Ancient and famous tree identification method based on multi-source remote sensing data

Abstract

The invention discloses an ancient and famous tree identification method based on multi-source remote sensing data. According to the method, an ancient tree ecological niche model is introduced, the association degree of the actual habitat parameters of the regional ancient trees and the remote sensing features is quantified as the weight, and the feature selection process of the random forest is fused, so that the screened features have the capability of distinguishing different vegetation and meet the real living environment requirements of the ancient trees. According to the method, the problem of high distinction degree but irrelevance caused by the fact that a traditional method only depends on feature distinction degree screening is avoided, the key features directly related to the ancient tree can be more accurately captured, and the recognition pertinence of a subsequent model to the target tree species is improved; the ecological niche constrained feature screening mechanism ensures the strong relevance between the features of the input model and the ancient tree habitat from the source, the multi-modal feature fusion network is combined to comprehensively extract space, spectrum and structure information, the conditions of missing detection and false detection caused by habitat changes or new ancient trees are reduced, and the recognition effect is more stable and reliable in long-term application.

Description

Technical Field

[0001] This invention belongs to the field of ancient tree identification technology, specifically a method for identifying ancient and famous trees based on multi-source remote sensing data. Background Technology

[0002] Ancient and famous tree identification refers to the process of classifying, identifying, and recording ancient and famous trees with historical, ecological, or cultural value using modern technology. This technology integrates image recognition, botanical knowledge, remote sensing technology, and artificial intelligence algorithms, enabling efficient and accurate identification and analysis of key information such as tree species, age, shape, and growth status. In practical applications, ancient and famous tree identification typically utilizes high-resolution image acquisition equipment, drone aerial photography, and 3D modeling to obtain the external features of trees. These features are then compared and classified using deep learning models, thereby achieving automated identification and digital management of ancient and famous trees. This technology is widely used in forestry resource surveys, ecological environmental protection, the construction of archives for ancient and famous trees, and cultural heritage protection, and is of great significance for improving the protection level of ancient and famous trees and promoting ecological civilization.

[0003] However, existing technologies typically rely solely on the distinguishing ability of features during feature selection, without fully considering their correlation with the actual habitat parameters of ancient trees. This results in selected features that, while capable of distinguishing vegetation, lack specificity. Furthermore, model training often depends on fixed datasets and lacks adaptive optimization mechanisms for different regional environments and seasonal changes. In long-term applications, this can easily lead to missed or false detections due to habitat changes or the addition of new ancient trees. Summary of the Invention

[0004] The purpose of this invention is to provide a method for identifying ancient and famous trees based on multi-source remote sensing data in order to solve the problems mentioned above.

[0005] The technical solution adopted in this invention is as follows: A method for identifying ancient and famous trees based on multi-source remote sensing data, characterized in that the method includes the following steps:

[0006] S1: Simultaneously acquire optical remote sensing images, SAR synthetic aperture radar data, and hyperspectral data, focusing on canopy texture, tree structure, and fine spectral features, respectively; the acquisition time covers the growing season and the dormant season, providing a foundation for subsequent time-series analysis.

[0007] S2: Perform radiometric calibration and atmospheric correction on optical images to eliminate illumination interference; perform multi-view processing and noise suppression on SAR data to preserve tree edges; perform band redundancy removal on hyperspectral data and screen characteristic bands; record error parameters of each data to provide a correction basis for S3 spatiotemporal registration.

[0008] S3: Based on the high-precision DEM, spatial registration of three-source data is achieved through feature point matching; the time dimension is based on the optical image acquisition time, aligning the timestamps of SAR and hyperspectral data to ensure that they correspond to the same phenological period, providing input for S4 feature fusion.

[0009] S4: Extract features from preprocessed data in layers: extract canopy area and canopy shape index from optical images; extract backscattering coefficient and coherence index from SAR data; extract red edge position and chlorophyll absorption peak depth from hyperspectral data; store in a three-level system of spatial structure-physiological characteristics-physiological state to provide a candidate set for feature screening in S5;

[0010] S5: Introduce an ancient tree niche model and construct a weight matrix by combining regional ancient tree habitat parameters; use a weighted random forest algorithm to screen core features that are highly correlated with the ecological niche, reduce model complexity and enhance target specificity;

[0011] S6: Based on the features selected in S5, a spatial-spectral-structural three-branch fusion network is constructed: optical features are input into a CNN to extract spatial context, hyperspectral features are input into a fully connected layer to extract physiological indicators, and SAR features are input into an LSTM to extract temporal structural changes; the outputs of the three branches are weighted and fused through an attention mechanism to generate a comprehensive feature vector of ancient trees.

