A tree species identification method for urban fragmented green space and related equipment

Abstract

The application discloses a kind of urban fragmented green space-oriented tree species identification method and related equipment, the method is by constructing the high-frequency unmanned aerial vehicle observation dataset covering complete autumn phenology process, fusion RGB image, canopy height model and phenology timing information, establish "spectrum-structure-timing" multidimensional feature expression system.On this basis, design end-to-end multimodal timing fusion model, realize the collaborative modeling of heterogeneous data in space and time dimensions.And by introducing the combination loss function of weighted cross-entropy and Dice, effectively alleviate the class imbalance problem, improve the rare tree species identification ability.The application significantly improves the tree species identification precision and model robustness, can realize pixel-level high-precision identification in complex urban green space environment, provides technical support for urban green space resource management and ecological assessment.

Description

Technical Field

[0001] This specification relates to the field of image recognition, and more specifically, this application relates to a method and related equipment for identifying tree species in fragmented urban green spaces. Background Technology

[0002] Urban green spaces, as an important component of urban ecosystems, play a vital role in carbon sequestration, microclimate regulation, biodiversity maintenance, and improving residents' quality of life. However, with the accelerating pace of urbanization, urban green spaces are increasingly characterized by spatial fragmentation and structural complexity. Traditional tree species identification methods, which rely on manual field surveys and visual interpretation, suffer from low efficiency, high costs, and difficulty in achieving large-scale dynamic monitoring, making them insufficient to meet the needs of refined urban ecological management.

[0003] In recent years, the development of remote sensing technology has provided new technical means for urban tree species identification. Unmanned aerial vehicle (UAV) remote sensing, due to its advantages such as high spatial resolution and flexible acquisition, has gradually become an important data source for urban green space monitoring. By acquiring high-resolution RGB imagery and canopy height models, vegetation characteristics can be characterized from both spectral and structural information perspectives. Simultaneously, multi-temporal remote sensing data can reflect plant phenological changes, providing important evidence for distinguishing spectrally similar tree species.

[0004] However, existing technologies still have certain shortcomings. On the one hand, most studies use single-phase or limited-phase data, which makes it difficult to fully characterize the dynamic phenological characteristics of tree species. On the other hand, the methods for fusing RGB imagery with multimodal data such as CHM are still relatively simple and lack the ability to collaboratively model temporal and spatial information. In addition, existing models are mostly "black box" structures, lacking systematic analysis of the role of different modal information in tree species identification, which limits further improvement in identification accuracy and model interpretability.

[0005] Therefore, there is an urgent need for a method for identifying urban tree species that can integrate multimodal and multi-temporal information and has good interpretability. Summary of the Invention

[0006] The summary section introduces a series of simplified concepts, which will be further explained in detail in the detailed description section. This summary section is not intended to limit the key and essential technical features of the claimed technical solution, nor is it intended to determine the scope of protection of the claimed technical solution.

[0007] Firstly, this application proposes a tree species identification method for fragmented urban green spaces, including: Acquire remote sensing image data of the study area within a preset phenological cycle and simultaneously generate canopy height model data to construct a multimodal dataset containing spectral, structural, and time-series information; Geometric correction, temporal registration, and spatial alignment were performed on the above multimodal dataset to obtain pixel-level aligned multimodal temporal data; A training sample dataset is constructed by performing pixel-level annotations on the aligned multimodal time-series data; A multimodal temporal deep learning model is constructed to extract features and fuse them into a model for the aforementioned multimodal temporal data. The trained multimodal temporal deep learning model performs inference calculations on the region to be identified and outputs pixel-level tree species classification results.

[0008] In one feasible implementation, the above-mentioned acquisition of multiple remote sensing image data of the study area within a preset phenological cycle, and simultaneous generation of canopy height model data, are used to construct a multimodal dataset containing spectral information, structural information, and time series information, including: Using a drone platform, multiple aerial photography missions were carried out within the aforementioned preset phenological cycle at a preset sampling frequency to acquire the aforementioned multi-period remote sensing image data, wherein the aforementioned multi-period remote sensing image data are multi-period RGB image data; Three-dimensional point cloud data is generated based on the above-mentioned multi-period RGB image data, and the above-mentioned three-dimensional point cloud data is classified into ground points and vegetation points. A digital elevation model is constructed based on the classified ground points, and a digital surface model is constructed based on the aforementioned vegetation points. The canopy height model is obtained by performing pixel-by-pixel difference calculations on the above-mentioned digital surface model and the above-mentioned digital elevation model. The aforementioned canopy height model was paired with the aforementioned multi-period RGB image data in chronological order to construct the aforementioned multimodal dataset.

