A high-resolution urban green space identification method

Abstract

The present invention relates to the technical field of urban green space identification, and specifically to a high-resolution urban green space identification method. The present invention can identify urban green spaces more finely by utilizing the two steps of "preliminary identification" and "fine identification", especially the green spaces on both sides of roads. The present invention also utilizes the progress prediction and extraction of deep learning models, and has higher accuracy and migration capabilities than the prior art, and has a faster processing speed, thereby providing a high-quality and efficient method for solving the deficiencies in urban greening identification.

Description

Technical Field

[0001] The invention relates to the technical field of urban green space identification, in particular to a high-resolution urban green space identification method. Background Art

[0002] Urban green space is an important part of the urban ecosystem and is of great significance to the per capita environmental health of the city. Due to the vast area of ​​the city, how to have a macro-identification of the green space in the city affects the statistics of the overall green space distribution in the city.

[0003] With the development of science and technology, remote sensing platforms and machine learning technologies are usually used in existing technologies to identify urban green spaces. However, due to the lack of targeted deficiencies in the identification methods of urban green spaces by remote sensing platforms and machine learning models, such as the lack of labeled data, there is a lack of high-resolution and precise identification and monitoring in the technical field of urban green space identification.

[0004] Therefore, it is necessary to design a high-resolution urban green space identification method. Through a fast, large-scale, high spatial resolution (1m) urban green space identification method, the detailed distribution of urban green spaces in the target city can be obtained, providing accurate data and basis for urban green space statistics. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a high-resolution urban green space identification method. Through a fast, large-scale, high spatial resolution (1m) urban green space identification method, the detailed distribution of urban green spaces in the target city is obtained, providing accurate data and basis for urban green space statistics.

[0006] In order to achieve the above object, the present invention provides a high-resolution urban green space identification method:

[0007] The following steps are involved:

[0008] S1, Satellite data processing:

[0009] S1-1, satellite data download:

[0010] Obtain clear and unobstructed satellite image data with a spatial resolution of 1m through remote sensing satellites and earth image databases; S1-2, sample point production:

[0011] The built-up area is obtained through manual visual interpretation of images and urban green space sample points are generated. After cross-validation to confirm the accuracy, the sample points are randomly divided into training set, validation set and test set.

[0012] S2, Deep Learning Methods:

[0013] S2-1, Preliminary Identification:

[0014] A feature extractor is used to capture fine features from the image and generate preliminary recognition results of urban green spaces, specifically:

[0015] S2-1-1, the design consists of a 7×7 convolutional layer, a batch normalization layer, a rectified linear unit ReLU function, and a maximum pooling layer with a stride of 2;

[0016] S2-1-2 uses two one-dimensional global pooling layers to encode the input features into two direction-aware features in the vertical and horizontal directions to capture the spatial coordinate information;

[0017] S2-1-3, applies 4 1×1 convolution blocks to the fine features and unifies their output channels to 96; then after resizing, the fused features are input into 2 3×3 convolution layers to generate preliminary recognition results of urban green spaces;

[0018] S2-2, fine identification:

[0019] S2-2-1, select a sampling point with lower precision, and construct the point features of the selected sampling point with lower precision based on the preliminary recognition in S2-1 and the fine features extracted by the feature extractor;

[0020] S2-2-2, reclassifying the results of the selected sampling points with lower accuracy through a multi-layer perceptron to achieve fine identification of urban green space, where N points are randomly generated and then M points with the largest uncertainty are selected from the N points;

[0021] S2-2-3, construct 5 convolutional layers, and the output channels of each convolutional layer are 64, 128, 256, 512 and 1 respectively; update the preliminary recognition result of urban green space according to the training results of M points with the largest uncertainty in the randomly selected points after reclassification in S2-2-2, and obtain the refined recognition result of urban green space;

[0022] S2-3, Deep Learning Optimization:

[0023] S2-3-1, loss feature optimization function: Assuming that the initial recognition result of the urban green space in the training set is X, and the actual situation is Y, then the loss between Y and X is calculated using the following formula:

[0024]

[0025] Where L Dice is the loss function of the training set, X∩Y is the intersection of X and Y, |X| and |Y| represent the number of pixels in X and Y respectively;

[0026] The classification result loss of N sampling points is measured by cross entropy loss, which is calculated using the following formula:

[0027]

[0028] Where L CE is the loss function of N sampling points, x i and i Represent the result of the i-th point among N sampling points and the ground truth respectively; on this basis, the loss characteristics of urban green space are learned from them through the deep learning framework and the loss characteristic optimization function is established, which is calculated as follows:

[0029] L ugs =L Dice +L CE ;

[0030] S2-3-2, adversarial training optimization: Input the test set into the trained deep learning model, and predict the loss feature optimization function of the result P, which is calculated using the following formula:

[0031] L D (P) = -[(1-y)log(D(P) (h,w,0) )+ylog(D(P) (h,w,1) )],;

[0032] Where P is the prediction result, y is the test set, y=0 is the feature information consistent with the training set, y=1 is the feature information inconsistent with the training set, h is the test set input, w is the training set input, and D is the fine recognition result;

[0033] Calculate the adversarial loss to generate more training results, using the following formula:

[0034] L adv = -log(D(P) (h,w,1) );

[0035] Where L adv is the loss function of the test set, and D is the fine recognition result;

[0036] Use the following formula for adversarial training optimization:

[0037] L G =L ugs +L adv ;

[0038] Where L G Optimize functions for adversarial training;

[0039] S3, accuracy verification:

[0040] S3-1, accuracy verification indicators:

[0041] Five indicators are used to evaluate accuracy, calculated using the following formulas 7 to 11:

[0042]

[0043]

[0044]

[0045] Where Pre is precision, Rec is recall, F1 is F1 score, IoU is intersection over union, OA is overall accuracy; TP, FP, TN and FN are the number of actual positive predictions as positive, the number of actual negative predictions as positive, the number of actual negative predictions as negative and the number of actual positive predictions as negative, respectively.

[0046] The specific determination of low accuracy in S2-2-1 is the accuracy of the bottom 10%.

[0047] The feature extractor in S2-1 is ResNet.

[0048] The specific steps of S2-2-2 are:

[0049] S2-2-2-1, randomly generate N points from the preliminary recognition results, and then select M points with the largest uncertainty from the N points;

[0050] S2-2-2-2, establish point features based on the sampling points obtained in S2-2-2-1, correspond to the M corresponding fine features, and obtain point features that contain both local details and global background;

[0051] S2-2-2-3, input the point features into the multilayer perceptron to obtain new classification results for M points.

[0052] The multilayer perceptron is a 1×1 convolutional layer.

[0053] M=230, N=1024.

[0054] The kernel size of the convolutional layer in S2-2-3 is 4 and the stride is 2.

[0055] The proportions of training set, validation set and test set in S1-2 are 50%, 20% and 30% respectively.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] By utilizing the two steps of "preliminary identification" and "fine identification", the present invention can identify urban green spaces more finely, especially the green spaces on both sides of the roads, and utilizes the progress prediction and extraction of the deep learning model. Compared with the existing technology, it has higher accuracy and migration capabilities, and faster processing speed, providing a high-quality and efficient method for solving the deficiencies in urban greening identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is the overall flow chart of the present invention;

[0059] Figure 2 It is a schematic diagram for comparing various processes of the present invention; DETAILED DESCRIPTION

[0060] The present invention will now be further described with reference to the accompanying drawings.

