A Systematic Monitoring and Evaluation Method for the Functional Diversity of Urban Forest Stands

Abstract

This application discloses a method for systematically monitoring and evaluating the structural and functional diversity of urban forest stands. It identifies the plant type by analyzing the ratio of leaf to leaf root length among different plants, combined with pigment information of plants of the same type. The ratio of leaf to leaf root length is a prominent feature of plant leaves, and identifying this feature through binarized images is fast and accurate. However, identifying plant types solely based on the ratio of leaf to leaf root length or plant pigment information can lead to low accuracy because some plants have similar ratios, and different plants may have similar pigment information. This application combines these two features, using the ratio of leaf to leaf root length and plant pigment information to comprehensively determine the plant type, thus greatly improving the comprehensiveness of the features and increasing the accuracy of identification.

Description

Technical Field

[0001] This invention relates to the field of forestry science and technology, specifically to a method for systematically monitoring and evaluating the structural and functional diversity of urban forest stands. Background Technology

[0002] With social development and rising levels of human civilization, living standards are improving, leading to higher demands for living environments. A healthy urban ecological climate has become a shared aspiration, and urban forests have emerged to meet this need. However, with globalization and industrialization, the concentration of greenhouse gases in the atmosphere is constantly rising, and global warming is attracting increasing attention. Forest plants, as a major component of terrestrial ecosystems, play a crucial role in the structure of urban forest stands, including composition, horizontal and vertical dimensions, age structure, and evenness. These factors are primarily determined by tree species composition, stand density, tree configuration, and tree age. Current technologies for monitoring the evenness of stand structure still rely mainly on manual monitoring. However, manual monitoring is time-consuming and labor-intensive, and it is difficult to monitor large areas of urban forests. Furthermore, the accuracy of manual monitoring is low. Summary of the Invention

[0003] The purpose of this invention is to address the problem of low accuracy in manual monitoring of forest evenness in existing technologies by proposing a systematic monitoring and evaluation method for the structural and functional diversity of urban forest stands.

[0004] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0005] A method for systematically monitoring and evaluating the structural and functional diversity of urban forest stands includes the following steps:

[0006] Step 1: Obtain the leaf image dataset and perform binarization on the leaf images in the dataset to obtain binary images;

[0007] Step 2: Obtain the leaf tip position, leaf base position, and leaf root length from the binary image. The leaf base is the connection point between the leaf blade and the leaf root. Then, obtain the ratio of the length from the leaf tip to the leaf base to the length of the leaf root.

[0008] Step 3: Based on the leaf images in the leaf image dataset, obtain the pigment information of the plant in the leaf images;

[0009] The steps for obtaining the pigment information are as follows:

[0010] Step 31: Select different fixed areas according to the type of plant, obtain all pixels in the fixed areas, and extract the grayscale values ​​of the three primary colors of all pixels;

[0011] Step 32: Obtain the number of pixels at one gray level in the three primary color gray values, and take the ratio of the number of pixels at one gray level to the total number of pixels in the plant image as the frequency of the three primary colors at one gray level in the three primary color gray values;

[0012] Step 33: Repeat step 32 to obtain the frequency of all gray levels, and obtain the gray level histogram based on the frequency of all gray levels, which is the pigment information of the plant.

[0013] Step 4: Label the types of leaves in the leaf images, and use the ratio of the length from the leaf tip to the leaf base to the length of the leaf root and the pigment information of the plant as inputs, and the type of leaf as outputs to train the neural network.

[0014] Step 5: Record the monitored area as the area to be tested, acquire images above the area to be tested and determine the boundary of the area to be tested, extract the images within the boundary of the area to be tested and record them as the image of the area to be tested;

[0015] Step 6: Use the watershed algorithm to segment the image of the region to be tested, and record each region in the segmentation result as a unit segmentation region;

[0016] Step 7: Based on the unit segmentation region, process it using the trained neural network to obtain the type of plant in the unit segmentation region;

[0017] Step 8: Obtain the plant type in all unit segmentation regions. Based on a plant type, obtain all unit segmentation regions containing that plant type and treat each region as a feature point.