[0012] S7: Using historical tree samples as the training set, cross-validation is used to divide the training / validation sets; a dynamic learning rate strategy is introduced during training: fast convergence in the initial stage, and when the validation loss slows down, the learning rate is reduced and regularization is increased to suppress overfitting; the model parameters with the highest validation accuracy are retained.

[0013] S8: Validate in typical areas not involved in training, and count the false negative and false positive rates; if the false negative rate is high, backtrack to S5 to check the niche characteristics; if the false positive rate is high, analyze the ability of the S6 fusion network to distinguish non-target vegetation, and adjust the feature selection rules or network structure accordingly to form a validation-optimization closed loop.

[0014] S9: Deploy the model to the remote sensing monitoring platform, access real-time remote sensing data, automatically execute the S1-S8 process, and output a heat map of ancient tree distribution; collect manual verification data every six months, retrain the model to adapt to tree species renewal and environmental changes, and ensure long-term accuracy stability.

[0015] In a preferred embodiment, in step S1, the data acquisition stage requires the simultaneous acquisition of three types of remote sensing data: optical remote sensing images with a resolution of 0.5 meters to capture the texture details of the ancient tree canopy; SAR synthetic aperture radar data using VV / VH dual polarization mode to penetrate clouds and fog to obtain three-dimensional structural information of the tree; and hyperspectral data covering more than 200 bands to extract fine spectral features of the tree. The acquisition time is strictly divided into the growing season (April to October) and the dormant season (November to March of the following year) to ensure that subsequent time series analysis can compare the changes in tree characteristics at different phenological stages. The growing season data is collected every 15 days, and the dormant season data is collected every 30 days to ensure the continuity of the time series.

[0016] In a preferred embodiment, step S2 includes optical image preprocessing, which involves radiometric calibration and atmospheric correction. Radiometric calibration uses an absolute calibration method to convert pixel gray values ​​into apparent reflectance. Atmospheric correction uses a 6S model to eliminate aerosol and water vapor interference, and the radiometric error of the processed image is controlled within 5%. SAR data preprocessing uses multi-view processing, taking the number of views to 3×3 to reduce speckle noise, and combines Lee filtering to suppress random noise while retaining tree edge information. The noise suppression rate of the processed data is ≥80%. Hyperspectral data preprocessing uses a continuum removal method to screen 10 to 15 characteristic bands, prioritizing bands sensitive to vegetation such as red edge and near-infrared to remove redundant information.

[0017] In a preferred embodiment, in step S3, spatial registration is based on a high-precision DEM with a resolution of 1 meter to achieve spatial alignment of the three source data through feature point matching: optical images select canopy edge feature points, SAR data select tree outline feature points, and the SIFT feature matching algorithm is used to match the spatial position deviation of the three source data after registration is ≤1 pixel.

[0018] In a preferred embodiment, in step S4, three types of features are extracted from the preprocessed data in layers: the canopy area is extracted from optical images by calculating the number of pixels in the closed region and the canopy shape index circularity = 4π × area / perimeter 2 through an edge detection algorithm; the backscattering coefficient VV / VH polarization channel is extracted from SAR data by the average backscattering value and the coherence index phase correlation of adjacent SAR data, with a value range of 0 to 1; and the wavelength corresponding to the maximum value of the first derivative of reflectance at the red edge position and the difference between the minimum value of reflectance in the deep red band of the chlorophyll absorption peak and the average value of reflectance on both sides are extracted from hyperspectral data.

[0019] In a preferred embodiment, in step S5, firstly, the correlation between the typical habitat parameters recorded in the regional ancient tree list and these habitat parameters is quantified into weights; secondly, in the feature subset selection stage of the random forest algorithm, the weight matrix is ​​introduced to adjust the probability of the feature being selected, so that features more closely related to the ecological niche receive higher priority when the tree splits, and finally the core features that can reflect the physical characteristics of the tree and meet its ecological needs are selected.

[0020] The feature selection probability formula for weighted random forest is:

[0021]

[0022] In the formula, p j Wj represents the probability that the j-th feature is selected when the random forest node splits, Wj is the niche weight value of the j-th feature, and m is the total number of selectable features for the current node.

