Urban evergreen forest and deciduous forest extraction method

Abstract

The invention provides an urban evergreen forest and fallen leaf forest extraction method, which comprises the following steps of: obtaining image blocks in two seasons of late winter and early summer, and forming a data set I by all the image blocks in the two seasons; each image block comprises an urban forest land parcel and a non-urban forest land parcel; creating a label value of a non-urban forest land parcel; based on the trained random forest model, acquiring respective pseudo label data of the evergreen forest and the deciduous forest corresponding to the data set I; the label values of the non-urban forest land parcels and the label values corresponding to the pseudo label data of the evergreen forest and the deciduous forest form a data set III; training a deep learning model based on the data set I and the data set III; and based on the trained deep learning model, obtaining a probability graph of the urban evergreen forest and the deciduous forest. According to the method, the drawing precision of urban evergreen forest and deciduous forest distribution is improved.

Description

Technical Field

[0001] The invention belongs to the technical field of ecological environment monitoring, and in particular relates to a method for extracting urban evergreen forests and deciduous forests. Background Art

[0002] Countries around the world are facing the environmental impacts of climate change, with the contribution of greenhouse gases to global temperature rise receiving particular attention. Humanity's burning of fossil fuels has increased carbon sources, while deforestation has reduced carbon sinks, impacting the global carbon cycle. Estimating carbon stocks, a cornerstone of carbon cycle research, requires a critical metric to determine how much carbon is stored in land-surface vegetation. Traditional carbon accounting research has focused primarily on primary forests, while urban forests, as areas of forest use with the highest human activity, have received relatively little attention. Identifying the carbon storage of urban forests is a crucial and urgent task.

[0003] Due to their distinct physiological characteristics, deciduous and evergreen forest plots exhibit significant differences in key parameters that significantly influence carbon storage, such as photosynthetic absorption rates and carbon uptake rates. Currently, there is a lack of high-precision spatial distribution maps of deciduous and evergreen forests across large urban areas. This remains an obstacle to further refinement of research on the relationship between forest type and carbon storage, and on the impact of forest type on the carbon cycle.

[0004] Currently, there are two main methods for obtaining the spatial distribution of urban evergreen and deciduous forests: field surveys and image data extraction. The former requires a lot of manpower and material resources, and the reporting of survey data at all levels also requires a lot of time. The data obtained is usually tabular data that records the spatial location, type, and other parameters of trees and shrubs, and is converted into GIS data in the form of point features. The latter uses RGB data, multispectral data, and even hyperspectral data of the study area obtained from different platforms to identify and extract forest types. Among them, high-resolution satellite remote sensing multispectral data combines the advantages of high spatial resolution and multispectral bands, and is suitable for identifying and extracting forest types in urban scenes with high spatial heterogeneity.

[0005] Deep learning has achieved remarkable results in image segmentation in recent years, and has also shown promising results for feature segmentation in remote sensing images. This patent focuses on how to further combine point feature data from traditional forestry surveys with deep learning networks to generate large-scale, high-resolution, and highly accurate spatial distribution maps of urban evergreen and deciduous forest plots. Summary of the Invention

[0006] The purpose of this invention is to provide a method for extracting urban evergreen and deciduous forests, combining field survey point element data with high-resolution multispectral remote sensing data and using deep learning technology to improve the mapping accuracy of urban forest spatial distribution. The technical solution adopted is: A method for extracting urban evergreen forests and deciduous forests comprises the following steps: Obtain image blocks from late winter and early summer. All image blocks from the two seasons constitute Dataset 1. Each image block includes both urban forest plots and non-urban forest plots. Create label values ​​for non-urban forest plots. Based on the trained random forest model, the pseudo-label data of evergreen forest and deciduous forest corresponding to dataset one are obtained; the label values ​​of non-urban forest plots, evergreen forest, and deciduous forest corresponding to their pseudo-label data form dataset three; Based on Dataset 1 and Dataset 3, train the deep learning model; Based on the trained deep learning model, the probability map of urban evergreen forests and deciduous forests is obtained.