[0018] Step 9: Select feature points of the same plant type and obtain the minimum bounding rectangle of all feature points, i.e., the feature point distribution map;

[0019] Step 10: Obtain the midpoint of each edge in the feature point distribution map, then connect the midpoints of two opposite edges to obtain two segmented region images. Then connect the midpoints of the other two opposite edges to obtain two more segmented region images, for a total of four region images.

[0020] Step 11: Connect the diagonals of the feature point distribution map to obtain 4 region images;

[0021] Step 12: Count the number of feature points in the 8 regional images obtained in Step 10 and Step 11, and form a vector from the number of feature points in the 8 regional images, namely the regional statistical distribution vector;

[0022] Step 13: Calculate the variance of the regional statistical distribution vector. The larger the variance, the better for monitoring and evaluating the urban forest stand structure.

[0023] Furthermore, the specific steps of step one are as follows:

[0024] Step 11: Scale the leaf image according to its aspect ratio;

[0025] Steps 1 and 2: Perform a 3×3 median filter on the scaled image;

[0026] Step 13: Perform brightness normalization on the image after median filtering so that the pixel values ​​are distributed between 0 and 255;

[0027] Step 14: Binarize the image after brightness normalization;

[0028] Step 15: Perform erosion and dilation processing on the binarized image.

[0029] Furthermore, the neural network is a Cascade-RCN network, and the Cascade-RCN network specifically comprises:

[0030] The input image is first processed by Conv1 for feature extraction, and then a feature map is generated by 3×3 max pooling.

[0031] The feature map generated by Conv1 is input into Conv2 for feature extraction, and then a feature map is generated by 3×3 max pooling.

[0032] The feature map generated by Conv2 is input into Conv3 for feature extraction, and then a feature map is generated by 3×3 max pooling.

[0033] The feature map generated by Conv3 is input into Conv4 for feature extraction, and then a feature map is generated by 3×3 max pooling.

[0034] The feature map generated by Conv4 is input into Conv5 for feature extraction, and then a feature map is generated by 3×3 max pooling.

[0035] The input feature maps in Conv2, Conv3, Conv4 and Conv5 are first reduced in dimensionality by 1×1 convolution, then by 3×3 convolution, and finally restored in dimensionality by 1×1 convolution, with an output channel number of 256.

[0036] The feature map generated by Conv2 is used as feature map A;

[0037] Feature map A is downsampled and then fused with the feature map generated by Conv3 to form feature map B;

[0038] Feature map B is downsampled and then fused with the feature map generated by Conv4 to form feature map C;

[0039] Feature map C is downsampled and then fused with the feature map generated by Conv5 to form feature map D;

[0040] Feature map D is upsampled and then fused with feature map C to form feature map CC;

[0041] Feature map CC is upsampled and then fused with feature map B to form feature map BB.

[0042] Feature map BB is upsampled and then fused with feature map A to form feature map AA.

[0043] Feature maps BB, CC, and D are input into the backbone network as common feature maps.

[0044] Feature map AA is downsampled and then fused with feature map BB to form feature map BBB;

[0045] The feature map BBB is downsampled and then fused with the feature map CC to form the feature map CCC.

[0046] The feature map CCC is downsampled and then fused with the feature map DD to form the feature map DDD.

[0047] After performing ROI pooling on feature maps AA, BBB, CCC, and DDD, the output category and location are the detection results.

[0048] Furthermore, the loss function of the Cascade-R-CNN network is CS Loss:

[0049] CS Loss=βCIoU Loss+(1β)SmoothL1Loss

[0050] Where CIoU Loss represents the crossover ratio loss function for coverage area, center point distance, and aspect ratio, and SmoothL1 represents the smoothed average absolute error loss function, with a parameter β value of 0.7.

