a general term for green vegetation in the city. It has many functions for urban society, not only in terms of ecological environment protection and improvement, but also in terms of residents' quality of life, social interaction, urban image, economic development and other aspects. Therefore, it is of great significance to rationally plan and manage urban green space and give full play to its functions in order to build a livable, business-friendly and tourist-friendly urban environment. In order to scientifically plan urban green space, it is necessary to quantify and measure the green space area, coverage rate and even the specific types of green space. These data are helpful for ecological city calculation planning and become an important data support for maintaining the sustainable development of future cities.

[0003] Green space fine classification can be divided into pixel-based classification and plot-based (parcel-level) classification. Pixel-based classification dominates existing research because it is easy to implement and is in continuous development and improvement. However, it has many disadvantages. Pixel-level classification often fails to fully consider the spatial continuity and contextual information of objects, and is easily confused in places such as blurred boundaries and spectral mixing, and there is salt and pepper noise. Plot-based methods have gained momentum and are widely used under the impetus of extremely high spatial resolution remote sensing images. At present, the green space classification methods based on plots mainly rely on three plot extraction technologies: edge-based methods, image segmentation methods, and deep learning algorithms. Edge-based methods use pixel differences to extract plot boundaries, which are suitable for simple scenes, but are easily affected by noise in complex environments and lack boundary accuracy; image segmentation methods divide images into homogeneous areas through texture and color features, which can generate geometric patches, but ignore the actual geographical significance and deviate from the distribution of real green space plots; deep learning methods combine edge and texture features to extract plots, which have higher accuracy and more complete semantic information, but rely heavily on sample annotation quality and computing resources. In addition, existing methods still have shortcomings in dealing with complex green space types and significant temporal and spatial characteristics. Summary of the invention

[0004] In view of the above problems, the present invention aims to provide a method for fine classification of urban green spaces based on high-resolution remote sensing images.

[0005] The technical solution of the present invention is as follows:

[0006] A fine classification method for urban green space based on high-resolution remote sensing images includes the following steps:

[0007] S1: VHR images are obtained through Google Earth, and monthly Sentinel-2 remote sensing time series data are obtained through Google Earth Engine;

[0008] S2: obtaining green space boundary samples according to the VHR images, and obtaining green space type samples according to the green space boundary samples combined with the monthly Sentinel-2 remote sensing time series data;

[0009] S3: performing plot semantic segmentation on the green space boundary samples to obtain urban green space plots, performing plot superpixel segmentation on the urban green space plots to obtain superpixel plots, and combining the urban green space plots and the superpixel plots to obtain green space candidate classification plots;

[0010] S4: extracting the time series features of the green space type samples based on the monthly NDVI changes of the monthly Sentinel-2 remote sensing time series data to obtain time series feature pixels;

[0011] S5: Calculate the morphological features of the green space phenological classification plots to obtain a plot morphological feature subset, and perform feature engineering extraction on the temporal feature pixels to obtain a plot temporal element feature subset;

[0012] S6: Using a classifier to perform fine classification of urban green space according to the morphological feature subset of the land parcel and the temporal feature subset of the land parcel, to obtain a fine classification result of the urban green space.

[0013] Preferably, in step S3, the D-LinkNet model is used to perform semantic segmentation of land parcels.

[0014] Preferably, in step S3, when performing superpixel segmentation of the plot, the SLIC method is first used to segment the superpixel blocks, and then the hierarchical clustering method is used to merge the RAGs to obtain the superpixel plot.

[0015] Preferably, in step S4, an LSTM model is used to extract time series features.

[0016] Preferably, in step S5, the morphological features include plot area, plot perimeter, plot roundness, morphological index, minimum enveloping rectangle length, minimum enveloping rectangle width and minimum enveloping rectangle aspect ratio, and the plot roundness and the morphological index are calculated respectively by the following formulas:

[0017]

[0018] Where: Roundness is the roundness of the plot; Area is the area of ​​the plot; Perimeter is the perimeter of the plot; ShapeIndex is the morphological index.