[0007] Preferably, the steps include: Step 1: Obtain high spatial resolution multispectral satellite remote sensing data of the city in late winter and early summer, and preprocess the data; Step 2: Based on the urban administrative boundary vector data, forest plots were selected from remote sensing images covering the city. The remote sensing data were divided into image blocks of the same size. All image blocks from the two seasons constituted Dataset 1; each image block included both urban forest plots and non-urban forest plots. Step 3: Create a label image corresponding to each image segment through visual interpretation method; The labeled images include polygonal areas of urban forest plots and polygonal areas of non-urban forest plots; all labeled images constitute dataset 2; Among them, the polygonal areas of urban forest plots and the polygonal areas of non-urban forest plots are distinguished by different label values; Step 4: Obtain a point feature vector file containing specific geographic coordinates and corresponding evergreen or deciduous forest attribute information through field sampling or visual interpretation. Use this point feature vector file to extract the band reflectance of multispectral satellite images for the two seasons, calculate the vegetation index, and record and save the results in a table. Step 5: Divide all the sample data containing the vegetation indices of the two periods constructed in Step 4 into a training set and a test set. Use the random forest model to train the vegetation indices contained in the training set to obtain the trained random forest model, and use the test set data to evaluate the model's performance. Step 6: Calculate the vegetation index using the image segmentation dataset in step 2, and classify the vegetation index using the random forest model trained in step 5 to obtain the pseudo-label data of evergreen and deciduous forests corresponding to the image segmentation dataset, as well as the confidence data corresponding to the pseudo-label data. The pseudo-label data, the polygonal area data of the non-urban forest plots in the second dataset, and the confidence data constitute the third dataset; The polygonal areas of non-urban forest plots, evergreen forest plots, and deciduous forest plots are set with different label values, each corresponding to an integer value. In addition, the confidence value contained in the dataset is a floating point value between 0 and 1, representing the degree of confidence in the type label, with 0 representing complete disbelief and 1 representing complete trust. Step 7: Build a deep learning model; Step 8: Train the deep learning model based on dataset 1 and dataset 3. Step 9: Input the pre-processed urban high-spatial-resolution multispectral satellite remote sensing image block set into the trained deep learning model to obtain the probability map of urban non-forest plots and forest plots. If it is greater than the set threshold, the urban forest spatial distribution result can be obtained; obtain the probability map of evergreen forest plots and deciduous forest plots. If it is greater than the set threshold, the evergreen forest spatial distribution result can be obtained. Multiply the two results to obtain the final urban evergreen forest and deciduous forest spatial mapping results, respectively.

[0008] Preferably, the high spatial resolution multispectral satellite remote sensing data obtained in step 1 has a spatial resolution of 0.5 meters, which is the actual distance on the ground corresponding to the pixel size of the image taken when the satellite is at an altitude of 535 kilometers from the ground, and its multispectral spectrum includes four bands: blue, green, red, and near-infrared.

[0009] Preferably, the image block size in step 2 is 1024 pixels × 1024 pixels, the actual ground range corresponding to each image block is 512 meters × 512 meters, and 500 image blocks are randomly generated in the entire data.

[0010] Preferably, step 3 specifically includes the following steps: for each image segment, professionals identify urban forest plots and outline the polygonal boundaries of the plots, and convert the polygonal vector data of the plots into labeled images corresponding to the image resolution, wherein: the label attribute of non-urban forest plots is assigned a value of 0, and the label attribute of urban forest plots is assigned a value of 1.

[0011] Preferably, step 4 specifically includes the following steps: professionals identify evergreen forest and deciduous forest plots through field surveys or visual interpretation, and save the spatial locations in the form of point features and save the forest types in the form of attribute tables, wherein the label value of deciduous forest is 1 and the label value of evergreen forest is 2.