[0023] The calculation of Wj is based on the formula for calculating niche compatibility:

[0024]

[0025] n is the number of habitat parameters, β i Let s be the importance coefficient of the i-th habitat parameter. ij Let be the correlation coefficient between the j-th feature and the i-th habitat parameter.

[0026] In a preferred embodiment, in step S6, a three-branch fusion network is constructed based on the selected core features: the optical feature input convolutional neural network (CNN) network structure contains 3 convolutional layers with 3×3 kernels and a stride of 1 and 2 fully connected layers to extract spatial context features of the canopy and surrounding vegetation; the hyperspectral feature input fully connected layer contains 512 neurons to extract physiological indicators such as chlorophyll content; and the SAR feature input long short-term memory network (LSTM) network contains 2 hidden layers with 128 neurons per layer to extract temporal structural features of tree volume changes in different seasons.

[0027] In a preferred embodiment, in step S7, the model training uses a dataset of more than 500 historical ancient trees as the dataset and employs a 5-fold cross-validation method to divide the training set into 70% and the validation set into 30%. The training objective is to minimize the cross-entropy loss function. During the training process, a dynamic learning rate strategy is adopted: the initial learning rate is set to 0.01, and this learning rate is maintained for the first 20 rounds to achieve rapid convergence; when the change in validation set loss is less than 0.01 for 3 consecutive rounds, the learning rate is reduced to 0.001, and the L2 regularization coefficient is gradually increased from 0.005 to 0.01 to suppress overfitting; the training termination condition is that the validation set accuracy does not improve for 5 consecutive rounds or the maximum number of iterations of 200 rounds is reached, and finally, the model parameters with a validation set accuracy ≥ 92% are retained.

[0028] In a preferred embodiment, in step S8, model validation selects three typical regions: mountainous areas with an altitude of 500-1000 meters and a slope of 15-30°; plains with an altitude of <200 meters and a slope of <5°; and urban areas with a building coverage rate of >30%. For each region, 100 ancient tree samples that were not used in training are selected. Validation indicators include the false negative rate (number of unidentified ancient trees / total number of ancient trees) and the false positive rate (number of non-ancient trees mistakenly identified as ancient trees / total number of non-ancient trees). If the false negative rate is >5%, the process is backtracked to S5 to check whether key habitat parameters, such as the slope aspect information of ancient trees in mountainous areas, are missed in the niche feature screening. If the false positive rate is >5%, the process is analyzed in S6, focusing on the fusion network's ability to distinguish features of non-target vegetation, such as tall trees, by adjusting attention weights or supplementing feature samples of non-target vegetation.

[0029] In a preferred embodiment, in step S9, the optimized model is deployed to the remote sensing monitoring platform to access real-time remote sensing data once a month during the growing season and once a quarter during the dormant season. The data format is uniformly GeoTIFF. The system automatically executes processes S1 to S8 to output a heat map of ancient tree distribution. The map marks the latitude and longitude of each ancient tree, the estimated age is calculated based on the crown width and growth model, and the health status is determined by the depth of the chlorophyll absorption peak: depth <0.1 indicates sub-health and ≥0.1 indicates health.

[0030] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0031] 1. In this invention, by introducing an ancient tree niche model, the correlation between the actual habitat parameters of regional ancient trees and remote sensing features is quantified into weights and integrated into the feature selection process of random forest. This ensures that the selected features not only have the ability to distinguish different vegetation types but also conform to the actual living environment requirements of ancient trees. This avoids the problem of high discriminative power but irrelevance caused by traditional methods that rely solely on feature discriminative power for selection. In practical applications, it can more accurately capture key features directly related to ancient trees, improving the subsequent model's targeting accuracy for identifying target tree species.

[0032] 2. In this invention, the niche-constrained feature selection mechanism ensures a strong correlation between the features of the input model and the ancient tree habitat from the source. Combined with the comprehensive extraction of spatial, spectral, and structural information by a multimodal feature fusion network, and the suppression of overfitting by a dynamic learning rate training strategy, the model can maintain a high recognition accuracy even when facing complex environments in different regions and seasons. Meanwhile, the operational system iteratively updates every six months based on manually verified data, further ensuring that the model can adapt to natural tree species updates and environmental changes, reducing missed or false detections caused by habitat changes or the addition of new ancient trees. This results in more stable and reliable recognition performance in long-term applications. Attached Figure Description

[0033] Figure 1This is a schematic diagram illustrating the process principle of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0035] Example:

[0036] Reference Figure 1 ,

[0037] A method for identifying ancient and famous trees based on multi-source remote sensing data, the method includes the following steps:

[0038] S1: Simultaneously acquire optical remote sensing images, SAR synthetic aperture radar data, and hyperspectral data, focusing on canopy texture, tree structure, and fine spectral features, respectively; the acquisition time covers the growing season and the dormant season, providing a foundation for subsequent time-series analysis.