[0051] Furthermore, the CIoU Loss is represented as:

[0052]

[0053]

[0054]

[0055] Where b represents the distance between the center point of the predicted bounding box and the center point of the ground truth bounding box, c represents the diagonal distance of the minimum closure region; ρ represents the Euclidean distance between the two center points, α represents a balancing parameter that does not participate in the gradient calculation, v represents the tilt angle of the rectangle's diagonal, and w gth represents the width of the actual bounding box. gt w represents the height of the ground truth bounding box, w represents the width of the predicted bounding box, and h represents the height of the predicted bounding box.

[0056] Furthermore, the SmoothL1Loss is represented as:

[0057]

[0058] Where X represents the numerical difference between the predicted bounding box and the ground truth bounding box, and Otherwise represents the case where |x| is not less than 1.

[0059] Furthermore, before the binarization process in step one, the step of adjusting the leaf image is included to ensure that the tilt angle of the leaf image in the horizontal or vertical direction is within 15°.

[0060] Furthermore, in step one, the leaf image is scaled according to the aspect ratio, and the height is scaled to 128 pixels.

[0061] Furthermore, the binarization process is performed using the Otsu's method.

[0062] The beneficial effects of this invention are:

[0063] This application identifies the plant type of a leaf by analyzing the ratio of leaf to leaf root length, combined with pigment information of plants within the same category. The ratio of leaf to leaf root length is a prominent feature of plant leaves, and identifying this feature through binarized images is fast and accurate. However, identifying plant types solely based on the ratio of leaf to leaf root length or plant pigment information can lead to low accuracy because some plants have similar ratios, and different plants may have similar pigment information. This application combines these two features, using the ratio of leaf to leaf root length and plant pigment information to comprehensively determine the plant type. This significantly improves the comprehensiveness of the features and increases the accuracy of identification. Furthermore, this technical solution replaces traditional manual monitoring, enabling large-area monitoring with high accuracy. Attached Figure Description

[0064] Figure 1 A schematic diagram showing the length ratio of the leaf blade to the leaf root;

[0065] Figure 2 This is a schematic diagram of a binarized image;

[0066] Figure 3 This is a schematic diagram of the binary image after erosion and dilation processing.

[0067] Figure 4This is a schematic diagram of region image segmentation. Detailed Implementation

[0068] It should be noted that, where there is no conflict, the various embodiments disclosed in this application can be combined with each other.

[0069] Specific implementation method one: Refer to Figure 1 This embodiment describes a method for systematically monitoring and evaluating the structural and functional diversity of urban forest stands, comprising the following steps:

[0070] Step 1: Obtain the leaf image dataset and perform binarization on the leaf images in the dataset to obtain binary images;

[0071] Step 2: Obtain the leaf tip position, leaf base position, and leaf root length from the binary image. The leaf base is the connection point between the leaf blade and the leaf root. Then, obtain the ratio of the length from the leaf tip to the leaf base to the length of the leaf root.

[0072] Step 3: Based on the leaf images in the leaf image dataset, obtain the pigment information of the plant in the leaf images;

[0073] The steps for obtaining the pigment information are as follows:

[0074] Step 31: Select different fixed areas according to the type of plant, obtain all pixels in the fixed areas, and extract the grayscale values ​​of the three primary colors of all pixels;

[0075] Step 32: Obtain the number of pixels at one gray level in the three primary color gray values, and take the ratio of the number of pixels at one gray level to the total number of pixels in the plant image as the frequency of the three primary colors at one gray level in the three primary color gray values;

[0076] Step 33: Repeat step 32 to obtain the frequency of all gray levels, and obtain the gray level histogram based on the frequency of all gray levels, which is the pigment information of the plant.

[0077] Step 4: Label the types of leaves in the leaf images, and use the ratio of the length from the leaf tip to the leaf base to the length of the leaf root and the pigment information of the plant as inputs, and the type of leaf as outputs to train the neural network.