[0012] Preferably, the vegetation index calculated in step 4 can maximally distinguish evergreen forests from deciduous forests in terms of spectral properties. Each row of the saved table data represents a point feature, with a total of three columns of data, corresponding to the vegetation index of the first period, the vegetation index of the second period, and the forest type corresponding to the point feature.

[0013] Preferably, the ratio of the training set and the test set of the point feature samples in step 5 is 6:4. The partitioning of the data set and the construction of the random forest model are implemented using the Scikit-learn library in the programming language Python 3.11.

[0014] Preferably, the type confidence in step 6 is calculated based on a random forest voting mechanism. The type confidence is the ratio of the number of decision trees supporting the type to the total number of decision trees. The label categories in Dataset 3 are 0, 1, and 2, corresponding to non-forest plots, deciduous forest plots, and evergreen forest plots, respectively.

[0015] Preferably, the programming language used in step 7 is Python 3.11, and the artificial intelligence model is written using the PyTorch library; the overall framework of the model is selected as the twin structure U-Net to realize multi-level image feature extraction in winter and summer.

[0016] Preferably, a wavelet transform module is added to the model with the twin architecture as the overall architecture to extract image texture features, and an improved deformable convolution module is used to ensure feature alignment of images from the two periods; two binary segmentation output heads are used to output the forest and non-forest segmentation results, as well as the evergreen forest and deciduous forest segmentation results respectively; finally, the results of the two are superimposed to obtain the final urban evergreen forest and deciduous forest segmentation results.

[0017] Preferably, the parameters of model training are set as follows: learning rate 0.0001, number of training rounds 200, number of input channels 4, number of output channels 1, batch size 16, number of processes 8; Adam is selected as the model optimizer, and the two decay factors are 0.9 and 0.999 respectively.

[0018] Preferably, the model is trained to predict forest type structures using the model every 50 global steps, where pixels with a probability greater than a threshold and not previously included in the pseudo-label are included in the pseudo-label for subsequent training.

[0019] Preferably, during model training, two loss functions are calculated to evaluate the forest segmentation results and the evergreen and deciduous leaf segmentation results, and are used for backpropagation to update model parameters. Both loss functions are Dice Loss. For forest segmentation, the loss is calculated using all pixels, while for evergreen and deciduous leaf segmentation, the loss is calculated using pixels with a confidence score greater than a threshold. The overall loss is the weighted sum of the two loss results, with the weights being manually set hyperparameters.

[0020] Preferably, the actual space size corresponding to the pixel in step 9 is 0.5 meters × 0.5 meters.

[0021] Compared with the prior art, the advantages of the present invention are: The method for extracting urban evergreen and deciduous forests combines field survey data with high-resolution multispectral remote sensing data. Leveraging artificial intelligence technology and the four most common multispectral bands in remote sensing imagery, the method uses a lightweight deep learning model to rapidly, accurately, and comprehensively map the spatial distribution of urban evergreen and deciduous forests at a high spatial resolution of 0.5 meters. This method overcomes the time-consuming and labor-intensive, low-precision, and incomplete nature of previous manual interpretation methods for extracting urban evergreen and deciduous forests.

[0022] 2. The method is highly reusable and can meet the future demand for extracting urban evergreen and deciduous forests from high-spatial-resolution satellite multispectral data. The highly precise extraction of urban forest types can help reveal the spatial distribution characteristics of urban forest types, the impact of urban forest types on forest carbon storage, monitor the impact of urban forest types on environmental changes at the interannual scale, and provide high-quality basic data support for subsequent quantitative research on urban forests. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 Schematic diagram of the original remote sensing image blocks and their corresponding pseudo-label images for two seasons used for model training. The green and red pseudo-labels correspond to two forest types, respectively, and the brightness and darkness represent the confidence level. Figure 2 This is a schematic diagram of the deep learning model architecture; Figure 3 Schematic diagram of the distribution of urban evergreen forests and deciduous forests in an extracted image slice. DETAILED DESCRIPTION

[0024] The following is a more detailed description of a method for extracting urban evergreen and deciduous forests according to the present invention, with reference to schematic diagrams. These schematic diagrams illustrate preferred embodiments of the present invention. It should be understood that those skilled in the art may modify the present invention described herein while still achieving the beneficial effects of the present invention. Therefore, the following description should be understood as generally known to those skilled in the art and is not intended to limit the present invention.