[0019] Preferably, in step S5, when performing feature engineering extraction, the design coverage, deviation, maximum independent distribution score and independent distribution number are the sequential meta-features of feature engineering extraction, and each sequential meta-feature of feature engineering extraction is calculated by the following formula:

[0020]

[0021] Where: area i The area of ​​green space in the plot; area p is the area of ​​the plot; n is the number of IFDs in the plot; IFD is an independent feature pixel; ω i is the area weighting coefficient; DIS() means to find the distance; c p is the centroid of the plot; c IFD is the centroid of IFD; r is the maximum distance; MIFD is the maximum independent distribution fraction; MAX() means finding the maximum value; area IFD is the area of ​​IFD; SUM() means sum; IDN is the independent distribution number; NUMBER() means the total number.

[0022] Preferably, in step S6, the classifier adopts a random forest classifier.

[0023] Preferably, when step S3 performs semantic segmentation of land parcels, step S4 performs temporal feature extraction of land parcels, and step S6 performs fine classification of urban green space, a step of performing precision analysis is also included.

[0024] Preferably, when performing precision analysis in step S3 and step S5, pixels are used as the unit of precision evaluation, verification sample points of each category are uniformly selected in the entire image, and evaluation indicators based on the confusion matrix are calculated.

[0025] Preferably, the evaluation index based on the confusion matrix includes accuracy, precision, recall, F1 score and intersection-over-union ratio, and each evaluation index is calculated by the following formula:

[0026]

[0027] Where: Accuracy is the accuracy rate; TP is the number of samples whose true category is positive and correctly classified as positive; TN is the number of samples whose true category is negative and correctly classified as negative; FP is the number of samples whose true category is negative but incorrectly classified as positive; FN is the number of samples whose true category is positive but incorrectly classified as negative; Precision is the precision rate; Recall is the recall rate; F1 is the F1 score; IoU is the intersection over union ratio; Area o Area is the intersection area of ​​the segmented area and the actual segmented area; Uis the union area of ​​the segmented area and the actual segmented area.

[0028] The beneficial effects of the present invention are:

[0029] The present invention can carry out fine extraction and classification of urban green spaces, and provide technical support for promoting the development of urban planning, environmental monitoring, ecological protection and other fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0031] Figure 1 It is a schematic diagram of the process of the urban green space fine classification method based on high-resolution remote sensing images of the present invention;

[0032] Figure 2 A schematic diagram of the network structure of a D-LinkNet model in a specific embodiment;

[0033] Figure 3 A schematic diagram of the network structure of an LSTM model in a specific embodiment;

[0034] Figure 4 A schematic diagram of a training sample of a D-LinkNet model in a specific embodiment;

[0035] Figure 5 A schematic diagram of the extraction results of the D-LinkNet model in a specific embodiment;

[0036] Figure 6 It is a schematic diagram of superpixel segmentation and merging effect in a specific embodiment;

[0037] Figure 7 It is a schematic diagram of the effect of extracting green space phenological classification plots in a specific embodiment;

[0038] Figure 8 A schematic diagram of time-series characteristic pixels of an urban area and surrounding farmland in a specific embodiment;

[0039] Fig. 9 is a schematic diagram of a VHR image in a specific embodiment;

[0040] Fig.10 A schematic diagram of the classification effect of urban areas in a specific embodiment and the comparison with real objects;

[0041] Fig.11A schematic diagram of the classification effect of the urban periphery area and the comparison with the real ground objects in a specific embodiment;

[0042] Fig.12 It is a detailed schematic diagram of the urban green space fine category classification result in a specific embodiment. DETAILED DESCRIPTION

[0043] The present invention is further described below in conjunction with the accompanying drawings and embodiments. It should be noted that, in the absence of conflict, the embodiments in this application and the technical features in the embodiments can be combined with each other. It should be noted that, unless otherwise specified, all technical and scientific terms used in this application have the same meanings as those generally understood by those of ordinary skill in the art to which this application belongs. The words "including" or "comprising" and the like used in the disclosure of the present invention mean that the elements or objects appearing before the word cover the elements or objects listed after the word and their equivalents, without excluding other elements or objects.

