Urban wetland fine classification method considering combination of different seasonal facies

Abstract

The invention relates to the technical field of environmental monitoring, in particular to an urban wetland fine classification method considering different seasonal phase combinations, which comprises the following steps: S1, establishing a wetland fine classification system suitable for urban characteristics; s2, comprehensively collecting wetland sample data, and obtaining a sample point data set; s3, adopting a remote sensing image weighted fusion method to carry out standardized preprocessing on seasonal remote sensing image data of wetland features convenient to extract in the sample point data set, and establishing a multi-seasonal image data set; s4, screening out an optimal feature set according to the multi-season image data set in combination with multi-season features; s5, performing parameter optimization on the pre-training model through the optimal feature set, inputting a multi-season image data set and a preset urban wetland feature library containing typical features of different urban wetlands, and constructing a classification model; and S6, deploying the classification model to carry out urban wetland fine classification work. According to the invention, wetland types can be subdivided rapidly and precisely, and the classification effect is good.

Description

Technical Field

[0001] The present invention relates to the technical field of environmental monitoring, and in particular to a fine classification method for urban wetlands considering different seasonal combinations. Background Art

[0002] Urban wetlands are an important part of the urban ecosystem. They not only play an important ecological function in the urban environment and provide rich ecological services, but also play a key role in maintaining the urban ecological balance, improving environmental quality, and promoting biodiversity. However, due to the acceleration of urbanization and the impact of human activities, urban wetlands are facing increasingly severe threats, including filling, pollution, and ecological environmental damage.

[0003] In order to achieve targeted protection of urban wetland resources, it is necessary to adopt optimized and refined classification schemes to obtain detailed wetland datasets. However, many land cover datasets currently only contain large-scale data at the global or national level. These datasets do not classify wetlands in detail or only divide them into two categories: water bodies and submerged vegetation. For example, the AGLC-2000-2015 global 30-meter resolution land cover change dataset developed by Xu Xiaocong et al. and the improved global land cover change monitoring platform FROM-GLC Plus developed by Le Y et al. only include two land types related to wetlands: water bodies and wetlands; the globally consistent, high-resolution and near-real-time land use and land cover classification product Dynamic World developed by Brown et al. based on deep learning and 10-meter Sentinel-2 images also only includes two wetland types: water bodies and submerged vegetation. Due to the complexity and diversity of wetland types, it has become more difficult to classify and protect them. In their study on fine wetland classification in Changde, an international wetland city, Deng Yawen et al. pointed out that there is currently a lack of fine wetland classification research for international wetland cities. At the same time, the wetland classification categories in existing studies are relatively rough and fail to fully meet the needs of wetland resource management and protection. This means that existing research cannot provide sufficient detail and accuracy to support effective wetland protection policies and decision-making. Therefore, it is crucial to design high-precision and detailed classification schemes and conduct more in-depth and refined wetland classification research.

[0004] At the same time, the existing wetland classification schemes lack comprehensive consideration of seasonal changes, resulting in inaccurate classification results. In autumn and winter, as wetland plants turn yellow and fall, their spectral characteristics become closer to water bodies or other landforms, which can easily lead to model confusion and misclassification. At the same time, seasonal precipitation and evaporation can also affect the hydrological conditions of wetlands. For example, in the dry season, wetland water levels drop or dry up, which changes the hydrological characteristics and plant growth conditions of the wetlands, making it easy for existing classification models to fail to monitor in real time, resulting in erroneous recognition results and poor recognition effects. Summary of the invention

[0005] The technical problem to be solved by the present invention is as follows: the existing wetland classification methods are relatively rough and fail to fully meet the needs of wetland resource management and protection. At the same time, when the growth cycle of wetland plants and the spectral characteristics of water bodies change significantly in different seasons, the existing wetland classification methods have low recognition accuracy and are prone to confusion and misclassification, resulting in poor recognition effect.

[0006] In order to solve the above technical problems, the present invention adopts the following technical solution: a method for fine classification of urban wetlands considering different seasonal combinations, comprising the following steps: S1: Combine the existing classification of urban wetlands with the characteristics of the study area to establish a detailed wetland classification system suitable for urban characteristics; S2: Comprehensively collect wetland sample data, mark sample points after preprocessing, and obtain sample point data sets; S3: Using the weighted fusion method of remote sensing images, the seasonal remote sensing image data for easy extraction of wetland features in the sample point data set are standardized and preprocessed to establish a multi-seasonal image data set; S4: Based on the multi-season image dataset combined with multi-season features, feature extraction, correlation and importance analysis are performed to screen out the optimal feature set; S5: After optimizing the parameters of the pre-trained model through the optimal feature set, a multi-season image dataset and a preset urban wetland feature library containing typical features of different urban wetlands are input to construct a classification model; S6: Deploy the classification model to perform fine classification of urban wetlands.

[0007] When the present invention is working, it can fuse image data of different seasons, and improve the classification ability of fuzzy areas of multi-season wetlands through the weighted fusion method of remote sensing images. The classification model constructed by this can quickly and accurately subdivide wetland types and has good robustness. At the same time, an urban wetland feature library containing typical characteristics of different urban wetlands is used for standardized input of the classification model, which can further ensure the consistency of input data and the stability of classification results. It can be applicable to the classification work of complex wetland scenes, with good classification effect and high recognition accuracy.