[0025] In this embodiment, the evergreen forest plot and the evergreen forest have the same meaning.

[0026] Deciduous forest plot and deciduous forest have the same meaning.

[0027] A method for extracting urban evergreen forests and deciduous forests comprises the following steps: Step 1: Obtain high spatial resolution multispectral satellite remote sensing data of the city in late winter and early summer, and preprocess the data.

[0028] The difference between evergreen and deciduous forests lies in the presence or absence of leaves in different seasons, so the two seasons are convenient for distinguishing these two forest types.

[0029] Late winter: January to February, early summer: May to June.

[0030] The high-spatial-resolution multispectral satellite remote sensing data obtained has a 0.5-meter resolution, which corresponds to the spatial resolution when the satellite is shot at an altitude of 535 kilometers from the earth's surface. It includes four bands: blue, green, red, and near-infrared.

[0031] Preprocessing includes radiometric calibration, atmospheric correction, terrain correction, geometric correction, mosaicking and stitching of multispectral images.

[0032] Step 2: Based on the city administrative boundary vector data, forest plots were selected from remote sensing images covering the city. The remote sensing data were divided into image blocks of the same size. All image blocks from the two seasons constituted Dataset 1. Among them, the city’s administrative boundaries are existing public data.

[0033] Both urban forest plots and non-urban forest plots were included in each image slice; Dataset 1 contains two folders, one folder stores image tiles from late winter, and the other folder stores image tiles from early summer.

[0034] In this embodiment, the image block size is 1024 pixels×1024 pixels, the actual ground range corresponding to each image block is 512 meters×512 meters, and 500 image blocks are randomly generated in the entire data.

[0035] Step 3: Create a label image corresponding to dataset one through visual interpretation method.

[0036] The labeled images include polygonal areas of urban forest plots and polygonal areas of non-urban forest plots; all labeled images constitute dataset 2.

[0037] Among them, the label values ​​of the polygonal areas of urban forest plots and the polygonal areas of non-urban forest plots are different.

[0038] The specific steps include: For each image segment, professionals identify urban forest plots and outline the vector polygon boundaries of the plots, converting the vector polygons of the plots into labeled images corresponding to the image resolution; The label value of non-urban forest plots is 0, and the label value of urban forest plots is 1.

[0039] Step 4: Obtain a point feature vector file containing specific geographic coordinates and corresponding evergreen forest or deciduous forest attribute information through field sampling or visual interpretation; That is, “point elements” include: geographic coordinate location and forest attribute information.

[0040] The point vector file was used to extract the band reflectance of the multispectral satellite images in late winter and early summer, and the vegetation index was calculated. The results were recorded and saved as a table, as described in Table 1.

[0041] The specific steps include: Professionals identify evergreen forest and deciduous forest plots through field surveys or visual interpretation, and save the spatial locations as point features and the forest types (evergreen forest, deciduous forest) in the form of attribute tables.

[0042] Among them, the label value of deciduous forest is 1, and the label value of evergreen forest is 2.

[0043] In this embodiment, the calculated vegetation index (Soil Adjusted Vegetation Index) can distinguish between evergreen forests and deciduous forests to the greatest extent in terms of spectral attributes.

[0044] As described in Table 1, each row of the saved table data represents a point feature, with a total of three columns of data, corresponding to the vegetation index of the first season (early summer), the vegetation index of the second season (late winter), and the forest type corresponding to the point feature.