[0044] like Figure 1 As shown, the present invention provides a method for fine classification of urban green space based on high-resolution remote sensing images, comprising the following steps:

[0045] S1: VHR images were obtained through Google Earth, and monthly Sentinel-2 remote sensing time series data were obtained through Google Earth Engine.

[0046] S2: Obtain green space boundary samples according to the VHR image, and obtain green space type samples according to the green space boundary samples combined with the monthly Sentinel-2 remote sensing time series data.

[0047] S3: Performing plot semantic segmentation on the green space boundary samples to obtain urban green space plots, and performing plot superpixel segmentation on the urban green space plots to obtain superpixel plots, and combining the urban green space plots and the superpixel plots to obtain green space candidate classification plots.

[0048] In a specific embodiment, the D-LinkNet model is used for semantic segmentation of land parcels. The D-LinkNet model is a model inherited and developed from LinkNet. It embeds a hole convolution layer in the central part of the model, which can expand the receptive field as much as possible without sacrificing the resolution of the feature map, realize the effective fusion of multi-scale features, and retain the fine spatial information of the green space. The core idea of ​​the model is to first encode the road information into the features, and then remap this information to the spatial domain for image segmentation in the decoding stage. By introducing convolution layers in the central area without adding additional learning parameters, the D-LinkNet model simplifies the training process. Compared with the pooling layer in the LinkNet model, the advantage of the hole convolution layer is that it maintains the resolution of the feature map without losing spatial information. It introduces continuous dilated convolutions with different expansion rates and adds them to the core structure of the model. Dilated convolution is a method to expand the receptive field of the convolution kernel by calculating the linear combination of the kernel parameters and a limited number of pixels within its range. Therefore, dilated convolution is often used to replace traditional pooling techniques (such as maximum pooling and average pooling), which is highly praised for its simplicity and can effectively expand information without reducing the resolution of feature maps. In a specific embodiment, the network structure diagram of the D-LinkNet model used is as follows: Figure 2 shown.

[0049] In a specific embodiment, when performing superpixel segmentation of a land block, the SLIC method is first used to segment the superpixel blocks, and then the hierarchical clustering method is used to merge the RAGs to obtain the superpixel land block.

[0050] In the above embodiment, the SLIC method (simple linear iterative clustering) is used to segment superpixel blocks, which can convert the RGB image into CIELAB color and XY coordinate space, so that the image pixel information is represented by a five-dimensional vector, and then a multi-dimensional space distance metric is constructed to perform similarity clustering on these vectors representing the image pixel information. Then, a compact and uniformly distributed superpixel pattern is generated, which has the advantages of being compact, suitable for color and grayscale images, simple parameters, fast speed and maintaining clear contours. In terms of boundary fitting, computational complexity and regional compactness, the use of this algorithm can achieve results that are superior to other segmentation algorithms. It is still very effective when applied to the segmentation of urban green spaces on RGB images.

[0051] The SLIC method first generates starting points (cluster centers) uniformly over the entire image: the number of superpixels is uniformly generated in the image according to the parameter setting or the number of superpixels calculated by the parameters. For example, if the image has N pixels, and the parameter setting is to segment K superpixels of uniform size and distribution, it can be calculated that each superpixel contains N / K image pixels, then the initial distance (step length) between the generated adjacent starting points is approximately To prevent the starting point from falling on the edge of the superpixel patch, the edge image gradient is generally large, so it is necessary to ensure that the seed point is located at the position where the gradient of the starting point is the smallest. Therefore, it is necessary to search the 3×3 pixel space around the seed point with the starting point as the center and find the position with the smallest gradient in this area.