[0078] Step 5: Record the monitored area as the area to be tested, acquire images above the area to be tested and determine the boundary of the area to be tested, extract the images within the boundary of the area to be tested and record them as the image of the area to be tested;

[0079] Step 6: Use the watershed algorithm to segment the image of the region to be tested, and record each region in the segmentation result as a unit segmentation region;

[0080] Step 7: Based on the unit segmentation region, process it using the trained neural network to obtain the type of plant in the unit segmentation region;

[0081] Step 8: Obtain the plant type in all unit segmentation regions. Based on a plant type, obtain all unit segmentation regions containing that plant type and treat each region as a feature point.

[0082] Step 9: Select feature points of the same plant type and obtain the minimum bounding rectangle of all feature points, i.e., the feature point distribution map;

[0083] Step 10: Obtain the midpoint of each edge in the feature point distribution map, then connect the midpoints of two opposite edges to obtain two segmented region images. Then connect the midpoints of the other two opposite edges to obtain two more segmented region images, for a total of four region images.

[0084] Step 11: Connect the diagonals of the feature point distribution map to obtain 4 region images;

[0085] Step 12: Count the number of feature points in the 8 regional images obtained in Step 10 and Step 11, and form a vector from the number of feature points in the 8 regional images, namely the regional statistical distribution vector;

[0086] Step 13: Calculate the variance of the regional statistical distribution vector. The larger the variance, the worse the uniformity; the smaller the variance, the stronger the uniformity. This is used to monitor and evaluate the stand structure of urban forests.

[0087] This application obtains pigment information of plants at different stages through the grayscale values ​​of the three primary colors, and trains a neural network by mapping the pigment information of different plant stages to the ratio of leaves to leaf roots. This allows for the identification of the plant type to which the leaves belong. The ratio of leaves to leaf roots varies among different plants, such as... Figure 1 As shown.

[0088] like Figure 4 As shown, generally speaking, when a person subjectively judges whether the distribution of feature points in an image is uniform, they mainly estimate whether the number of feature points in different regions (or directions) of the image is similar or equal. If the number of feature points in different regions of the image is similar or equal, then the distribution of feature points is uniform. Conversely, if the number of feature points in different regions of the image differs significantly, it indicates that the distribution of feature points is uneven. Therefore, this application combines this method with image recognition to evaluate and monitor the stand structure of urban forests.

[0089] Specific Implementation Method Two: This implementation method is a further explanation of Specific Implementation Method One. The difference between this implementation method and Specific Implementation Method One is that the specific steps of step one are as follows:

[0090] Step 11: Scale the leaf image according to its aspect ratio;

[0091] Steps 1 and 2: Perform a 3×3 median filter on the scaled image;

[0092] Step 13: Perform brightness normalization on the image after median filtering so that the pixel values ​​are distributed between 0 and 255;

[0093] Step 14: Binarize the image after brightness normalization;

[0094] Step 15: Perform erosion and dilation processing on the binarized image.

[0095] Binarized images typically contain small connected regions and other interference. For example... Figure 2 As shown. This application removes small connected regions and interfering factors from the binarized image through erosion and dilation processing. Figure 3 As shown in the image. This improves the accuracy of the recognition.

[0096] Specific Implementation Method Three: This implementation method is a further explanation of Specific Implementation Method One. The difference between this implementation method and Specific Implementation Method One is that the neural network is a Cascade-RCN network. Specifically, the Cascade-RCN network is as follows:

[0097] The input image is first processed by Conv1 for feature extraction, and then a feature map is generated by 3×3 max pooling.

[0098] The feature map generated by Conv1 is input into Conv2 for feature extraction, and then a feature map is generated by 3×3 max pooling.

[0099] The feature map generated by Conv2 is input into Conv3 for feature extraction, and then a feature map is generated by 3×3 max pooling.

[0100] The feature map generated by Conv3 is input into Conv4 for feature extraction, and then a feature map is generated by 3×3 max pooling.

[0101] The feature map generated by Conv4 is input into Conv5 for feature extraction, and then a feature map is generated by 3×3 max pooling.

[0102] The input feature maps in Conv2, Conv3, Conv4 and Conv5 are first reduced in dimensionality by 1×1 convolution, then by 3×3 convolution, and finally restored in dimensionality by 1×1 convolution, with an output channel number of 256.