[0008] Preferably, in step S1, when establishing a detailed wetland classification system suitable for urban characteristics by combining the existing classification of urban wetlands with the characteristics of the study area, the following steps are adopted: A1: Based on the geographical and ecological characteristics of urban wetlands, select the existing data of the corresponding land cover remote sensing monitoring system and the Ramsar Convention classification system; A2: Based on the existing data of urban wetlands and the characteristics of the study area, a classification scheme for urban wetland land cover types was established to divide the land cover types into wetlands and non-wetlands.

[0009] Preferably, in step A2, when the land cover type is divided into wetlands and non-wetlands, the following steps are also included: wetlands are divided into natural wetlands and artificial wetlands, the natural wetland subclasses include at least one of rivers, lakes, woody swamps, herbaceous swamps and beaches, the artificial wetland subclasses include at least one of reservoirs, ponds, canals, canals and breeding ponds, and the non-wetland subclasses include at least one of woodland, construction land, cultivated land, grassland and bare land.

[0010] Preferably, in step S2, wetland sample data is comprehensively collected, sample points are marked after preprocessing, and the following steps are adopted when obtaining the sample point data set: B1: Obtain wetland sample data, follow the sampling criteria of uniform collection in small areas and global distribution, and obtain several sample points; B2: Using classification uncertainty analysis, we automatically select additional sample points in a specific area to optimize the initial sampling distribution; B3: Based on the object-level classification method, the image of the study area is divided into several objects, and the sample points are preferably covered in the sample collection area with high heterogeneity of ground objects; B4: Obtain multi-source auxiliary data containing corresponding geographic data, and use the multi-source auxiliary data to mark the collected sample points; B5: Perform category balance on the sample points, divide a part of the sample point data set into training data, and divide another part of the sample point data set into verification data.

[0011] When the present invention is working, it can realize sample point collection and sample point supplementation of specific regional points, thereby improving the comprehensiveness of the sample point data set, thereby ensuring the classification accuracy of the classification model in complex areas, and at the same time, by annotating the collected sample points through multi-source auxiliary data containing corresponding geographic data, the multi-seasonal spectral characteristics, transformation characteristics, texture characteristics and terrain characteristics of each sample point can be accurately and completely annotated, which is convenient for the construction of the classification model and can further improve its model performance.

[0012] Preferably, in step B1, following the sampling criteria of uniform collection in small areas and global distribution, when obtaining a number of sample points, the following steps are adopted: according to the geographical distribution characteristics of the study area, the study area is divided into a number of sub-areas of equal area using a grid division method, and sample points are randomly collected in each sub-area. The initial number of sample points is calculated using the following formula: ; in: is the number of initial sample points, is the total area of ​​the study area, is the area of ​​a single grid, and R is the sampling ratio of each grid; In step B2, when the classification uncertainty analysis is used to automatically select supplementary sample points in a specific area to optimize the initial sampling distribution, the classification uncertainty analysis is calculated using the following formula: ; in: is the classification uncertainty index, x is the sample point to be classified, C is the set of all classification categories, is the probability that sample x belongs to category c; In step B3, based on the object-level classification method, the image of the study area is divided into several objects. When the sample points cover the sample collection area with high heterogeneity of the ground objects, the object with the highest sample representative value is selected as the sample sampling area. The sample representative value of the object is calculated using the following formula: ; in: is the sample representative value of the object, is the weight of the sample point x, is the average spectral difference between the sample point x and other categories, is the number of sample points in the object.

[0013] Preferably, in step S3, a weighted fusion method of remote sensing images is used to standardize and preprocess the seasonal remote sensing image data for convenient extraction of wetland features in the sample point data set, and the following steps are adopted when establishing a multi-seasonal image data set: C1: Obtain the remote sensing image data of the seasonal phase that is convenient for extracting wetland features from the sample point data set, and perform pixel-level averaging on multiple remote sensing images of the same seasonal phase in multiple time periods through time series synthesis to generate long-term images that can reflect the long-term status of the seasonal phase; C2: Perform radiometric calibration on the long-term image of the corresponding seasonal phase to generate the reflectivity of the long-term image, use the preset correction algorithm to perform atmospheric correction, calculate and remove the atmospheric effect, and convert it into apparent reflectivity; C3: Remove the redundant bands during atmospheric correction, merge the remaining bands, and generate a stable comprehensive band combination; C4: Obtain multi-source auxiliary data containing corresponding geographic data, integrate the processed remote sensing image data with the corresponding radar image data in the multi-source auxiliary data through a weighted fusion method, and obtain multivariate image data; C5: After band alignment and feature fusion of multivariate image data of at least two consecutive seasons, a multi-season image dataset is established for subsequent classification model input.

[0014] When the present invention is working, it can realize the fusion of multiple remote sensing images of the same season in multiple time periods and generate long-term images that can reflect the long-term state of the season, which can reduce the subsequent workload, improve work efficiency and facilitate the extraction of typical features to accurately perform classification work. At the same time, atmospheric correction and filtering of redundant bands can reduce the impact of data dimension and noise. By fusing the remaining bands, a more stable and comprehensive band combination can be generated, which can improve the accuracy and reliability of wetland type classification. Band alignment and feature fusion are performed on multivariate image data of at least two consecutive seasons, which is convenient for building a classification model and can further improve its model performance.

[0015] Preferably, in step C4, the multivariate image data is calculated using the following formula: ; in: is the fused multi-source image data, is the processed remote sensing image data, is the corresponding radar image data in the multi-source auxiliary data, is the weight factor.