[0052] The SLIC method aggregates based on the similarity between CIELAB color and xy coordinate 5-dimensional space. First, the RGB image pixels are converted to LAB color 3-dimensional space, and then the x-coordinate and y-coordinate dimensions of the pixels are added so that each pixel generates a vector C in the 5-dimensional space. i = {l i , a i , b i , x i ,y i}, the weighted distance D between the pixel and the seed point pixel and the calculation formula of the corresponding parameters are:

[0053]

[0054] Where: m is the weight.

[0055] Based on the weighted distance, the K-means method is used to cluster pixels, that is, iteratively update the calculation, and aggregate isolated points or small clusters into the nearest neighbor cluster, that is, to complete the superpixel segmentation. Finally, the hierarchical clustering method is used to create a regional adjacency graph (RAG), and gradually merge adjacent superpixels based on the similarity between adjacent superpixels. The adjacency graph and similarity are continuously updated to achieve step-by-step merging, and finally a hierarchical segmentation result is output, which can effectively capture the large-scale structure of the image and reduce over-segmentation.

[0056] S4: extracting the time series features of the green space type samples based on the monthly NDVI changes of the monthly Sentinel-2 remote sensing time series data to obtain time series feature pixels.

[0057] In the present invention, the plot time series feature extraction of the green space type sample is performed using the monthly NDVI change. NDVI is specially designed to measure and monitor vegetation health and coverage. It calculates the biomass and health of plants by using the reflectivity of red light and near-infrared light. Therefore, using NDVI as a feature can help the model focus more on vegetation-related information and filter out other irrelevant ground features (such as buildings, bare soil, water bodies, etc.), thereby improving the accuracy of fine classification of green space. And the one-dimensional training feature greatly reduces the dimension of the data, making model training more efficient and reducing the demand for computing resources. This advantage is more prominent when processing high spatial resolution remote sensing images. Finally, NDVI is suitable for multi-temporal analysis, that is, analyzing vegetation changes at the same place at different times. This is of great significance for monitoring seasonal changes, crop growth cycles, disaster impacts, and vegetation type distinctions. Models trained using NDVI can better capture these changes, which is of great significance for fine classification of green space. Using NDVI as a feature helps to improve the focus, accuracy and efficiency of the model, while having strong noise resistance and the advantage of adapting to multi-temporal analysis.

[0058] In a specific embodiment, an LSTM model is used for time series feature extraction. As a special recursive neural network (RNN), LSTM can improve the problems of gradient vanishing and gradient exploding that exist in traditional RNN when processing long sequence information. Gradient explosion means that the gradient will gradually increase with the process of back propagation, making the model unstable, and gradient vanishing means that as the sequence length increases, the influence and capture ability of the gradient in long-distance situations gradually weaken or even disappear. By introducing memory units and gating mechanisms, LSTM can capture long-term dependencies in sequences and is widely used in time series prediction, natural language processing and other fields. In a specific embodiment, the LSTM model network framework used consists of two layers and 18 hidden units. The network structure diagram is shown in the figure. Figure 3 shown.

[0059] S5: Calculate the morphological features of the green space phenological classification plots to obtain a plot morphological feature subset, and perform feature engineering extraction on the temporal feature pixels to obtain a plot temporal element feature subset.

[0060] In a specific embodiment, the morphological features include the plot area, plot perimeter, plot roundness, morphological index, minimum enveloping rectangle length, minimum enveloping rectangle width and minimum enveloping rectangle aspect ratio, and the plot roundness and the morphological index are calculated by the following formulae respectively:

[0061]

[0062] Where: Roundness is the roundness of the plot; Area is the area of ​​the plot; Perimeter is the perimeter of the plot; ShapeIndex is the morphological index.

[0063] In a specific embodiment, when performing feature engineering extraction, the design coverage, deviation, maximum independent distribution score and independent distribution number are the sequential meta-features of feature engineering extraction, and each sequential meta-feature of feature engineering extraction is calculated by the following formula:

[0064]

[0065] Where: area i The area of ​​green space in the plot; area p is the area of ​​the plot; n is the number of IFDs in the plot; IFD is an independent feature pixel; ω i is the area weighting coefficient; DIS() means to find the distance; c p is the centroid of the plot; c IFD is the centroid of IFD; r is the maximum distance; MIFD is the maximum independent distribution fraction; MAX() means finding the maximum value; area IFD is the area of ​​IFD; SUM() means sum; IDN is the independent distribution number; NUMBER() means the total number.