[0045] Table 1 Vegetation index In Table 1, the numbers corresponding to the bands are the reflectivity of the bands.

[0046] Step 5: Divide all the sample data containing vegetation indices of the two periods constructed in Step 4 into a training set and a test set. The image block set corresponding to the sample data is independent of the first dataset.

[0047] Among them, both the training set and the test set include: vegetation index and forest type label values ​​(1 and 2); The random forest model is trained using the training set to obtain a trained random forest model, and the performance of the model is evaluated using the test set data.

[0048] In this embodiment, the ratio of the training set and the test set of the point feature samples is 6:4.

[0049] The data set partitioning and random forest model construction were both implemented using the Scikit-learn library in the programming language Python 3.11.

[0050] Step 6: Calculate the vegetation index of the evergreen forest and deciduous forest in dataset 1, and then use the random forest model trained in step 5 to classify the vegetation index to obtain the pseudo-label data and confidence data of the pseudo-label data corresponding to dataset 1. The pseudo-label data and the polygonal area data and confidence data of the non-urban forest plots in the second dataset constitute the third dataset.

[0051] The label value categories of the pseudo-label data of dataset 3 are 0, 1, and 2, corresponding to non-forest plots, deciduous forest plots, and evergreen forest plots, respectively.

[0052] Among them, the polygon area label values ​​of non-urban forest plots, evergreen forest plots, and deciduous forest plots are different, and each corresponds to an integer value.

[0053] like Figure 1 As shown, for non-forest plots, the label value of their pseudo-label data is 0, which is black; For deciduous forest plots, the pseudo-label data has a label value of 1, which is green; The evergreen forest plot, whose pseudo-label data has a label value of 2, is colored orange.

[0054] In addition, the confidence value contained in the dataset is a floating point value between 0 and 1, which represents the degree of confidence in the type label, 0 represents no belief at all, and 1 represents complete belief.

[0055] The calculation of type confidence is based on the random forest voting mechanism. The ratio of the number of decision trees supporting the type to the total number of decision trees is the type confidence.

[0056] Among them, the image blocks and pseudo-label data images of the two periods are as follows Figure 1 As shown, the red and green colors in the label image represent two forest types, and the brightness and darkness represent the confidence level.

[0057] The vegetation index of evergreen forest and deciduous forest in dataset 1 is calculated. The specific steps include: The band reflectance of the multispectral satellite image is extracted from the image blocks in dataset 1, and the vegetation index is calculated.

[0058] Step 7: Build a deep learning model.

[0059] like Figure 2 As shown in the figure, the programming language used is Python 3.11, and the artificial intelligence model is written using the PyTorch library; the overall framework of the model is selected as the twin structure U-Net to realize multi-level image feature extraction in two seasons.

[0060] A wavelet transform module is added to the model with the twin architecture as the overall architecture to extract image texture features, and an improved deformable convolution module is used to ensure feature alignment of images from the two periods.

[0061] Use two binary segmentation output heads to output forest and non-forest segmentation results and evergreen forest and deciduous forest segmentation results respectively, and superimpose the two results to obtain the final urban evergreen forest and deciduous forest segmentation results, such as Figure 3 shown.

[0062] Step 8: Train the deep learning model based on dataset 1 and dataset 3.

[0063] The parameters for model training are set as follows: learning rate 0.0001, number of training rounds 50, number of input channels 4, number of output channels 1, batch size 16, and number of processes 8; Adam is selected as the model optimizer, and the two decay factors are 0.9 and 0.999 respectively.

[0064] The model is trained to predict forest type structure every 50 global steps; Among them, pixels with a probability (confidence) greater than the threshold (0.99) are selected for subsequent training.

[0065] That is, the confidence of the labels changes dynamically during model training, and may include some labels that are considered to have low confidence in the original dataset.

[0066] The probability obtained during model training is considered to be the confidence of the result, so the probability here is equivalent to the confidence in the dataset.