[0016] Preferably, in step S4, feature extraction, correlation and importance analysis are performed based on the multi-season image data set combined with multi-season features to screen out the optimal feature set, and the following steps are adopted: D1: Extract the original bands of the multi-season image dataset as the basis for calculating spectral features. Based on the surface reflectance of the original bands, several spectral feature indices are calculated; D2: Normalize each spectral feature index, calculate the correlation between each spectral feature, and identify redundant features; D3: Use the feature importance analysis in the preset analysis algorithm to quantify the contribution of each spectral feature to the classification effect; D4: Eliminate redundant features and select spectral features whose contribution is higher than a preset threshold to form the optimal feature set.

[0017] When the present invention is working, through the feature importance analysis in the preset analysis algorithm, the contribution of several spectral feature indexes to the classification effect can be quantified, and redundant features can be eliminated, and spectral features with contributions higher than a preset threshold are selected to form an optimal feature set, which can further improve the classification efficiency of the classification model and avoid the influence of redundant features on the classification work.

[0018] Preferably, in step S5, after optimizing the parameters of the pre-trained model through the optimal feature set, a multi-season image dataset and a preset urban wetland feature library containing typical features of different urban wetlands are input, and the following steps are adopted when constructing the classification model: E1: Optimize the parameters of the pre-trained model using the preset optimization algorithm through the optimal feature set, and select the parameter configuration with the best overall effect; E2: Construct an urban wetland feature library containing typical features of different urban wetlands and spectral features, transformation features, texture features and terrain features after Z-score standardization; E3: Input a multi-season image dataset and a preset urban wetland feature library containing typical characteristics of different urban wetlands. After the training performance is qualified, output the classification model.

[0019] Preferably, in step E1, the parameters of the pre-trained model are optimized by using a preset optimization algorithm through the optimal feature set. When selecting the parameter configuration with the best overall effect, the following steps are adopted: using k-fold cross validation, the optimal feature set is divided into several subsets, and each parameter optimization sets out some subsets as validation sets, and the random forest algorithm parameter optimization is performed several times. For several groups of parameter configurations for random forest algorithm parameter optimization, their performance on each subset is evaluated respectively, and the performance score and the average score are calculated. The parameter configuration with the best overall effect is selected, and the following formula is used to calculate the cross validation score: ; in: is the average score, is the performance score of the ith verification; In step E3, a multi-season image dataset and a preset urban wetland feature library containing typical features of different urban wetlands are input. After the training performance is qualified, when the classification model is output, the following steps are adopted to train the pre-trained model using the multi-season image dataset and the urban wetland feature library to generate a classifier. After the training performance is qualified, the classification model is output. When the classifier outputs the category label, the following formula is adopted: ; in: is the final classification result of the classifier; It is the majority voting operation function; for The prediction result given by an independent decision tree model for classifying the input sample point x.

[0020] When the present invention works, it can improve the model performance of the classification model as much as possible while avoiding overfitting of the classification model through the optimization of random forest algorithm parameters. It has strong generalization ability, high computational efficiency, and can adapt to data sets of different types and sizes, and has strong versatility. At the same time, by inputting a multi-season image data set and a preset urban wetland feature library containing typical characteristics of different urban wetlands, the trained classification model has the ability to classify wetlands in all time periods and all seasons. Whether in daily work or in complex environments, it can quickly and accurately subdivide wetland types, and has good robustness, a wide range of applications, and strong versatility.

[0021] The beneficial technical effects of the present invention include: 1. The present invention can fuse image data of different seasons, and improve the classification ability of fuzzy areas of multi-season wetlands through the weighted fusion method of remote sensing images. The classification model constructed by this can quickly and accurately subdivide wetland types and has good robustness. At the same time, an urban wetland feature library containing typical characteristics of different urban wetlands is used for standardized input of the classification model, which can further ensure the consistency of input data and the stability of classification results. It can be applied to the classification work of complex wetland scenes, with good classification effect and high recognition accuracy.

[0022] 2. The present invention can realize sample point collection and sample point supplementation in specific areas, thereby improving the comprehensiveness of the sample point data set, thereby ensuring the classification accuracy of the classification model in complex areas, and at the same time, by annotating the collected sample points with multi-source auxiliary data containing corresponding geographic data, the multi-seasonal spectral characteristics, transformation characteristics, texture characteristics and terrain characteristics of each sample point can be accurately and completely annotated, which is convenient for the construction of the classification model and can further improve its model performance.

[0023] 3. The present invention can realize the fusion of multiple remote sensing images of the same season in multiple time periods and generate long-term images that can reflect the long-term state of the season, which can reduce the subsequent workload, improve work efficiency and facilitate the extraction of typical features to accurately perform classification work. At the same time, atmospheric correction and filtering of redundant bands can reduce the impact of data dimension and noise. By fusing the remaining bands, a more stable and comprehensive band combination can be generated, which can improve the accuracy and reliability of wetland type classification. Band alignment and feature fusion are performed on multivariate image data of at least two consecutive seasons, which is convenient for building a classification model and can further improve its model performance.

[0024] 4. The present invention can quantify the contribution of several spectral feature indexes to the classification effect through feature importance analysis in a preset analysis algorithm, eliminate redundant features, and select spectral features with contributions higher than a preset threshold to form an optimal feature set, which can further improve the classification efficiency of the classification model and avoid the influence of redundant features on the classification work.

[0025] 5. The present invention can improve the model performance of the classification model as much as possible while avoiding overfitting of the classification model through random forest algorithm parameter optimization. It has strong generalization ability, high computational efficiency, and can adapt to data sets of different types and sizes. It has strong versatility. At the same time, by inputting multi-season image data sets and a preset urban wetland feature library containing typical characteristics of different urban wetlands, the trained classification model has the ability to classify wetlands in all time periods and all seasons. Whether in daily work or in complex environments, it can quickly and accurately subdivide wetland types, and has good robustness, a wide range of applications, and strong versatility.