[0066] In the above embodiment, the coverage describes the proportion of feature pixels in the plot. The larger the coverage value, the more corresponding feature pixels there are in the plot. The deviation rate describes the degree of deviation of the distribution of independent feature pixels in the plot, where independent feature pixels refer to the connected feature pixel set in the plot. The maximum independent distribution score is the proportion of the maximum IFD. The number of independent distributions is the number of IFDs. The combination of MIDF and IDN can describe the degree of discreteness of IFD. These time series meta-features can describe the distribution of time series feature pixels in the plot area, which is essentially an abstraction of the time series information in the plot.

[0067] S6: Using a classifier to perform fine classification of urban green space according to the morphological feature subset of the land parcel and the temporal feature subset of the land parcel, to obtain a fine classification result of the urban green space.

[0068] In a specific embodiment, the classifier adopts a random forest classifier. Random forest (RF) is based on decision tree (DT) and introduces a random feature selection process, using a bagged ensemble method. RF extracts self-help samples from the original data set and constructs multiple separate decision trees. When constructing a decision tree, each decision tree randomly selects a part of the variables and selects the best splitting variable as the splitting node. RF integrates a large number of decision trees and finally obtains the classification result by majority voting.

[0069] In a specific embodiment, when step S3 performs semantic segmentation of land parcels, step S4 performs temporal feature extraction of land parcels, and step S6 performs fine classification of urban green space, a step of performing precision analysis is also included.

[0070] For the accuracy evaluation of land parcel classification, although the final green land parcel classification is at the land parcel level, considering that the green land parcels obtained by superpixel segmentation are of different sizes and there are some partial land parcels with large area differences, if the land parcel is used as the accuracy evaluation unit, the final evaluation and calculation index will fluctuate greatly due to the area proportion of the selected verification-level sample land parcels in the entire image, and thus cannot accurately reflect the image classification effect. Therefore, in a specific embodiment, when performing accuracy analysis in steps S3 and S5, pixels are used as the unit of accuracy evaluation, and verification sample points of each category are evenly selected in the entire image, and evaluation indicators based on the confusion matrix are calculated.

[0071] Optionally, the evaluation index based on the confusion matrix includes accuracy, precision, recall, F1 score and intersection-over-union ratio, and each evaluation index is calculated by the following formula:

[0072]

[0073] Where: Accuracy is the accuracy rate; TP is the number of samples whose true category is positive and correctly classified as positive; TN is the number of samples whose true category is negative and correctly classified as negative; FP is the number of samples whose true category is negative but incorrectly classified as positive; FN is the number of samples whose true category is positive but incorrectly classified as negative; Precision is the precision rate; Recall is the recall rate; F1 is the F1 score; IoU is the intersection over union ratio; Area o Area is the intersection area of ​​the segmented area and the actual segmented area; U is the union area of ​​the segmented area and the actual segmented area.

[0074] In a specific embodiment, a certain area in Chengdu is used as a research area, and the urban green space fine classification method based on high-resolution remote sensing images of the present invention is used to finely classify the urban green space, which specifically includes the following steps:

[0075] (1) Data preparation

[0076] The RGB three-band remote sensing data captured at screen level 20 on Google Earth (corresponding to a spatial resolution of 0.27m, during 2022) were cropped, spliced, and reprojected to the WGS-84 coordinate system to obtain high spatial resolution remote sensing data (VHR images).

[0077] Remote sensing time series data In Google Earth Engine (GEE), cloud cover screening was performed on 10m spatial resolution Sentinel-2 images from January to December 2022 to ensure that there was an image with ideal cloud cover and cloud distribution in each month as much as possible (finally 9 remote sensing data were selected). After declouding, atmospheric correction, cropping and other operations, the near infrared band (B8) and infrared band (B4) of Sentinel-2 data were used to calculate the normalized vegetation index:

[0078]

[0079] The 9 calculated NDVI data are used as the basic data for time series feature extraction.