[0039] S2: Perform radiometric calibration and atmospheric correction on optical images to eliminate illumination interference; perform multi-view processing and noise suppression on SAR data to preserve tree edges; perform band redundancy removal on hyperspectral data and select characteristic bands; record error parameters of each data to provide a correction basis for S3 spatiotemporal registration.

[0040] S3: Based on the high-precision DEM, spatial registration of three-source data is achieved through feature point matching; the time dimension is based on the optical image acquisition time, aligning the SAR and hyperspectral data timestamps to ensure that they correspond to the same phenological period, providing input for S4 feature fusion.

[0041] S4: Extract features from preprocessed data in layers: extract canopy area and canopy shape index from optical images; extract backscattering coefficient and coherence index from SAR data; extract red edge position and chlorophyll absorption peak depth from hyperspectral data; store in a three-level system of spatial structure-physical characteristics-physiological state to provide a candidate set for feature screening in S5.

[0042] S5: Introduce an ancient tree niche model and construct a weight matrix by combining regional ancient tree habitat parameters; use a weighted random forest algorithm to screen core features that are highly correlated with the ecological niche, reduce model complexity and enhance target specificity.

[0043] S6: Based on the features selected in S5, a spatial-spectral-structural three-branch fusion network is constructed: optical features are input into a CNN to extract spatial context, hyperspectral features are input into a fully connected layer to extract physiological indicators, and SAR features are input into an LSTM to extract temporal structural changes; the outputs of the three branches are weighted and fused through an attention mechanism to generate a comprehensive feature vector of ancient trees.

[0044] S7: Using historical tree samples as the training set, cross-validation is used to divide the training / validation sets; a dynamic learning rate strategy is introduced during training: fast convergence in the initial stage, and when the validation loss slows down, the learning rate is reduced and regularization is increased to suppress overfitting; the model parameters with the highest validation accuracy are retained.

[0045] S8: Validate in typical areas not involved in training, and count the false negative and false positive rates; if the false negative rate is high, backtrack to S5 to check the niche characteristics; if the false positive rate is high, analyze the ability of the S6 fusion network to distinguish non-target vegetation, and adjust the feature screening rules or network structure accordingly to form a validation-optimization closed loop.

[0046] S9: Deploy the model to the remote sensing monitoring platform, access real-time remote sensing data, automatically execute the S1-S8 process, and output a heat map of ancient tree distribution; collect manual verification data every six months, retrain the model to adapt to tree species renewal and environmental changes, and ensure long-term accuracy stability.

[0047] In step S1, the data acquisition phase requires the simultaneous acquisition of three types of remote sensing data: optical remote sensing images at a resolution of 0.5 meters to capture the texture details of the ancient tree canopy; SAR synthetic aperture radar data in VV / VH dual polarization mode to penetrate clouds and fog and obtain three-dimensional structural information of the tree; and hyperspectral data covering more than 200 bands to extract fine spectral features of the tree. The acquisition time is strictly divided into the growing season (April to October) and the dormant season (November to March of the following year) to ensure that subsequent time-series analysis can compare changes in tree characteristics at different phenological stages. Growing season data is collected every 15 days, and dormant season data is collected every 30 days to ensure the continuity of the time series.

[0048] In step S2, optical image preprocessing includes radiometric calibration and atmospheric correction. Radiometric calibration uses an absolute calibration method to convert pixel gray values ​​into apparent reflectance. Atmospheric correction uses a 6S model to eliminate aerosol and water vapor interference, and the radiometric error of the processed image is controlled within 5%. SAR data preprocessing uses multi-look processing, taking 3×3 looks to reduce speckle noise, and combines Lee filtering to suppress random noise while retaining tree edge information. The noise suppression rate of the processed data is ≥80%. Hyperspectral data preprocessing uses a continuum removal method to select 10 to 15 characteristic bands, prioritizing bands sensitive to vegetation such as red edges and near-infrared to remove redundant information. During preprocessing, the geometric error parameters of each data are recorded simultaneously. The geometric error of optical images is ≤0.5 pixels, and the geometric error of SAR data is ≤1 pixel, providing a correction basis for subsequent spatiotemporal registration.