[0026] Other features and advantages of the present invention will be disclosed in detail in the following specific embodiments and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The present invention will be further described below in conjunction with the accompanying drawings: Figure 1 The figure is a workflow diagram of a fine classification method of urban wetlands considering different seasonal combinations; Figure 2 A workflow diagram for establishing a detailed wetland classification system suitable for urban characteristics; Figure 3 Workflow diagram for constructing the optimal feature set; Figure 4 A method for fine classification of urban wetlands considering different seasonal combinations is presented. The results of fine classification of wetlands in Hangzhou in 2020 based on spring and summer fusion images are as follows; Figure 5 This is a method for fine-grained classification of urban wetlands that considers different seasonal combinations. The second result of fine-grained classification of wetlands in Hangzhou in 2020 based on spring and summer fusion images. DETAILED DESCRIPTION

[0028] The technical solutions of the embodiments of the present invention are explained and described below in conjunction with the drawings of the embodiments of the present invention, but the following embodiments are only preferred embodiments of the present invention, not all. Based on the embodiments in the implementation mode, other embodiments obtained by those skilled in the art without creative work are all within the protection scope of the present invention.

[0029] In the following description, terms such as "inside", "outside", "up", "down", "left", "right", etc. that indicate directions or positional relationships are only used to facilitate the description of the embodiments and simplify the description, and do not indicate or imply that the referred device or element must have a specific direction, be constructed and operate in a specific direction. Therefore, it should not be understood as a limitation of the present invention. Embodiment 1:

[0030] Please see attached Figure 1 This embodiment discloses a method for fine classification of urban wetlands considering different seasonal combinations, including the following steps: S1: Combine the existing classification of urban wetlands with the characteristics of the study area to establish a detailed wetland classification system suitable for urban characteristics; S2: Comprehensively collect wetland sample data, mark sample points after preprocessing, and obtain sample point data sets; S3: Using the weighted fusion method of remote sensing images, the seasonal remote sensing image data for easy extraction of wetland features in the sample point data set are standardized and preprocessed to establish a multi-seasonal image data set; S4: Based on the multi-season image dataset combined with multi-season features, feature extraction, correlation and importance analysis are performed to screen out the optimal feature set; S5: After optimizing the parameters of the pre-trained model through the optimal feature set, a multi-season image dataset and a preset urban wetland feature library containing typical features of different urban wetlands are input to construct a classification model; S6: Deploy the classification model to perform fine classification of urban wetlands.

[0031] When this embodiment is working, it can fuse image data of different seasons, and improve the classification ability of fuzzy areas of multi-season wetlands through the weighted fusion method of remote sensing images. The classification model constructed in this way can quickly and accurately subdivide wetland types and has good robustness. At the same time, an urban wetland feature library containing typical characteristics of different urban wetlands is used for standardized input of the classification model, which can further ensure the consistency of input data and the stability of classification results. It can be applied to the classification work of complex wetland scenes, with good classification effect and high recognition accuracy.

[0032] Please see attached Figure 2 and attached Figure 3 In this embodiment, in step S1, the following steps are adopted to establish a detailed wetland classification system suitable for urban characteristics by combining the existing classification of urban wetlands with the characteristics of the study area: A1: Based on the geographical and ecological characteristics of urban wetlands, select the existing data of the corresponding land cover remote sensing monitoring system and the Ramsar Convention classification system; A2: In combination with the existing data of urban wetlands and the characteristics of the study area, a classification scheme for urban wetland land cover types is established to divide the land cover types into wetlands and non-wetlands. In the specific implementation, in the step A2, when dividing the land cover types into wetlands and non-wetlands, the following steps are also included: dividing wetlands into natural wetlands and artificial wetlands, the natural wetland subclasses include at least one of rivers, lakes, woody swamps, herbaceous swamps and beaches, the artificial wetland subclasses include at least one of reservoirs, ponds, canals, ditches and breeding ponds, and the non-wetland subclasses include at least one of forest land, construction land, cultivated land, grassland and bare land. Embodiment 2:

[0033] Please refer to the attached Figure 1 This embodiment provides a method for fine classification of urban wetlands taking into account different seasonal combinations. The similarities with other aspects will not be repeated here. The differences will be described in detail below in conjunction with the accompanying drawings.

[0034] Please refer to the attached Figure 1 To Attachment Figure 5 In this embodiment, in step S2, wetland sample data is collected comprehensively, sample points are marked after preprocessing, and the following steps are adopted to obtain the sample point data set: B1: Obtain wetland sample data, follow the sampling criteria of uniform collection in small areas and global distribution, and obtain several sample points; B2: Using classification uncertainty analysis, we automatically select additional sample points in a specific area to optimize the initial sampling distribution; B3: Based on the object-level classification method, the image of the study area is divided into several objects, and the sample points are preferably covered in the sample collection area with high heterogeneity of ground objects; B4: Obtain multi-source auxiliary data containing corresponding geographic data, and use the multi-source auxiliary data to mark the collected sample points; B5: Perform category balance on the sample points, divide a part of the sample point data set into training data, and divide another part of the sample point data set into verification data.