[0103] The feature map generated by Conv2 is used as feature map A;

[0104] Feature map A is downsampled and then fused with the feature map generated by Conv3 to form feature map B;

[0105] Feature map B is downsampled and then fused with the feature map generated by Conv4 to form feature map C;

[0106] Feature map C is downsampled and then fused with the feature map generated by Conv5 to form feature map D;

[0107] Feature map D is upsampled and then fused with feature map C to form feature map CC;

[0108] Feature map CC is upsampled and then fused with feature map B to form feature map BB.

[0109] Feature map BB is upsampled and then fused with feature map A to form feature map AA.

[0110] Feature maps BB, CC, and D are input into the backbone network as common feature maps.

[0111] Feature map AA is downsampled and then fused with feature map BB to form feature map BBB;

[0112] The feature map BBB is downsampled and then fused with the feature map CC to form the feature map CCC.

[0113] The feature map CCC is downsampled and then fused with the feature map DD to form the feature map DDD.

[0114] After performing ROI pooling on feature maps AA, BBB, CCC, and DDD, the output category and location are the detection results.

[0115] To fully utilize the extracted features and improve the network's feature extraction capabilities, this application employs the concepts of residual blocks, spatial blending, and PANet to improve the Feature Pyramid Network (FPN). The FPN optimizes the features extracted by the backbone network ResNet-50, enabling the fusion of feature maps from convolutional layers of different sizes in ResNet-50. While the most basic FPN network only implements top-down unidirectional fusion, this application improves the FPN network by incorporating the bottom-up fusion concept of PANet, forming a simple bidirectional fusion. Furthermore, the residual block concept is applied to the fusion process, and the idea of ​​spatial blending of the extracted feature maps is also incorporated.

[0116] The FPN-bd network primarily improves the feature map fusion method, changing from a simple top-down downsampling fusion approach at multiple scales to a simple bidirectional upsampling and downsampling fusion approach at multiple scales. Referring to the residual concept, after spatial mixing, the features from before mixing are fused with the residual blocks to form small-scale residual blocks, thus preserving previous features and increasing feature diversity. Compared to the traditional FPN feature pyramid, the modified FPN-bd feature extraction network is more diverse, significantly improving feature fusion and extraction capabilities, and thus noticeably enhancing object detection performance.

[0117] Specific Implementation Method Four: This implementation method is a further explanation of Specific Implementation Method Three. The difference between this implementation method and Specific Implementation Method Three is that the loss function of the Cascade-R-CNN network is CS Loss.

[0118] CS Loss=βCIoU Loss+(1β)SmoothL1Loss

[0119] Where CIoU Loss represents the crossover ratio loss function for coverage area, center point distance, and aspect ratio, and SmoothL1 represents the smoothed average absolute error loss function, with a parameter β value of 0.7.

[0120] Specific Implementation Method Five: This implementation method is a further explanation of Specific Implementation Method Four. The difference between this implementation method and Specific Implementation Method Four is that the CIoU Loss is represented as follows:

[0121]

[0122]

[0123]

[0124] Where b represents the distance between the center point of the predicted bounding box and the center point of the ground truth bounding box, c represents the diagonal distance of the minimum closure region; ρ represents the Euclidean distance between the two center points, α represents a balancing parameter that does not participate in the gradient calculation, v represents the tilt angle of the rectangle's diagonal, and w gt h represents the width of the actual bounding box. gt w represents the height of the ground truth bounding box, w represents the width of the predicted bounding box, and h represents the height of the predicted bounding box.

[0125] Specific Implementation Method Six: This implementation method is a further explanation of Specific Implementation Method Five. The difference between this implementation method and Specific Implementation Method Five is that the SmoothL1Loss is represented as:

[0126]

[0127] Where X represents the numerical difference between the predicted bounding box and the ground truth bounding box, and Otherwise represents the case where |x| is not less than 1.