[0049] In step S3, spatial registration is performed using a high-precision DEM with a resolution of 1 meter as a reference. Spatial alignment of the three source data is achieved through feature point matching: optical images select canopy edge feature points, and SAR data selects tree contour feature points. The SIFT feature matching algorithm is used, and the spatial position deviation of the three source data after registration is ≤1 pixel. Temporal registration uses the acquisition time of the optical images as a reference and resamples the timestamps of the SAR and hyperspectral data to ensure that multi-source data from the same area correspond to the same phenological period. For example, optical images acquired in April require their corresponding SAR and hyperspectral data to be adjusted to within a time window of ±3 days from April 1st to provide spatiotemporally consistent input data for subsequent feature fusion.

[0050] In step S4, three types of features are extracted from the preprocessed data in layers: For optical images, the canopy area is extracted by calculating the number of pixels in the closed region and the canopy shape index (circularity = 4π × area / perimeter²) using an edge detection algorithm; for SAR data, the backscattering coefficient (VV / VH polarization channel average backscattering value) and coherence index are extracted, with the phase correlation of adjacent SAR data ranging from 0 to 1; for hyperspectral data, the wavelength corresponding to the maximum first derivative of reflectance at the red edge position and the depth of the chlorophyll absorption peak are extracted, along with the difference between the minimum red band reflectance and the average reflectance on both sides. All features are stored in a three-level system: spatial structure – physical characteristics – physiological state. The spatial structure layer includes canopy area and shape index; the physical characteristics layer includes backscattering coefficient and coherence index; and the physiological state layer includes red edge position and chlorophyll absorption peak depth, providing a multi-dimensional candidate set for feature selection.

[0051] In step S5, firstly, based on typical habitat parameters recorded in the regional ancient tree register (such as altitude 300–800 meters, slope ≤25°, and no tall buildings within 50 meters), an ecological niche matching index is constructed. The correlation between each remote sensing feature (such as hyperspectral chlorophyll index, SAR coherence index, etc.) and these habitat parameters is quantified into a weight. Secondly, in the feature subset selection stage of the random forest algorithm, this weight matrix is ​​introduced to adjust the probability of feature selection, so that features more closely related to the ecological niche receive higher priority during tree splitting. Finally, core features that reflect both the physical characteristics of the tree and meet its ecological needs are selected. For example, when ancient trees in a certain area are mainly distributed at altitudes of 400–600 meters, hyperspectral red-edge location features that are sensitive to altitude will be given higher weight, making them easier to retain in feature selection.

[0052] The feature selection probability formula for weighted random forest is:

[0053]

[0054] In the formula, p jWj represents the probability that the j-th feature is selected when the random forest node splits, Wj is the niche weight value of the j-th feature, and m is the total number of selectable features for the current node.

[0055] The calculation of Wj is based on the formula for calculating niche compatibility:

[0056]

[0057] n represents the number of habitat parameters (such as altitude, slope, distance from nearby buildings, etc.), β i Let s be the importance coefficient of the i-th habitat parameter (determined by expert experience or historical sample statistics). ij Let be the correlation coefficient between the j-th feature and the i-th habitat parameter;

[0058] This formula breaks through the limitations of traditional methods that rely solely on feature discrimination by embedding niche information into the feature selection mechanism of random forests. It enables the selected features to have both strong classification capabilities and to conform to the actual living environment constraints of ancient trees, thus significantly improving the targeting of subsequent identification models for the target tree species.

[0059] In step S6, a three-branch fusion network is constructed based on the selected core features: Optical features are input into a convolutional neural network (CNN) containing three convolutional layers (3×3 kernels, stride 1) and two fully connected layers to extract spatial context features of the canopy and surrounding vegetation; hyperspectral features are input into a fully connected layer containing 512 neurons to extract physiological indicators such as chlorophyll content; SAR features are input into a long short-term memory (LSTM) network containing two hidden layers (128 neurons per layer) to extract temporal structural features of tree volume changes in different seasons. The three-branch outputs are weighted and fused using an attention mechanism. The attention weights are calculated using a softmax function, with a weight sum of 1, ultimately generating a 256-dimensional comprehensive feature vector of ancient trees containing spatial, spectral, and structural information for subsequent model training.