[0035] When this embodiment is working, it can realize sample point collection and sample point supplementation of specific regional points, thereby improving the comprehensiveness of the sample point data set, thereby ensuring the classification accuracy of the classification model in complex areas, and at the same time, by annotating the collected sample points through multi-source auxiliary data containing corresponding geographic data, it is possible to accurately and completely annotate the multi-seasonal spectral characteristics, transformation characteristics, texture characteristics and terrain characteristics of each sample point, which is convenient for the construction of the classification model and can further improve its model performance.

[0036] In the specific implementation, in step B1, following the sampling criteria of uniform collection in small areas and global distribution, when obtaining a number of sample points, the following steps are adopted: according to the geographical distribution characteristics of the study area, the study area is divided into a number of sub-areas of equal area using the grid division method, and sample points are randomly collected in each sub-area. The initial number of sample points is calculated using the following formula: ; in: is the number of initial sample points, is the total area of ​​the study area, is the area of ​​a single grid, and R is the sampling ratio of each grid; In step B2, when the classification uncertainty analysis is used to automatically select supplementary sample points in a specific area to optimize the initial sampling distribution, the classification uncertainty analysis is calculated using the following formula: ; in: is the classification uncertainty index, x is the sample point to be classified, C is the set of all classification categories, is the probability that sample x belongs to category c, when When it is greater than the preset uncertainty threshold, more sample points will be collected in the specific area, thus ensuring the improvement of the classification model’s performance in complex areas; In step B3, based on the object-level classification method, the image of the study area is divided into several objects. When the sample points cover the sample collection area with high heterogeneity of the ground objects, the object with the highest sample representative value is selected as the sample sampling area. The sample representative value of the object is calculated using the following formula: ; in: is the sample representative value of the object, is the weight of the sample point x, is the average spectral difference between the sample point x and other categories, The number of sample points within the object is determined by object-level classification and dynamic weight method to screen sample points, which can optimize the distribution of sample points and ensure the robustness and accuracy of the classification model in complex wetland scenarios.

[0037] As a further improvement of this embodiment, in step B4, multi-source auxiliary data including corresponding geographic data is obtained. When the collected sample points are annotated using the multi-source auxiliary data, the multi-source auxiliary data can use appropriate databases such as Google Earth high-definition images and HydroLAKES, Chinese lake data sets, and OSM river network data to facilitate the annotation work. Preferably, the sample point annotation results can also be converted into the input format of the classification model, for example, label encoding , where C is the number of classification categories and the feature vector , which includes multi-seasonal spectral features, transformation features, texture features and terrain features.

[0038] Preferably, in step B5, when class balancing is performed on the sample points, a part of the sample point data set is divided into training data, and another part of the sample point data set is divided into verification data, the following steps are adopted to class balance the sample point data set, and the proportion of samples in each category is adjusted by undersampling and oversampling methods to ensure the balance of the model input data. Preferably, the leave-one-out method and cross-validation method can also be used to verify the representativeness and classification performance of the sample points. Embodiment three:

[0039] Please refer to the attached Figure 1 This embodiment provides a method for fine classification of urban wetlands taking into account different seasonal combinations. The similarities with other aspects will not be repeated here. The differences will be described in detail below in conjunction with the accompanying drawings.

[0040] Please refer to the attached Figure 1 To Attachment Figure 5 In this embodiment, in step S3, a weighted fusion method of remote sensing images is used to standardize and preprocess the seasonal remote sensing image data for convenient extraction of wetland features in the sample point data set. When establishing a multi-seasonal image data set, the following steps are adopted: C1: Obtain the seasonal remote sensing image data for easy extraction of wetland features in the sample point data set. Through time series synthesis, perform pixel-level average processing on multiple remote sensing images of the same seasonal phase in multiple time periods to generate long-term images that can reflect the long-term state of the seasonal phase. In specific work, the spring image data and summer image data in the remote sensing image data can be divided to ensure their temporal continuity and integrity. C2: Perform radiometric calibration on the long-term image of the corresponding seasonal phase to generate the reflectance of the long-term image. Perform atmospheric correction using a preset correction algorithm. After calculating and removing the atmospheric effect, convert it into apparent reflectance. In specific implementation, the Sen2Cor algorithm can be used for atmospheric correction. Based on the Top-of-Atmosphere (TOA) reflectance data, the atmospheric effect is calculated and removed, and converted into apparent reflectance. An automatic detection algorithm based on spectral features is applied, and parameters are set according to global and local thresholds to ensure the best cloud removal and shadow removal effects. C3: Remove the redundant bands during atmospheric correction, merge the remaining bands, and generate a stable comprehensive band combination; C4: Obtain multi-source auxiliary data containing corresponding geographic data, integrate the processed remote sensing image data with the corresponding radar image data in the multi-source auxiliary data through a weighted fusion method, and obtain multivariate image data. In actual work, the corresponding radar image data in the multi-source auxiliary data can use the VV and VH polarization data of Sentinel-1. Of course, other suitable data can also be used; C5: After band alignment and feature fusion of multivariate image data of at least two consecutive seasons, a multi-season image dataset is established for subsequent classification model input.

[0041] When this embodiment is working, it can realize the fusion of multiple remote sensing images of the same season in multiple time periods and generate long-term images that can reflect the long-term state of the season, which can reduce the subsequent workload, improve work efficiency and facilitate the extraction of typical features to accurately perform classification work. At the same time, atmospheric correction and filtering of redundant bands can reduce the impact of data dimension and noise. By fusing the remaining bands, a more stable and comprehensive band combination can be generated, which can improve the accuracy and reliability of wetland type classification. Band alignment and feature fusion are performed on multivariate image data of at least two consecutive seasons, which is convenient for building a classification model and can further improve its model performance.