[0061] See also Figures 1-2 The present invention provides a high-resolution urban green space identification method, comprising the following steps:

[0062] S1, Satellite data processing:

[0063] S1-1, satellite data download:

[0064] Obtain clear and unobstructed satellite image data with a spatial resolution of 1m through the Gaofen-2 (GF2) and Google Earth image database;

[0065] S1-2, sample point preparation:

[0066] This example uses Guangzhou in 2022 as a case study area, obtains the built-up area through manual visual interpretation of images and generates 5523 urban green space sample points. After cross-validation to confirm that the accuracy is 98%, the sample points are randomly divided into 50% training set, 20% validation set and 30% test set;

[0067] S2, Deep Learning Methods:

[0068] S2-1, Preliminary Identification:

[0069] ResNet is used as a feature extractor to capture fine features from images and generate preliminary recognition results of urban green spaces, specifically:

[0070] S2-1-1, ResNet consists of a 7×7 convolutional layer, a batch normalization layer, a rectified linear unit ReLU function, and a maximum pooling layer with a stride of 2, which is used for urban green space feature enhancement;

[0071] S2-1-2, the maximum pooling layer includes two one-dimensional global pooling layers, which encode the input features into two direction-aware features along the vertical and horizontal directions to capture the spatial coordinate information, so as to capture fine features;

[0072] S2-1-3, four 1×1 convolution blocks are applied to fuse the fine features, unifying their output channels into 96; then, after resizing, the fused features are input into two 3×3 convolution layers to generate the preliminary recognition results of urban green spaces;

[0073] S2-2, fine identification:

[0074] S2-2-1, select 552 sampling points with the lowest 10% accuracy, and construct the point features of the selected sampling points with lower accuracy based on the preliminary recognition of S2-1 and the fine features extracted by the feature extractor;

[0075] S2-2-2, reclassifying the results of the selected sampling points with lower precision through a multi-layer perceptron to achieve fine identification of urban green space, wherein N points are randomly generated, and then M points with the largest uncertainty are selected from the N points. In this embodiment, M=230, N=1024;

[0076] According to the sampling points obtained in S2-2-2-1, point features are established, corresponding to the M corresponding fine features, and point features containing both local details and global background are obtained;

[0077] The point features are input into a multilayer perceptron, where the multilayer perceptron is a 1×1 convolutional layer, to obtain new classification results for M points.

[0078] S2-2-3, construct 5 convolutional layers, each with a kernel size of 4 and a stride of 2; and the output channels of each convolutional layer are 64, 128, 256, 512 and 1 respectively; according to the training results of the M points with the largest uncertainty in the randomly selected points after reclassification in S2-2-2, transfer them to images of other regions to update the preliminary recognition results of urban green spaces and obtain the refined recognition results of urban green spaces;

[0079] S2-3, Deep Learning Optimization:

[0080] S2-3-1, loss feature optimization function: Assuming that the initial recognition result of the urban green space in the training set is X, and the actual situation is Y, then the loss between Y and X is calculated using the following formula:

[0081]

[0082] Where L Dice is the loss function of the training set, X∩Y is the intersection of X and Y, |X| and |Y| represent the number of pixels in X and Y respectively;

[0083] The classification result loss of N sampling points is measured by cross entropy loss, which is calculated using the following formula:

[0084]

[0085] Where L CE is the loss function of N sampling points, x i and iRepresent the result of the i-th point among N sampling points and the ground truth respectively; on this basis, the loss characteristics of urban green space are learned from them through the deep learning framework and the loss characteristic optimization function is established, which is calculated as follows:

[0086] L ugs =L Dice +L CE ;

[0087] S2-3-2, adversarial training optimization: Input the test set into the trained deep learning model, and predict the loss feature optimization function of the result P, which is calculated using the following formula:

[0088] L D (P) = -[(1-y)log(D(P) (h,w,0) )+ylog(D(P) h,w,1 )],;

[0089] Where P is the prediction result, y is the test set, y=0 is the feature information consistent with the training set, y=1 is the feature information inconsistent with the training set, h is the test set input, w is the training set input, and D is the fine recognition result;

[0090] Calculate the adversarial loss to generate more training results, using the following formula:

[0091] L adv = -log(D(P t ) (h,w,1) );