[0060] In step S7, the model training uses a dataset of over 500 historical trees as the dataset and employs a 5-fold cross-validation method to divide the dataset into a 70% training set and a 30% validation set. The training objective is to minimize the cross-entropy loss function. A dynamic learning rate strategy is used during training: the initial learning rate is set to 0.01, and this rate is maintained for the first 20 rounds to ensure rapid convergence; when the validation set loss changes by less than 0.01 for three consecutive rounds, the learning rate is reduced to 0.001, and the L2 regularization coefficient is gradually increased from 0.005 to 0.01 to suppress overfitting. The training terminates when the validation set accuracy shows no improvement for five consecutive rounds or when the maximum number of iterations (200 rounds) is reached. Finally, model parameters with a validation set accuracy ≥ 92% are retained.

[0061] In step S8, model validation selects three typical regions: mountainous areas (altitude 500-1000 meters, slope 15-30°), plains (altitude <200 meters, slope <5°), and urban areas (building coverage >30%). For each region, 100 ancient trees not used in training are selected as samples. Validation metrics include the false negative rate (number of unidentified ancient trees / total number of ancient trees) and the false positive rate (number of non-ancient trees mistakenly identified as ancient trees / total number of non-ancient trees). If the false negative rate is >5%, the process reverts to S5 to check if key habitat parameters, such as slope aspect information for mountainous ancient trees, are missed in the niche feature selection. If the false positive rate is >5%, the S6 fusion network's ability to distinguish features of non-target vegetation, such as tall trees, is analyzed, and attention weights are adjusted or feature samples of non-target vegetation are supplemented. The optimized feature selection rules and network structure need to be retrained in S7 and validated in S8 until both the false negative rate and false positive rate are ≤5%, forming a closed-loop optimization process of validation-optimization-revalidation.

[0062] In step S9, the optimized model is deployed to the remote sensing monitoring platform. The platform accesses real-time remote sensing data monthly during the growing season and quarterly during the dormant season, with the data format uniformly set to GeoTIFF. The system automatically executes processes S1 to S8, outputting a heat map of ancient tree distribution. The map marks the latitude and longitude of each ancient tree, the estimated age calculated based on crown width and growth model, and the health status determined by the depth of chlorophyll absorption peaks (depth <0.1 indicates sub-health, ≥0.1 indicates health). Every six months, 50 to 100 trees are manually checked for verification through on-site measurements to confirm location and health status. The model is then retrained to adapt to natural tree regeneration, such as the addition of new ancient trees and environmental changes, such as changes in light conditions due to surrounding buildings blocking light, ensuring that the model's recognition accuracy remains stable above 90% over the long term.

[0063] From the above, we can conclude that:

[0064] This invention introduces an ancient tree niche model, quantifying the correlation between actual habitat parameters and remote sensing features of regional ancient trees as weights, and incorporating this into the feature selection process of a random forest. This ensures that the selected features not only have the ability to distinguish different vegetation types but also closely match the actual living environment requirements of ancient trees. This avoids the problem of high discriminative power but irrelevance caused by traditional methods that rely solely on feature discriminative power for selection. In practical applications, it can more accurately capture key features directly related to ancient trees, improving the subsequent model's targeting accuracy for identifying the target tree species.

[0065] In this invention, the niche-constrained feature selection mechanism ensures a strong correlation between the features of the input model and the ancient tree habitat from the source. Combined with the comprehensive extraction of spatial, spectral, and structural information by a multimodal feature fusion network, and the suppression of overfitting by a dynamic learning rate training strategy, the model can maintain a high recognition accuracy even when facing complex environments in different regions and seasons. Meanwhile, the operational system is iteratively updated every six months based on manually verified data, further ensuring that the model can adapt to natural tree species renewal and environmental changes, reducing missed detections and false detections caused by habitat changes or the addition of new ancient trees. This results in more stable and reliable recognition performance in long-term applications.