[0067] During the model training process, two loss functions are calculated to evaluate the forest segmentation results and the evergreen and deciduous segmentation results, and are used for backpropagation to update the model parameters.

[0068] Both loss functions are Dice Loss. Forest segmentation uses all pixels for loss calculation, while evergreen and deciduous segmentation uses pixels with confidence greater than a threshold (0.99) for loss calculation.

[0069] The overall loss is the weighted sum of the two loss results, and the weight is a manually set hyperparameter.

[0070] In this embodiment, the weight of forest segmentation is 0.7, and the weight of evergreen and deciduous segmentation is 0.3.

[0071] like Figure 2 As shown, the pseudo-label data in dataset 3 are divided into two groups (forest / non-forest corresponds to 0, 1; deciduous / evergreen corresponds to 1, 2), which are the forest label square and forest type pseudo-label square in the figure respectively.

[0072] Step 9: Input the pre-processed high spatial resolution multispectral satellite remote sensing image block set of the city (including the image block sets of two seasons) into the trained deep learning model; Obtain the probability map of urban non-forest plots and forest plots. If the probability is greater than the set threshold, the spatial distribution result of urban forests can be obtained, that is, Figure 2 “Forest probability” in the dataset includes both urban non-forest plot data and forest plot data; Obtain the probability map of evergreen forest plots and deciduous forest plots. If the probability is greater than the set threshold, the spatial distribution result of evergreen forest plots can be obtained, that is, Figure 2 The "evergreen forest probability" in the above example includes evergreen forest plot data and deciduous forest plot data. In other embodiments, the spatial distribution result of deciduous forest can be obtained if the probability is less than a set threshold.

[0073] Multiplying the two results can obtain the final spatial mapping results of urban evergreen forests and deciduous forests respectively.

[0074] Among them, the image can be cropped into a size of 256 pixels × 256 pixels to adapt to the memory size of the graphics card device; the spatial resolution of satellite remote sensing data is the actual ground size corresponding to one pixel, which is also 0.5 meters × 0.5 meters.

[0075] like Figure 2 As shown, the forest label is 1, and the forest type pseudo labels include 1 and 2.

[0076] The above description is merely a preferred embodiment of the present invention and does not limit the present invention in any way. Any person skilled in the art who, without departing from the scope of the present invention, makes any equivalent substitution, modification, or other changes to the technical solution and technical content disclosed in the present invention shall be deemed to be within the scope of the present invention and still fall within the scope of protection of the present invention.

Claims

1. A method for extracting urban evergreen forests and deciduous forests, characterized in that: The following steps are involved: Obtain image blocks from late winter and early summer. All image blocks from the two seasons constitute Dataset 1. Each image block includes both urban forest plots and non-urban forest plots. Create label values ​​for non-urban forest plots. Based on the trained random forest model, the pseudo-label data of evergreen forest and deciduous forest corresponding to dataset one are obtained; the label values ​​of non-urban forest plots, evergreen forest, and deciduous forest corresponding to their pseudo-label data form dataset three; Based on Dataset 1 and Dataset 3, train the deep learning model; Based on the trained deep learning model, the probability map of urban evergreen forests and deciduous forests is obtained.