[0009] In one feasible implementation, the above-mentioned geometric correction, temporal registration, and spatial alignment processing of the multimodal dataset yields pixel-level aligned multimodal temporal data, including: Geometric fine correction was performed on the aforementioned multi-phase remote sensing image data based on ground control points to eliminate spatial positioning errors in the aforementioned multi-phase remote sensing image data; Using a preset reference image as a reference, a feature point extraction and matching algorithm is used to perform time-series registration of the above-mentioned multi-period remote sensing image data, and an abnormal matching point is removed by a mismatch elimination algorithm. Based on the registration results, the remote sensing image data of the above multiple periods are resampled to ensure that the images of each period are consistent in spatial location. The canopy height model was resampled at a preset resolution to ensure that the canopy height model and the multi-period remote sensing image data were consistent in spatial resolution. Based on the above processing results, the above multi-period remote sensing image data and the above canopy height model are aligned at the pixel level in the time and spatial dimensions to obtain the above multimodal time series data.

[0010] In one feasible implementation, the above-mentioned construction of a multimodal temporal deep learning model, and the feature extraction and fusion modeling of the above-mentioned multimodal temporal data, includes: The above-mentioned multi-period remote sensing image data at each time point are fused with the above-mentioned canopy height model in the channel dimension to form multi-channel input data; The above multi-channel input data is normalized to construct a multimodal temporal input tensor containing the time dimension; The multimodal temporal input tensor input encoder module described above uses a Transformer-based backbone network to extract multi-scale spatial features. A time-adaptor module is introduced into the encoder module to model the multi-scale spatial features in the time dimension, so as to capture the feature dependencies between different time steps. The features processed by the aforementioned timing adapter module are fused and then input into the decoder module for upsampling and feature fusion to generate pixel-level semantic segmentation results.

[0011] In one feasible implementation, the features processed by the timing adapter module are fused and then input into the decoder module for upsampling and feature fusion to generate pixel-level semantic segmentation results, including: The multi-scale features output by the above timing adapter module are uniformly mapped according to the preset number of channels; The multi-scale features are upsampled based on the multi-scale feature fusion structure, and high-level semantic features are fused with low-level detail features. Contextual information enhancement processing is introduced during the feature fusion process to supplement the fused features with global information. The enhanced fusion features are input into the classification layer, and the probability distribution map of tree species corresponding to each pixel is output. The pixel-level semantic segmentation results are determined based on the above tree species category probability distribution map.

[0012] In one feasible implementation, the above-mentioned multimodal temporal deep learning model training process includes: Based on the above training sample dataset, construct training input data and corresponding label data; A combined loss function of weighted cross-entropy loss and Dice loss is constructed to alleviate the class imbalance problem; Set the category weight parameters in the weighted cross-entropy loss function based on the pixel proportion of each category; The multimodal temporal deep learning model is trained end-to-end based on the above combined loss function; During training, a preset optimizer is used to update the model parameters, and a preset learning rate adjustment strategy is used to dynamically adjust the learning rate. In the training process described above, auxiliary branch outputs are introduced to enhance gradient propagation and improve model convergence stability.

[0013] In one feasible implementation, the above-mentioned multimodal temporal deep learning model, based on the trained model, performs inference calculations on the region to be identified and outputs pixel-level tree species classification results, including: Acquire multi-period remote sensing image data of the area to be identified, and simultaneously acquire the corresponding canopy height model data; The above-mentioned multi-period remote sensing image data and the above-mentioned canopy height model data are preprocessed and aligned to obtain the multimodal time series data to be identified; The multimodal temporal data is input into the trained multimodal temporal deep learning model for forward inference calculation to obtain the category probability distribution corresponding to each pixel. The tree species category label for each pixel is determined based on the above category probability distribution; Output pixel-level tree species classification results containing spatial distribution information of tree species.

[0014] Secondly, the present invention also proposes a tree species identification system for fragmented urban green spaces, used to implement the tree species identification method for fragmented urban green spaces as described in any one of the first aspects of the claims, comprising: The building unit is used to acquire remote sensing image data of the study area in multiple periods within a preset phenological cycle, and simultaneously generate canopy height model data to construct a multimodal dataset containing spectral information, structural information and time series information. The processing unit is used to perform geometric correction, temporal registration and spatial alignment processing on the above multimodal dataset to obtain pixel-level aligned multimodal temporal data; The annotation unit is used to perform pixel-level annotations on aligned multimodal time-series data to construct a training sample dataset; The modeling unit is used to construct a multimodal temporal deep learning model and to perform feature extraction and fusion modeling on the aforementioned multimodal temporal data. The classification unit is used to perform inference calculations on the region to be identified based on the trained multimodal temporal deep learning model, and output pixel-level tree species classification results.