[0066] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the term "comprising" or any other variations thereof is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0067] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying ancient and famous trees based on multi-source remote sensing data, characterized in that: The method includes the following steps: S1: Simultaneously acquire optical remote sensing images, SAR synthetic aperture radar data, and hyperspectral data, focusing on canopy texture, tree structure, and fine spectral features, respectively; the acquisition time covers the growing season and the dormant season, providing a foundation for subsequent time-series analysis. S2: Perform radiometric calibration and atmospheric correction on optical images to eliminate illumination interference; perform multi-view processing and noise suppression on SAR data to preserve tree edges; perform band redundancy removal on hyperspectral data and screen characteristic bands; record error parameters of each data to provide a correction basis for S3 spatiotemporal registration. S3: Based on the high-precision DEM, spatial registration of three-source data is achieved through feature point matching; the time dimension is based on the optical image acquisition time, aligning the timestamps of SAR and hyperspectral data to ensure that they correspond to the same phenological period, providing input for S4 feature fusion. S4: Extract features from preprocessed data in layers: extract canopy area and canopy shape index from optical images; extract backscattering coefficient and coherence index from SAR data; extract red edge position and chlorophyll absorption peak depth from hyperspectral data; store in a three-level system of spatial structure-physiological characteristics-physiological state to provide a candidate set for feature screening in S5; S5: Introduce an ancient tree niche model and construct a weight matrix by combining regional ancient tree habitat parameters; use a weighted random forest algorithm to screen core features that are highly correlated with the ecological niche, reduce model complexity and enhance target specificity; S6: Based on the features selected in S5, a spatial-spectral-structural three-branch fusion network is constructed: optical features are input into a CNN to extract spatial context, hyperspectral features are input into a fully connected layer to extract physiological indicators, and SAR features are input into an LSTM to extract temporal structural changes; the outputs of the three branches are weighted and fused through an attention mechanism to generate a comprehensive feature vector of ancient trees. S7: Using historical tree samples as the training set, cross-validation is used to divide the training / validation sets; a dynamic learning rate strategy is introduced during training: fast convergence in the initial stage, and when the validation loss slows down, the learning rate is reduced and regularization is increased to suppress overfitting; the model parameters with the highest validation accuracy are retained. S8: Validate in typical areas not involved in training, and count the false negative and false positive rates; if the false negative rate is high, backtrack to S5 to check the niche characteristics; if the false positive rate is high, analyze the ability of the S6 fusion network to distinguish non-target vegetation, and adjust the feature selection rules or network structure accordingly to form a validation-optimization closed loop. S9: Deploy the model to the remote sensing monitoring platform, access real-time remote sensing data, automatically execute the S1-S8 process, and output a heat map of ancient tree distribution; collect manual verification data every six months, retrain the model to adapt to tree species renewal and environmental changes, and ensure long-term accuracy stability.

2. The method for identifying ancient and famous trees based on multi-source remote sensing data as described in claim 1, characterized in that: In step S1, the data acquisition phase requires the simultaneous acquisition of three types of remote sensing data: optical remote sensing images with a resolution of 0.5 meters to capture the texture details of the ancient tree canopy; SAR synthetic aperture radar data using VV / VH dual polarization mode to penetrate clouds and fog and obtain three-dimensional structural information of the tree; and hyperspectral data covering more than 200 bands to extract fine spectral features of the tree. The acquisition time is strictly divided into the growing season (April to October) and the dormant season (November to March of the following year) to ensure that subsequent time series analysis can compare the changes in tree characteristics at different phenological stages. Data from the growing season is collected every 15 days, and data from the dormant season is collected every 30 days to ensure the continuity of the time series.

3. The method for identifying ancient and famous trees based on multi-source remote sensing data as described in claim 1, characterized in that: In step S2, optical image preprocessing includes radiometric calibration and atmospheric correction. Radiometric calibration uses an absolute calibration method to convert pixel gray values ​​into apparent reflectance. Atmospheric correction uses a 6S model to eliminate aerosol and water vapor interference. The radiometric error of the processed image is controlled within 5%. SAR data preprocessing uses multi-look processing with a 3×3 look number to reduce speckle noise and combines Lee filtering to suppress random noise while retaining tree edge information. The noise suppression rate of the processed data is ≥80%. Hyperspectral data preprocessing uses a continuum removal method to screen 10 to 15 characteristic bands, prioritizing bands sensitive to vegetation such as red edge and near-infrared to remove redundant information.

4. The method for identifying ancient and famous trees based on multi-source remote sensing data as described in claim 1, characterized in that: In step S3, spatial registration is based on a high-precision DEM with a resolution of 1 meter. Spatial alignment of the three source data is achieved by feature point matching: optical images select canopy edge feature points, SAR data selects tree outline feature points. The matching algorithm uses SIFT feature matching. After registration, the spatial position deviation of the three source data is ≤1 pixel.

5. The method for identifying ancient and famous trees based on multi-source remote sensing data as described in claim 1, characterized in that: In step S4, three types of features are extracted from the preprocessed data in layers: the canopy area is extracted from optical images, and the number of pixels in the closed region is calculated using an edge detection algorithm, along with the canopy shape index (circularity = 4π × area / perimeter 2); the backscattering coefficient VV / VH polarization channel is extracted from SAR data, and the phase correlation between the backscattering coefficient and coherence index of adjacent SAR data is extracted, with the value ranging from 0 to 1; the wavelength corresponding to the maximum value of the first derivative of reflectance at the red edge position is extracted from hyperspectral data, along with the difference between the minimum reflectance value of the deep red band at the chlorophyll absorption peak and the average reflectance value on both sides.