[0128] Specific Implementation Method Seven: This implementation method is a further explanation of Specific Implementation Method One. The difference between this implementation method and Specific Implementation Method One is that before the binarization process in step one, it also includes a step of adjusting the blade image so that the tilt angle of the blade image in the horizontal or vertical direction is within 15°.

[0129] Specific Implementation Method Eight: This implementation method is a further explanation of Specific Implementation Method Two. The difference between this implementation method and Specific Implementation Method Two is that in step one, the blade image is scaled according to the aspect ratio, and scaled to a height of 128 pixels.

[0130] Specific Implementation Method Nine: This implementation method is a further explanation of Specific Implementation Method One. The difference between this implementation method and Specific Implementation Method One is that the binarization process is performed using the Otsu's method.

[0131] It should be noted that the specific embodiments are merely explanations and illustrations of the technical solution of the present invention and should not be used to limit the scope of protection. Any modifications made in accordance with the claims and specification of the present invention that are only partial should still fall within the protection scope of the present invention.

Claims

1. A method for systematically monitoring and evaluating the structural and functional diversity of urban forest stands, characterized in that... Includes the following steps: Step 1: Obtain the leaf image dataset and perform binarization on the leaf images in the dataset to obtain binary images; Step 2: Obtain the leaf tip position, leaf base position, and leaf root length from the binary image. The leaf base is the connection point between the leaf blade and the leaf root. Then, obtain the ratio of the length from the leaf tip to the leaf base to the length of the leaf root. Step 3: Based on the leaf images in the leaf image dataset, obtain the pigment information of the plant in the leaf images; The steps for obtaining the pigment information are as follows: Step 31: Select different fixed areas according to the type of plant, obtain all pixels in the fixed areas, and extract the grayscale values ​​of the three primary colors of all pixels; Step 32: Obtain the number of pixels at one gray level in the three primary color gray values, and take the ratio of the number of pixels at one gray level to the total number of pixels in the plant image as the frequency of the three primary colors at one gray level in the three primary color gray values; Step 33: Repeat step 32 to obtain the frequency of all gray levels, and obtain the gray level histogram based on the frequency of all gray levels, which is the pigment information of the plant. Step 4: Label the types of leaves in the leaf images, and use the ratio of the length from the leaf tip to the leaf base to the length of the leaf root and the pigment information of the plant as inputs, and the type of leaf as outputs to train the neural network. Step 5: Record the monitored area as the area to be tested, acquire images above the area to be tested and determine the boundary of the area to be tested, extract the images within the boundary of the area to be tested and record them as the image of the area to be tested; Step 6: Use the watershed algorithm to segment the image of the region to be tested, and record each region in the segmentation result as a unit segmentation region; Step 7: Based on the unit segmentation region, process it using the trained neural network to obtain the type of plant in the unit segmentation region; Step 8: Obtain the plant type in all unit segmentation regions. Based on a plant type, obtain all unit segmentation regions containing that plant type and treat each region as a feature point. Step 9: Select feature points of the same plant type and obtain the minimum bounding rectangle of all feature points, i.e., the feature point distribution map; Step 10: Obtain the midpoint of each edge in the feature point distribution map, then connect the midpoints of two opposite edges to obtain two segmented region images. Then connect the midpoints of the other two opposite edges to obtain two more segmented region images, for a total of four region images. Step 11: Connect the diagonals of the feature point distribution map to obtain 4 region images; Step 12: Count the number of feature points in the 8 regional images obtained in Step 10 and Step 11, and form a vector from the number of feature points in the 8 regional images, namely the regional statistical distribution vector; Step 13: Calculate the variance of the regional statistical distribution vector to monitor and evaluate the urban forest stand structure.

2. The method for systematic monitoring and evaluation of the structural and functional diversity of urban forest stands according to claim 1, characterized in that... The specific steps of step one are as follows: Step 11: Scale the leaf image according to its aspect ratio; Steps 1 and 2: Perform a 3×3 median filter on the scaled image; Step 13: Perform brightness normalization on the image after median filtering so that the pixel values ​​are distributed between 0 and 255; Step 14: Binarize the image after brightness normalization; Step 15: Perform erosion and dilation processing on the binarized image.