[0080] Green space boundary samples and green space type samples were obtained based on the VHR images and monthly Sentinel-2 remote sensing time series data obtained above.

[0081] (2) Performing semantic segmentation on the green space boundary samples to obtain urban green space plots

[0082] In this embodiment, the D-LinkNet model is used to extract linear land features, and its large receptive field can adapt to high-resolution images. The extremely high spatial resolution remote sensing image of Chengdu (0.13 meters) obtained from the Google Earth map in 2020 is used as the research data, and samples are collected on the image. In order to ensure the representativeness of the sample, the selected sample area contains as many types of urban green spaces as possible, including trees, grasslands, parks and other different situations. 1400 training sets and 200 validation sets of RGB images with 1024×1024 pixels and corresponding green space labels are used, and the sample images are manually labeled through visual interpretation. For model training and prediction, such as Figure 4 The training data set and the test data set are added to the D-LinkNet model for training and prediction. The total number of training rounds is 80,000, and the results are as follows Figure 5 shown.

[0083] The green space semantic segmentation results of the D-LinkNet model obtained have an accuracy of 0.924, a recall rate of 0.976, an F1 score of 0.941, and an IoU of 0.924. Judging from the indicator data of the precision analysis, the green space data obtained by this model has the potential to be used for fine classification of green spaces.

[0084] (3) Perform superpixel segmentation of the plot and obtain the green space classification plot

[0085] The SLIC method is used to segment superpixel blocks, and the superpixel segmentation granularity is selected as 800 for superpixel segmentation (the most ideal superpixel segmentation parameter should make the superpixel segmentation granularity close to the square of the minimum width of the object of interest). Then, the hierarchical merging algorithm is used to recursively merge superpixel blocks starting from the most similar adjacent superpixels. Specifically, for each pair of adjacent superpixels, if their color and other differences (weights) are less than a predetermined threshold, they are merged into a new superpixel, and the label map and adjacency map are updated. In this way, the number of superpixels is gradually reduced, similar areas are merged together, and finally a simplified label map is obtained, making each superpixel area larger, and still retaining the important features of the original image. This process not only improves the efficiency of image processing, but also lays the foundation for subsequent image analysis (such as classification, target detection, etc.). The patches merged by superpixel segmentation are used as plots, and the green land extracted by the semantic segmentation of the D-LinkNet model is used as the green land plot for filtering and correction, so as to extract the green land plots in the green land obtained by the D-LinkNet model, laying the foundation for the subsequent green land fine classification based on the parcel level. In this embodiment, superpixel segmentation and merging are as follows: Figure 6 As shown in Figure 2, the green space classification plots extracted by combining the semantic segmentation and superpixel segmentation of the D-LinkNet model are shown in Figure 2. Figure 7 shown.

[0086] (4) Extract the time series features of the plot and obtain the time series feature pixels

[0087] Based on 9 sets of time series Sentinel-2 data, the NDVI image corresponding to each image was preprocessed and calculated, and the NDVI feature was trained using the LSTM model. Using the VHR image as the reference image, green space type sample points were selected through visual interpretation on the entire image, including urban trees, urban grasslands, and surrounding farmland. The NDVI value and label of each sample point corresponding to the time series were written into a CSV file based on the sample point location information. This was used as the input data for LSTM model training. The number of LSTM network layers was set to 2, the learning rate was 0.001, and the dropout was set to 0.4 for training. The accuracy of the training model result was 0.9323, and the loss value was 0.2465. The model was then used for prediction to obtain the characteristic pixels of the urban green space time series. The mask results of the characteristic pixels of the green space time series in the urban area and the farmland around the city are shown in the figure. Figure 8 shown.