[0042] As a further improvement of this embodiment, in step C4, the multivariate image data is calculated using the following formula: ; in: is the fused multi-source image data, is the processed remote sensing image data, is the corresponding radar image data in the multi-source auxiliary data, is the weight factor. Embodiment 4:

[0043] Please see attached Figure 4 This embodiment provides a method for fine classification of urban wetlands taking into account different seasonal combinations. The similarities with other aspects will not be repeated here. The differences will be described in detail below in conjunction with the accompanying drawings.

[0044] Please see attached Figure 1 In this embodiment, in step S4, feature extraction, correlation and importance analysis are performed based on the multi-season image data set combined with multi-season features to screen out the optimal feature set, and the following steps are adopted: D1: Extract the original bands of the multi-season image dataset as the basis for calculating spectral features. Based on the surface reflectance of the original bands, several spectral feature indices are calculated; D2: Normalize each spectral feature index, calculate the correlation between each spectral feature, and identify redundant features; D3: Use the feature importance analysis in the preset analysis algorithm to quantify the contribution of each spectral feature to the classification effect; D4: Eliminate redundant features and select spectral features whose contribution is higher than a preset threshold to form the optimal feature set.

[0045] When this embodiment is working, through the feature importance analysis in the preset analysis algorithm, the contribution of several spectral feature indexes to the classification effect can be quantified, and redundant features can be eliminated, and spectral features with contributions higher than a preset threshold are selected to form an optimal feature set, which can further improve the classification efficiency of the classification model and avoid the influence of redundant features on the classification work.

[0046] In the specific implementation, in the step D1, when several spectral characteristic indexes are calculated based on the surface reflectance of the original band, the following steps are adopted to calculate the traditional spectral index and the red edge spectral index based on the surface reflectance of the original band, wherein the traditional spectral index includes at least one of NDVI, MNDWI, BSI, MSAVI, LSWI, EVI and NDBI, and the red edge spectral index includes at least one of MCARI, PSSRI, NDI45, NDre1 and NDre2. The spectral characteristic index calculation formula is shown in Table 1 below: Table 1 Spectral characteristic calculation formula

[0047] Among them: B2, B3, B4, B5, B6, B7, B8, and B11 are the original bands of Sentinel-2 remote sensing images.

[0048] As a further improvement of this embodiment, in step D1, the following steps are also included: The principal component analysis and tasseled cap transformation are used to transform the corresponding image data, and the extracted features include at least one of PC1, PC2, PC3, brightness, greenness and humidity; The texture features of the image data are extracted using a gray level co-occurrence matrix, wherein the extracted features include at least one of angular second moment, contrast, correlation, variance, inverse moment, sum mean, sum variance, entropy, and entropy; A preset digital elevation dataset, such as the NASADEM 30-meter resolution digital elevation dataset, is introduced to extract the corresponding slope and altitude features.

[0049] Preferably, in step D2, each spectral feature index is normalized, the correlation between each spectral feature is calculated, and when redundant features are identified, the following calculation formula can be used: ; in: is the Pearson correlation coefficient between features, n is the sample size, and is the i-th observed value of the characteristic variable, is the mean of variable x, is the mean of the variable y. Embodiment five:

[0050] Please see attached Figure 1 This embodiment provides a method for fine classification of urban wetlands taking into account different seasonal combinations. The similarities with other aspects will not be repeated here. The differences will be described in detail below in conjunction with the accompanying drawings.

[0051] Please see attached Figure 1 In this embodiment, in step S5, after optimizing the parameters of the pre-trained model through the optimal feature set, a multi-season image dataset and a preset urban wetland feature library containing typical features of different urban wetlands are input, and the following steps are adopted when constructing the classification model: E1: Optimize the parameters of the pre-trained model using the preset optimization algorithm through the optimal feature set, and select the parameter configuration with the best overall effect; E2: Construct an urban wetland feature library containing typical features of different urban wetlands and spectral features, transformation features, texture features and terrain features after Z-score standardization; E3: Input a multi-season image dataset and a preset urban wetland feature library containing typical characteristics of different urban wetlands. After the training performance is qualified, output the classification model.

[0052] In this embodiment, in step E1, the parameters of the pre-trained model are optimized by using a preset optimization algorithm through the optimal feature set. When selecting the parameter configuration with the best overall effect, the following steps are adopted: using k-fold cross validation, the optimal feature set is divided into several subsets, and each parameter optimization is performed with some subsets reserved as validation sets, and the random forest algorithm parameter optimization is performed several times. For several groups of parameter configurations for random forest algorithm parameter optimization, their performance on each subset is evaluated respectively, and the performance score and the average score are calculated. The parameter configuration with the best overall effect is selected, and the following formula is used to calculate the cross validation score: ; in: is the average score, is the performance score of the ith verification; In step E3, a multi-season image dataset and a preset urban wetland feature library containing typical features of different urban wetlands are input. After the training performance is qualified, when the classification model is output, the following steps are adopted to train the pre-trained model using the multi-season image dataset and the urban wetland feature library to generate a classifier. After the training performance is qualified, the classification model is output. When the classifier outputs the category label, the following formula is adopted: ; in: is the final classification result of the classifier; It is the majority voting operation function; for The prediction result given by an independent decision tree model for classifying the input sample point x.

[0053] Preferably, the Z-score standardization formula is as follows: ; Where: X is the eigenvalue, μ is the characteristic mean, and σ is the characteristic standard deviation.