2. The method for extracting urban evergreen forests and deciduous forests according to claim 1, wherein: The specific steps include: Step 1: Obtain high spatial resolution multispectral satellite remote sensing data of the city in late winter and early summer, and preprocess the data; Step 2: Select forest plots on remote sensing images covering the city and divide the remote sensing data into image blocks of the same size. All image blocks from the two seasons constitute Dataset 1. Both urban forest plots and non-urban forest plots were included in each image slice; Step 3: Create a label image corresponding to each image segment through visual interpretation method; The labeled images include polygonal areas of urban forest plots and polygonal areas of non-urban forest plots; All labeled images constitute dataset 2; Step 4: extract the band reflectance of multispectral satellite images in late winter and early summer, and calculate the vegetation index of evergreen forests and deciduous forests; Step 5: Divide all sample data of vegetation indices of two seasons constructed in step 4 to obtain a training set; Use the training set to train the vegetation index included in the random forest model to obtain a trained random forest model; Step 6: Calculate the vegetation index of evergreen forests and deciduous forests in dataset 1, and then use the random forest model trained in step 5 to classify the vegetation index to obtain the pseudo-label data of evergreen and deciduous forests corresponding to dataset 1, as well as the confidence data corresponding to the pseudo-label data; Pseudo-label data, polygonal area data of non-urban forest plots in dataset 2, and confidence data constitute dataset 3; Step 7: Build a deep learning model; Step 8: Train the deep learning model based on dataset 1 and dataset 3. Step 9: Input the pre-processed urban high spatial resolution multispectral satellite remote sensing image slice set into the trained deep learning model to obtain the probability map of urban non-forest plots and forest plots, and the probability map of evergreen forest plots and deciduous forest plots; Multiplying the two results gives the probability map of urban evergreen and deciduous forests.

3. The method for extracting urban evergreen forests and deciduous forests according to claim 2, wherein: The high-spatial-resolution multispectral satellite remote sensing data obtained has a 0.5-meter resolution, which corresponds to the spatial resolution when the satellite is shot at an altitude of 535 kilometers from the earth's surface. It includes four bands: blue, green, red, and near-infrared.

4. The method for extracting urban evergreen forests and deciduous forests according to claim 2, wherein: Step 3 specifically includes the following steps: for each image segment, professionals identify urban forest plots and outline the vector polygon boundaries of the plots, and convert the vector polygons of the plots into labeled images corresponding to the image resolution, where the label value of non-urban forest plots is 0 and the label value of urban forest plots is 1.

5. The method for extracting urban evergreen forests and deciduous forests according to claim 2, wherein: Step 4 specifically includes the following steps: professionals identify evergreen forest and deciduous forest plots through field surveys or visual interpretation, save the spatial locations as point features, and save the forest types in the form of an attribute table, where the label value of deciduous forest is 1 and the label value of evergreen forest is 2.

6. The method for extracting urban evergreen forests and deciduous forests according to claim 2, wherein: The vegetation index calculated in step 4 can maximize the distinction between evergreen forests and deciduous forests in terms of spectral properties.

7. The method for extracting urban evergreen forests and deciduous forests according to claim 2, wherein: The label values ​​of dataset 3 include 0, 1, and 2, corresponding to non-forest plots, deciduous forest plots, and evergreen forest plots, respectively.

8. The method for extracting urban evergreen forests and deciduous forests according to claim 2, wherein: The programming language used in step 7 is Python 3.11, and the artificial intelligence model is written using the PyTorch library; The overall framework of the model is chosen as the U-Net with a twin structure to realize multi-level image feature extraction for two seasons.

9. The method for extracting urban evergreen forests and deciduous forests according to claim 8, characterized in that: A wavelet transform module is added to the model with the twin architecture to extract image texture features, and an improved deformable convolution module is used to ensure feature alignment of images from the two periods. Two binary segmentation output heads are used to output forest-non-forest segmentation results and evergreen forest and deciduous forest segmentation results respectively. The results of the two are superimposed to obtain the final urban evergreen forest and deciduous forest segmentation results.

10. The method for extracting urban evergreen forests and deciduous forests according to claim 1, characterized in that: In step 8, two loss functions are calculated during model training to evaluate the forest segmentation results and the evergreen and deciduous segmentation results, and are used for backpropagation to update the model parameters; Both loss functions are Dice Loss. Forest segmentation uses all pixels for loss calculation, while evergreen and deciduous segmentation uses pixels with confidence greater than the threshold for loss calculation. The overall loss is the weighted sum of the two loss results, and the weight is a manually set hyperparameter.

Images