[0088] (5) Perform morphological feature calculation and feature engineering extraction to obtain the morphological feature subset and the time-series feature subset of the plot, and use the RF classifier to perform fine classification of urban green space

[0089] In this embodiment, the morphological features of the candidate classified green land parcels extracted in step (3) are calculated, including the parcel area, perimeter, roundness, morphological index, minimum enveloping rectangle length, width and aspect ratio. Coverage, deviation, maximum independent distribution fraction (MIDF) and independent distribution number (IDN) are extracted as time series meta-features through feature engineering.

[0090] Through visual interpretation, the corresponding green space type plots are selected as classifier training samples. In this embodiment, the selected samples include urban trees, urban grasslands and surrounding farmland. The green space plots are classified using a random forest classifier. Since there is almost no farmland in the urban core area, two images in the urban core area and the urban peripheral area are selected and the green space classification effects are displayed separately, as shown in FIG. Fig. 9 As shown, the left image is the VHR image of the urban area, and the right image is the VHR image of the urban periphery (including farmland). Fig.10 The results of urban core area classification and the comparison with real objects are shown. Fig.11 The classification effect of the urban surrounding area and the comparison with the real objects are shown. Fig.12 A detailed overview of the effect of fine classification of green spaces is shown.

[0091] Since the sizes of the green space plots are not uniform, if the plot-level accuracy evaluation is used, the evaluation results may not accurately represent the classification accuracy of the entire image due to the different proportions of the selected plots for evaluation. In addition, superpixel segmentation may confuse and aggregate different types of objects with similar pixels. In this embodiment, the present invention adopts a pixel-based evaluation method. Through careful visual interpretation on the image, the same number of verification-level data points are selected for each type of green space object, and the confusion evidence is created and the corresponding evaluation index is calculated. In this experiment, the urban core area image ( Fig.12 The accuracy of fine classification of the image (upper image in the figure) is 0.8296, the precision is 0.8333, the recall is 0.8296, and the F1 score is 0.8314. The image of the surrounding area of ​​the city ( Fig.12 The accuracy of fine classification for the lower image in the figure is 0.8083, the precision is 0.8085, the recall is 0.8083, and the F1 score is 0.8069.

[0092] In summary, the present invention adopts the D-LinkNet model for semantic segmentation of urban green space, which can effectively extract the spatial information of urban green space. By collecting and annotating samples of high-resolution remote sensing images of Chengdu, the D-LinkNet model is trained, and a relatively ideal accuracy is achieved on the validation set. The present invention introduces superpixel segmentation and merging technology, further corrects and optimizes the green space extracted by the D-LinkNet model, and improves the accuracy and integrity of the green space plots. The present invention uses the LSTM model to extract the NDVI time series features, which can effectively reflect the seasonal changes and growth status of urban green space. By calculating the NDVI image on the Sentinel-2 data and training the LSTM model, the spatiotemporal change characteristics of the urban green space are successfully captured, providing important time series information for the subsequent green space classification. Finally, the present invention uses a random forest classifier to classify the extracted green space plots, and then evaluates the classification results through a pixel-based evaluation method, and finally achieves an ideal classification effect. Compared with the prior art, the present invention has significant progress.

[0093] The above description is only a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as a preferred embodiment as above, it is not intended to limit the present invention. Any technician familiar with the profession can make some changes or modifications to equivalent embodiments of equivalent changes using the technical contents disclosed above without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of the technical solution of the present invention.

Claims

1. A fine classification method for urban green space based on high-resolution remote sensing images, characterized in that: The following steps are involved: S1: VHR images are obtained through Google Earth, and monthly Sentinel-2 remote sensing time series data are obtained through Google Earth Engine; S2: obtaining green space boundary samples according to the VHR images, and obtaining green space type samples according to the green space boundary samples combined with the monthly Sentinel-2 remote sensing time series data; S3: performing plot semantic segmentation on the green space boundary samples to obtain urban green space plots, performing plot superpixel segmentation on the urban green space plots to obtain superpixel plots, and combining the urban green space plots and the superpixel plots to obtain green space candidate classification plots; S4: extracting the time series features of the green space type samples based on the monthly NDVI changes of the monthly Sentinel-2 remote sensing time series data to obtain time series feature pixels; S5: Calculate the morphological features of the green space phenological classification plots to obtain a plot morphological feature subset, and perform feature engineering extraction on the temporal feature pixels to obtain a plot temporal element feature subset; S6: Using a classifier to perform fine classification of urban green space according to the morphological feature subset of the land parcel and the temporal feature subset of the land parcel, to obtain a fine classification result of the urban green space.