6. The method for identifying ancient and famous trees based on multi-source remote sensing data as described in claim 1, characterized in that: In step S5, firstly, the correlation between the typical habitat parameters recorded in the regional ancient tree list and these habitat parameters is quantified into weights; secondly, in the feature subset selection stage of the random forest algorithm, this weight matrix is ​​introduced to adjust the probability of the feature being selected, so that features more closely related to the ecological niche will have higher priority when the tree splits, and finally the core features that can reflect the physical characteristics of the tree and meet its ecological needs are selected. The feature selection probability formula for weighted random forest is: In the formula, p j Wj represents the probability that the j-th feature is selected when the random forest node splits, Wj is the niche weight value of the j-th feature, and m is the total number of selectable features for the current node. The calculation of Wj is based on the formula for calculating niche compatibility: n is the number of habitat parameters, β i Let s be the importance coefficient of the i-th habitat parameter. ij Let be the correlation coefficient between the j-th feature and the i-th habitat parameter.

7. The method for identifying ancient and famous trees based on multi-source remote sensing data as described in claim 1, characterized in that: In step S6, a three-branch fusion network is constructed based on the core features selected: the optical feature input convolutional neural network (CNN) network structure contains 3 convolutional layers with 3×3 kernels, a stride of 1, and 2 fully connected layers to extract the spatial context features of the canopy and surrounding vegetation; The hyperspectral feature input fully connected layer contains 512 neurons and extracts physiological indicators such as chlorophyll content; The SAR feature input Long Short-Term Memory (LSTM) network contains two hidden layers, each with 128 neurons, to extract temporal structural features of tree volume changes in different seasons.

8. The method for identifying ancient and famous trees based on multi-source remote sensing data as described in claim 1, characterized in that: In step S7, the model training uses a dataset of over 500 historical ancient trees as the dataset and employs a 5-fold cross-validation method to divide the training set into 70% and the validation set into 30%. The training objective is to minimize the cross-entropy loss function. During training, a dynamic learning rate strategy is adopted: the initial learning rate is set to 0.01, and this learning rate is maintained for the first 20 rounds to achieve rapid convergence; when the change in validation set loss is less than 0.01 for 3 consecutive rounds, the learning rate is reduced to 0.001, and the L2 regularization coefficient is gradually increased from 0.005 to 0.01 to suppress overfitting; the training termination condition is that the validation set accuracy does not improve for 5 consecutive rounds or the maximum number of iterations of 200 rounds is reached, and finally, the model parameters with a validation set accuracy ≥ 92% are retained.

9. The method for identifying ancient and famous trees based on multi-source remote sensing data as described in claim 1, characterized in that: In step S8, model validation selects three typical regions: mountainous areas with an altitude of 500-1000 meters and a slope of 15-30°; plains with an altitude of <200 meters and a slope of <5°; and urban areas with a building coverage rate of >30%. For each region, 100 ancient tree samples that were not used in training are selected. Validation indicators include the false negative rate (number of unidentified ancient trees / total number of ancient trees) and the false positive rate (number of non-ancient trees mistakenly identified as ancient trees / total number of non-ancient trees). If the false negative rate is >5%, the process is backtracked to S5 to check whether key habitat parameters, such as the slope aspect information of ancient trees in mountainous areas, are missed in the niche feature screening. If the false positive rate is >5%, the process is analyzed in S6 to assess the ability of the fusion network to distinguish features of non-target vegetation, such as tall trees, by adjusting the attention weights or supplementing feature samples of non-target vegetation.

10. The method for identifying ancient and famous trees based on multi-source remote sensing data as described in claim 1, characterized in that: In step S9, the optimized model is deployed to the remote sensing monitoring platform and accesses real-time remote sensing data once a month during the growing season and once a quarter during the dormant season. The data format is uniformly GeoTIFF. The system automatically executes processes S1 to S8 to output a heat map of ancient tree distribution. The map marks the latitude and longitude of each ancient tree, the estimated age is calculated based on the crown width and growth model, and the health status is judged by the depth of chlorophyll absorption peak. A depth <0.1 indicates sub-health and ≥0.1 indicates health.

Images