[0092] Where L adv is the loss function of the test set, and D is the fine recognition result;

[0093] Use the following formula for adversarial training optimization:

[0094] L G =L ugs +L adv ;

[0095] Where L G Optimize functions for adversarial training;

[0096] S3, accuracy verification:

[0097] S3-1, accuracy verification indicators:

[0098] Five indicators are used to evaluate accuracy, calculated using the following formulas 7 to 11:

[0099]

[0100]

[0101]

[0102] Where Pre is precision, Rec is recall, F1 is F1 score, IoU is intersection over union, OA is overall accuracy; TP, FP, TN and FN are the number of actual positive predictions as positive, the number of actual negative predictions as positive, the number of actual negative predictions as negative and the number of actual positive predictions as negative, respectively.

[0103] S3-2, accuracy verification results:

[0104] This embodiment uses a test set to verify the results (as shown in Table 1 below). The accuracy of the method of the present invention is between 93% and 98%, showing high applicability:

[0105] Pre 95.11％ Rec 94.06％ OA 97.79％ F1 94.61％ loU 93.02％

[0106] S3-3, Migration capability verification:

[0107] In order to verify the migration ability of the present invention, 251 images of Shenzhen were further selected, predictions were made in the trained deep learning model, urban green spaces were extracted and the classification accuracy was verified, and the results were shown in Table 2, where Table 2 is the migration result verified by the present invention:

[0108] Pre 96.03％ Rec 94.26％ OA 98.02％ F1 94.88％ loU 93.53％

[0109] The results show that the F1 value reaches 93.53%, indicating that the model has good robustness, and the OA value reaches 98.02%, indicating that even if the prediction area is changed, the patented method still maintains high accuracy and has good migration and generalization properties, and can be used for remote sensing image classification of large quantities of data.

[0110] The above are only preferred embodiments of the present invention, which are only used to help understand the method and core idea of ​​the present application. The protection scope of the present invention is not limited to the above embodiments. All technical solutions under the idea of ​​the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

[0111] In the description of the present invention, it should be noted that the terms "upper", "lower", "inside", "outside", etc. indicate directions or positional relationships based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore cannot be understood as a limitation on the present invention.

[0112] The present invention solves the defects of the existing technology in the identification of urban green spaces, such as the lack of high resolution and fine identification monitoring, resulting in insufficient accuracy and efficiency. By utilizing "preliminary identification" and "fine identification" as well as the prediction and extraction of deep learning models, urban green spaces can be identified more finely, especially the green spaces on both sides of the roads. It has higher accuracy and migration capabilities than the existing technology, and has a faster processing speed, providing a high-quality and efficient method for solving the deficiencies in urban greening identification.