[0054] When this embodiment is working, by optimizing the parameters of the random forest algorithm, it is possible to improve the model performance of the classification model as much as possible while avoiding overfitting of the classification model. It has strong generalization ability, high computational efficiency, and can adapt to data sets of different types and sizes. It has strong versatility. At the same time, by inputting a multi-season image data set and a preset urban wetland feature library containing typical characteristics of different urban wetlands, the trained classification model has the ability to classify wetlands in all time periods and all seasons. Whether in daily work or in complex environments, it can quickly and accurately subdivide wetland types, and has good robustness, a wide range of applicability, and strong versatility.

[0055] Please refer to the attached Figure 4 and attached Figure 5 , when working, the types of urban wetlands in Hangzhou are finely classified. In the multi-season experimental scheme, the spring and summer combination has the best wetland fine classification results. This combination helps to capture the significant differences in vegetation and water status in these two seasons, thereby improving the classification accuracy. In the specific implementation, the remote sensing image data of Hangzhou in 2020 is input into the classification model, and the fine classification results of wetlands in Hangzhou in 2020 based on the spring and summer fusion images can be obtained. Attached Figure 5The classification results of six typical urban wetland areas in Hangzhou were magnified, including Qiandao Lake, Qingshan Reservoir, Xixi Wetland, West Lake, Beijing-Hangzhou Grand Canal and Qiantang River. It can be interpreted that Hangzhou wetlands are mainly distributed in the northeastern and southwestern regions. In the northeastern region, they are mainly distributed in Xihu District, Yuhang District and Qiantang District, and in the southwestern region, they are mainly distributed in Chun'an County. At the same time, it can also be output that among all wetland types, rivers, lakes and herbaceous swamps have the highest distribution, which are 34.9%, 20.73% and 17.64% respectively. Therefore, on the whole, Hangzhou City is mainly based on wetland types such as rivers, lakes, herbaceous swamps, etc. to complete the classification work.

[0056] The beneficial technical effects of this embodiment include: the present invention can fuse image data of different seasons, and through the method of weighted fusion of remote sensing images, improve the classification ability of fuzzy areas of multi-seasonal wetlands, and the classification model constructed thereby can quickly and accurately subdivide wetland types, and has good robustness. At the same time, an urban wetland feature library containing typical characteristics of different urban wetlands is used for standardized input of the classification model, which can further ensure the consistency of input data and the stability of classification results, and can be applied to the classification work of complex wetland scenes, with good classification effect and high recognition accuracy.

[0057] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Those skilled in the art should understand that the present invention includes but is not limited to the contents described in the drawings and the above specific embodiments. Any modification that does not deviate from the functional and structural principles of the present invention will be included in the scope of the claims.

Claims

1. A fine classification method for urban wetlands considering different seasonal combinations, characterized in that: The following steps are involved: S1: Combine the existing classification of urban wetlands with the characteristics of the study area to establish a detailed wetland classification system suitable for urban characteristics; S2: Comprehensively collect wetland sample data, mark sample points after preprocessing, and obtain sample point data sets; S3: Using the weighted fusion method of remote sensing images, the seasonal remote sensing image data for easy extraction of wetland features in the sample point data set are standardized and preprocessed to establish a multi-seasonal image data set; S4: Based on the multi-season image dataset combined with multi-season features, feature extraction, correlation and importance analysis are performed to screen out the optimal feature set; S5: After optimizing the parameters of the pre-trained model through the optimal feature set, a multi-season image dataset and a preset urban wetland feature library containing typical features of different urban wetlands are input to construct a classification model; S6: Deploy the classification model to perform fine classification of urban wetlands.

2. The method for fine classification of urban wetlands considering different seasonal combinations according to claim 1, characterized in that: In step S1, the following steps are adopted to establish a detailed wetland classification system suitable for urban characteristics by combining the existing classification of urban wetlands with the characteristics of the study area: A1: Based on the geographical and ecological characteristics of urban wetlands, select the existing data of the corresponding land cover remote sensing monitoring system and the Ramsar Convention classification system; A2: Based on the existing data of urban wetlands and the characteristics of the study area, a classification scheme for urban wetland land cover types was established to divide the land cover types into wetlands and non-wetlands.

3. The method for fine classification of urban wetlands considering different seasonal combinations according to claim 2, characterized in that: In step A2, when the land cover type is divided into wetland and non-wetland, the following steps are also included: the wetland is divided into natural wetland and artificial wetland, the natural wetland subclass includes at least one of rivers, lakes, woody swamps, herbaceous swamps and beaches, the artificial wetland subclass includes at least one of reservoirs, ponds, canals, canals and breeding ponds, and the non-wetland subclass includes at least one of forest land, construction land, cultivated land, grassland and bare land.

4. The method for fine classification of urban wetlands considering different seasonal combinations according to claim 1, characterized in that: In step S2, wetland sample data is collected comprehensively, sample points are marked after preprocessing, and the following steps are adopted to obtain the sample point data set: B1: Obtain wetland sample data, follow the sampling criteria of uniform collection in small areas and global distribution, and obtain several sample points; B2: Using classification uncertainty analysis, we automatically select additional sample points in a specific area to optimize the initial sampling distribution; B3: Based on the object-level classification method, the image of the study area is divided into several objects, and the sample points are preferably covered in the sample collection area with high heterogeneity of ground objects; B4: Obtain multi-source auxiliary data containing corresponding geographic data, and use the multi-source auxiliary data to mark the collected sample points; B5: Perform category balance on the sample points, divide a part of the sample point data set into training data, and divide another part of the sample point data set into verification data.