2. The urban green space fine classification method based on high-resolution remote sensing images according to claim 1 is characterized in that: In step S3, the D-LinkNet model is used to perform semantic segmentation of land parcels.

3. The urban green space fine classification method based on high-resolution remote sensing images according to claim 1 is characterized in that: In step S3, when performing superpixel segmentation of the plot, the SLIC method is first used to segment the superpixel blocks, and then the hierarchical clustering method is used to merge the RAGs to obtain the superpixel plot.

4. The urban green space fine classification method based on high-resolution remote sensing images according to claim 1 is characterized in that: In step S4, the LSTM model is used to extract time series features.

5. The urban green space fine classification method based on high-resolution remote sensing images according to claim 1 is characterized in that: In step S5, the morphological features include the plot area, plot perimeter, plot roundness, morphological index, minimum enveloping rectangle length, minimum enveloping rectangle width and minimum enveloping rectangle aspect ratio. The plot roundness and the morphological index are calculated by the following formulae respectively: In the formula: Roundness is the roundness of the plot; Area is the area of ​​the plot; Perimeter is the perimeter of the plot; Shape Index is the morphological index.

6. The urban green space fine classification method based on high-resolution remote sensing images according to claim 1 is characterized in that: In step S5, when performing feature engineering extraction, the design coverage, deviation, maximum independent distribution score and independent distribution number are the sequential meta-features of feature engineering extraction, and each sequential meta-feature of feature engineering extraction is calculated by the following formula: Where: area i The area of ​​green space in the plot; area p is the area of ​​the plot; n is the number of IFDs in the plot; IFD is an independent feature pixel; ω i is the area weighting coefficient; DIS() means to find the distance; c p is the centroid of the plot; c IFD is the centroid of IFD; r is the maximum distance; MIFD is the maximum independent distribution fraction; MAX() means finding the maximum value; area IFD is the area of ​​IFD; SUM() means sum; IDN is the independent distribution number; NUMBER() means the total number.

7. The urban green space fine classification method based on high-resolution remote sensing images according to claim 1 is characterized in that: In step S6, the classifier adopts a random forest classifier.

8. The urban green space fine classification method based on high-resolution remote sensing images according to any one of claims 1 to 7, characterized in that: When performing semantic segmentation of land parcels in step S3, extracting temporal features of land parcels in step S4, and classifying urban green space into fine categories in step S6, a precision analysis step is also included.

9. The urban green space fine classification method based on high-resolution remote sensing images according to claim 8 is characterized in that: When performing precision analysis in step S3 and step S5, pixels are used as the unit of precision evaluation, verification sample points of each category are uniformly selected in the entire image, and evaluation indicators based on the confusion matrix are calculated.

10. The urban green space fine classification method based on high-resolution remote sensing images according to claim 9 is characterized in that: The evaluation indicators based on the confusion matrix include accuracy, precision, recall, F1 score and intersection-over-union ratio, and each evaluation indicator is calculated by the following formula: Where: Accuracy is the accuracy rate; TP is the number of samples whose true category is positive and correctly classified as positive; TN is the number of samples whose true category is negative and correctly classified as negative; FP is the number of samples whose true category is negative but incorrectly classified as positive; FN is the number of samples whose true category is positive but incorrectly classified as negative. Precision is the precision rate; Recall is the recall rate; F1 is the F1 score; IoU is the intersection over union ratio; Area o Area is the intersection area of ​​the segmented area and the actual segmented area; U is the union area of ​​the segmented area and the actual segmented area.

Images