Claims

1. A high-resolution urban green space identification method, characterized in that: The following steps are involved: S1, Satellite data processing: S1-1, satellite data download: Obtain clear and unobstructed satellite image data with a spatial resolution of 1m through remote sensing satellites and earth image databases; S1-2, sample point preparation: The built-up area is obtained through manual visual interpretation of images and urban green space sample points are generated. After cross-validation to confirm the accuracy, the sample points are randomly divided into training set, validation set and test set. S2, Deep Learning Methods: S2-1, Preliminary Identification: A feature extractor is used to capture fine features from the image and generate preliminary recognition results of urban green spaces, specifically: S2-1-1, the design consists of a 7×7 convolutional layer, a batch normalization layer, a rectified linear unit ReLU function, and a maximum pooling layer with a stride of 2; S2-1-2 uses two one-dimensional global pooling layers to encode the input features into two direction-aware features in the vertical and horizontal directions to capture the spatial coordinate information; S2-1-3, applies 4 1×1 convolution blocks to the fine features and unifies their output channels to 96; then after resizing, the fused features are input into 2 3×3 convolution layers to generate preliminary recognition results of urban green spaces; S2-2, fine identification: S2-2-1, select a sampling point with lower precision, and construct the point features of the selected sampling point with lower precision based on the preliminary recognition in S2-1 and the fine features extracted by the feature extractor; S2-2-2, reclassifying the results of the selected sampling points with lower accuracy through a multi-layer perceptron to achieve fine identification of urban green space, where N points are randomly generated and then M points with the largest uncertainty are selected from the N points; S2-2-3, construct 5 convolutional layers, and the output channels of each convolutional layer are 64, 128, 256, 512 and 1 respectively; According to the training results of M points with the largest uncertainty among the points randomly selected after reclassification in S2-2-2, the preliminary recognition results of urban green spaces are updated to obtain the refined recognition results of urban green spaces; S2-3, Deep Learning Optimization: S2-3-1, loss feature optimization function: Assuming that the initial recognition result of the urban green space in the training set is X, and the actual situation is Y, then the loss between Y and X is calculated using the following formula: Where L Dice is the loss function of the training set, X∩Y is the intersection of X and Y, |X| and |Y| represent the number of pixels in X and Y respectively; The classification result loss of N sampling points is measured by cross entropy loss, which is calculated using the following formula: Where L CE is the loss function of N sampling points, x i and i Represent the result of the i-th point among N sampling points and the ground truth respectively; on this basis, the loss characteristics of urban green space are learned from them through the deep learning framework and the loss characteristic optimization function is established, which is calculated as follows: L ugs =L Dice +L CE ; S2-3-2, adversarial training optimization: Input the test set into the trained deep learning model, and predict the loss feature optimization function of the result P, which is calculated using the following formula: L D (P)0-[(1-y)log(D(P) (h,w,0) )+ylog(D(P) h,w,1 )] ( Where P is the prediction result, y is the test set, y=0 is the feature information consistent with the training set, y=1 is the feature information inconsistent with the training set, h is the test set input, w is the training set input, and D is the fine recognition result; the adversarial loss is calculated to generate more training results, which is calculated using the following formula 5: L adv =-log(D(P) (h,w,1) ); Where L adv is the loss function of the test set, and D is the fine recognition result; Use the following formula for adversarial training optimization: L G =L ugs +L adv ; Where L G Optimize functions for adversarial training; S3, accuracy verification: S3-1, accuracy verification indicators: Five indicators are used to evaluate accuracy, calculated using the following formulas 7 to 11: Where Pre is precision, Rec is recall, F1 is F1 score, IoU is intersection over union, OA is overall accuracy; TP, FP, TN and FN are the number of actual positive predictions as positive, the number of actual negative predictions as positive, the number of actual negative predictions as negative and the number of actual positive predictions as negative, respectively.

2. The high-resolution urban green space identification method according to claim 1 is characterized in that: The specific determination of the lower accuracy in S2-2-1 is the accuracy of the last 10%.

3. The high-resolution urban green space identification method according to claim 1 is characterized in that: The feature extractor in S2-1 is ResNet.

4. The high-resolution urban green space identification method according to claim 1 is characterized in that: The specific steps of S2-2-2 are: S2-2-2-1, randomly generate N points from the preliminary recognition results, and then select M points with the largest uncertainty from the N points; S2-2-2-2, establish point features based on the sampling points obtained in S2-2-2-1, correspond to the M corresponding fine features, and obtain point features that contain both local details and global background; S2-2-2-3, input the point features into the multilayer perceptron to obtain new classification results for M points.

5. The high-resolution urban green space identification method according to claim 4 is characterized in that: The multilayer perceptron is a 1×1 convolutional layer.

6. The high-resolution urban green space identification method according to claim 5, characterized in that: The M=230, N＝1024。 7. The high-resolution urban green space identification method according to claim 1, characterized in that: The kernel size of the convolutional layer in S2-2-3 is 4 and the stride is 2.

8. The high-resolution urban green space identification method according to claim 1 is characterized in that: The proportions of the training set, validation set and test set in S1-2 are 50%, 20% and 30% respectively.

Images