5. The method for fine classification of urban wetlands considering different seasonal combinations according to claim 4, characterized in that: In step B1, following the sampling criteria of uniform collection in small areas and global distribution, when obtaining a number of sample points, the following steps are adopted: according to the geographical distribution characteristics of the study area, the study area is divided into a number of sub-areas of equal area using the grid division method, and sample points are randomly collected in each sub-area. The initial number of sample points is calculated using the following formula: ； in: is the number of initial sample points, is the total area of ​​the study area, is the area of ​​a single grid, and R is the sampling ratio of each grid; In step B2, when the classification uncertainty analysis is used to automatically select supplementary sample points in a specific area to optimize the initial sampling distribution, the classification uncertainty analysis is calculated using the following formula: ； in: is the classification uncertainty index, x is the sample point to be classified, C is the set of all classification categories, is the probability that sample x belongs to category c; In step B3, based on the object-level classification method, the image of the study area is divided into several objects. When the sample points cover the sample collection area with high heterogeneity of the ground objects, the object with the highest sample representative value is selected as the sample sampling area. The sample representative value of the object is calculated using the following formula: ； in: is the sample representative value of the object, is the weight of the sample point x, is the average spectral difference between the sample point x and other categories, is the number of sample points in the object.

6. The method for fine classification of urban wetlands considering different seasonal combinations according to claim 1, characterized in that: In step S3, a weighted fusion method of remote sensing images is used to standardize and preprocess the seasonal remote sensing image data for convenient extraction of wetland features in the sample point data set. When establishing a multi-seasonal image data set, the following steps are adopted: C1: Obtain the remote sensing image data of the seasonal phase that is convenient for extracting wetland features from the sample point data set, and perform pixel-level averaging on multiple remote sensing images of the same seasonal phase in multiple time periods through time series synthesis to generate long-term images that can reflect the long-term status of the seasonal phase; C2: Perform radiometric calibration on the long-term image of the corresponding seasonal phase to generate the reflectivity of the long-term image, use the preset correction algorithm to perform atmospheric correction, calculate and remove the atmospheric effect, and convert it into apparent reflectivity; C3: Remove the redundant bands during atmospheric correction, merge the remaining bands, and generate a stable comprehensive band combination; C4: Obtain multi-source auxiliary data containing corresponding geographic data, integrate the processed remote sensing image data with the corresponding radar image data in the multi-source auxiliary data through a weighted fusion method, and obtain multivariate image data; C5: After band alignment and feature fusion of multivariate image data of at least two consecutive seasons, a multi-season image dataset is established for subsequent classification model input.

7. The method for fine classification of urban wetlands considering different seasonal combinations according to claim 6, characterized in that: In step C4, the multivariate image data is calculated using the following formula: ； in: is the fused multi-source image data, is the processed remote sensing image data, is the corresponding radar image data in the multi-source auxiliary data, is the weight factor.

8. The method for fine classification of urban wetlands considering different seasonal combinations according to claim 1, characterized in that: In step S4, feature extraction, correlation and importance analysis are performed based on the multi-season image data set combined with multi-season features to screen out the optimal feature set, and the following steps are adopted: D1: Extract the original bands of the multi-season image dataset as the basis for calculating spectral features. Based on the surface reflectance of the original bands, several spectral feature indices are calculated; D2: Normalize each spectral feature index, calculate the correlation between each spectral feature, and identify redundant features; D3: Use the feature importance analysis in the preset analysis algorithm to quantify the contribution of each spectral feature to the classification effect; D4: Eliminate redundant features and select spectral features whose contribution is higher than a preset threshold to form the optimal feature set.

9. The method for fine classification of urban wetlands considering different seasonal combinations according to claim 1, characterized in that: In step S5, after optimizing the parameters of the pre-trained model through the optimal feature set, the multi-season image dataset and the preset urban wetland feature library containing typical features of different urban wetlands are input, and the following steps are adopted when constructing the classification model: E1: Optimize the parameters of the pre-trained model using the preset optimization algorithm through the optimal feature set, and select the parameter configuration with the best overall effect; E2: The urban wetland feature library containing typical characteristics of different urban wetlands is composed of spectral features, transformation features, texture features and terrain features after Z-score standardization; E3: Input a multi-season image dataset and a preset urban wetland feature library containing typical characteristics of different urban wetlands. After the training performance is qualified, output the classification model.

10. The method for fine classification of urban wetlands considering different seasonal combinations according to claim 9, characterized in that: In step E1, the parameters of the pre-trained model are optimized by using a preset optimization algorithm through the optimal feature set. When selecting the parameter configuration with the best overall effect, the following steps are adopted: using k-fold cross validation, the optimal feature set is divided into several subsets, and each parameter optimization is performed with some subsets reserved as validation sets, and the random forest algorithm parameter optimization is performed several times. For several groups of parameter configurations for random forest algorithm parameter optimization, their performance on each subset is evaluated respectively, and the performance score and the average score are calculated. The parameter configuration with the best overall effect is selected, and the following formula is used to calculate the cross validation score: ； in: is the average score, is the performance score of the ith verification; In step E3, a multi-season image dataset and a preset urban wetland feature library containing typical features of different urban wetlands are input. After the training performance is qualified, when the classification model is output, the following steps are adopted to train the pre-trained model using the multi-season image dataset and the urban wetland feature library to generate a classifier. After the training performance is qualified, the classification model is output. When the classifier outputs the category label, the following formula is adopted: ； in: is the final classification result of the classifier; It is the majority voting operation function; for The prediction result given by an independent decision tree model for classifying the input sample point x.

Images