3. The method for systematic monitoring and evaluation of the structural and functional diversity of urban forest stands according to claim 1, characterized in that... The neural network is a Cascade-RCNN network, and the Cascade-RCNN network is specifically as follows: The input image is first processed by Conv1 for feature extraction, and then a feature map is generated by 3×3 max pooling. The feature map generated by Conv1 is input into Conv2 for feature extraction, and then a feature map is generated by 3×3 max pooling. The feature map generated by Conv2 is input into Conv3 for feature extraction, and then a feature map is generated by 3×3 max pooling. The feature map generated by Conv3 is input into Conv4 for feature extraction, and then a feature map is generated by 3×3 max pooling. The feature map generated by Conv4 is input into Conv5 for feature extraction, and then a feature map is generated by 3×3 max pooling. The input feature maps in Conv2, Conv3, Conv4 and Conv5 are first reduced in dimensionality by 1×1 convolution, then by 3×3 convolution, and finally restored in dimensionality by 1×1 convolution, with an output channel number of 256. The feature map generated by Conv2 is used as feature map A; Feature map A is downsampled and then fused with the feature map generated by Conv3 to form feature map B; Feature map B is downsampled and then fused with the feature map generated by Conv4 to form feature map C; Feature map C is downsampled and then fused with the feature map generated by Conv5 to form feature map D; Feature map D is upsampled and then fused with feature map C to form feature map CC; Feature map CC is upsampled and then fused with feature map B to form feature map BB. Feature map BB is upsampled and then fused with feature map A to form feature map AA. Feature maps BB, CC, and D are input into the backbone network as common feature maps. Feature map AA is downsampled and then fused with feature map BB to form feature map BBB; The feature map BBB is downsampled and then fused with the feature map CC to form the feature map CCC. The feature map CCC is downsampled and then fused with the feature map DD to form the feature map DDD. After performing ROI pooling on feature maps AA, BBB, CCC, and DDD, the output category and location are the detection results.

4. The method for systematic monitoring and evaluation of the structural and functional diversity of urban forest stands according to claim 3, characterized in that... The loss function of the Cascade-rcnn network is CS Loss: CS Loss=βCIoU Loss+(1-β)SmoothL1 Loss Where CIoU Loss represents the crossover ratio loss function for coverage area, center point distance, and aspect ratio, and SmoothL1 represents the smoothed average absolute error loss function, with a parameter β value of 0.

7.

5. The method for systematic monitoring and evaluation of the structural and functional diversity of urban forest stands according to claim 4, characterized in that... The CIoU Loss is represented as: Where b represents the distance between the center point of the predicted bounding box and the center point of the ground truth bounding box, c represents the diagonal distance of the minimum closure region; ρ represents the Euclidean distance between the two center points, α represents a balancing parameter that does not participate in the gradient calculation, v represents the tilt angle of the rectangle's diagonal, and w gt h represents the width of the actual bounding box. gt w represents the height of the ground truth bounding box, w represents the width of the predicted bounding box, and h represents the height of the predicted bounding box.

6. The method for systematic monitoring and evaluation of the structural and functional diversity of urban forest stands according to claim 5, characterized in that... The SmoothL1 Loss is represented as: Where X represents the numerical difference between the predicted bounding box and the ground truth bounding box, and Otherwise represents the case where |x| is not less than 1.

7. The method for systematic monitoring and evaluation of the structural and functional diversity of urban forest stands according to claim 1, characterized in that... Before the binarization process in step one, the step also includes adjusting the leaf image so that the tilt angle of the leaf image in the horizontal or vertical direction is within 15°.

8. The method for systematic monitoring and evaluation of the structural and functional diversity of urban forest stands according to claim 2, characterized in that... In step one, the leaf image is scaled according to the aspect ratio until the height is 128 pixels.

9. The method for systematic monitoring and evaluation of the structural and functional diversity of urban forest stands according to claim 1, characterized in that... The binarization process is performed using the Otsu's method.

Images