[0015] Thirdly, the present invention also proposes an electronic device comprising: a memory and a processor, characterized in that the processor is used to execute a computer program stored in the memory to implement the steps of the tree species identification method for fragmented urban green spaces as described in any of the first aspects.

[0016] Fourthly, the present invention also proposes a computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the tree species identification method for fragmented urban green spaces as described in any of the first aspects.

[0017] In summary, this invention constructs a high-frequency, multimodal, time-series UAV dataset to comprehensively characterize the dynamic phenological features of tree species, effectively enhancing data representation capabilities. By designing a deep fusion model of "spectrum-structure-temporal sequence," it achieves collaborative modeling of multi-source information in the spatiotemporal dimensions, significantly improving tree species identification accuracy and reducing spectral confusion. Introducing a combined loss function alleviates class imbalance and enhances the identification capability of rare tree species. Simultaneously, ablation analysis reveals the mechanism of action of multimodal features, improving model interpretability. Ultimately, this invention achieves high-precision, low-cost tree species identification in complex urban green space scenarios, providing effective technical support for refined ecological management.

[0018] Other advantages, objectives and features of this application will be apparent in part from the description which follows, and in part from what those skilled in the art will understand through study and practice of this application. Attached Figure Description

[0019] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit this specification. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart illustrating a tree species identification method for fragmented urban green spaces, provided as an embodiment of this application; Figure 2 A schematic diagram of an MMF-KNet model architecture provided in an embodiment of this application; Figure 3 This application provides a schematic diagram of the workflow of a timing adapter. Figure 4 A schematic diagram comparing the accuracy changes at various phenological stages, provided as an embodiment of this application; Figure 5 A schematic diagram illustrating IoU evaluation metrics for various tree species in different modalities, provided for embodiments of this application; Figure 6A schematic diagram illustrating how CHM data improves the IoU index of different tree species, as provided in an embodiment of this application; Figure 7 A schematic diagram of a baseline task and a confusion matrix after adding CHM data is provided for an embodiment of this application; Figure 8 A structural schematic diagram of a tree species identification system for fragmented urban green spaces provided in this application embodiment; Figure 9 This is a schematic diagram of an electronic device structure provided in an embodiment of this application. Detailed Implementation

[0020] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus. The technical solutions of the embodiments of this application will now be clearly and completely described in conjunction with the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0021] Please see Figure 1 This is a flowchart illustrating a tree species identification method for fragmented urban green spaces provided in this application embodiment, which may specifically include: S110. Acquire remote sensing image data of the study area within a preset phenological cycle and simultaneously generate canopy height model data to construct a multimodal dataset containing spectral information, structural information and time series information. S120. Perform geometric correction, temporal registration and spatial alignment on the above multimodal dataset to obtain pixel-level aligned multimodal temporal data; S130. Pixel-level annotation is performed on the aligned multimodal time-series data to construct a training sample dataset; S140. Construct a multimodal temporal deep learning model to perform feature extraction and fusion modeling on the above multimodal temporal data; S150: Based on the trained multimodal temporal deep learning model, inference calculations are performed on the region to be identified, and pixel-level tree species classification results are output.

[0022] For example, in step S110, firstly, multiple remote sensing image data of the study area within a preset phenological cycle are acquired. Preferably, RGB images are acquired through multi-temporal aerial photography using a UAV platform, and three-dimensional point cloud data is generated based on image reconstruction. By separating ground points and vegetation points from the three-dimensional point cloud data, a digital elevation model (DEM) and a digital surface model (DSM) are constructed respectively. Furthermore, the canopy height model (CHM) data is obtained by calculating the pixel-by-pixel difference between the DSM and the DEM, thereby forming a data foundation that simultaneously contains spectral and structural information. By aligning the images and CHM data of each period in chronological order, a multimodal dataset with time-series characteristics is constructed, so that the differences in the phenological changes of different tree species can be reflected.

[0023] In step S120, the multimodal dataset is subjected to geometric correction, temporal registration, and spatial alignment. Specifically, geometric fine correction is performed on the multi-period remote sensing image data using ground control points to eliminate spatial positioning errors. Using a preset reference image as a reference, feature point extraction and matching algorithms are used to complete the registration between images from different times. At the same time, mismatch elimination is used to improve the registration accuracy. Furthermore, the image and canopy height model data are resampled at a uniform resolution to achieve pixel-level alignment of the multimodal data in both spatial and temporal dimensions, resulting in consistent multimodal temporal data.

[0024] In step S130, pixel-level annotation is performed on the aligned multimodal time-series data to construct a training sample dataset. Specifically, different tree species regions can be finely labeled through manual annotation or semi-automatic annotation methods, thereby providing high-quality supervision information for subsequent model training.

[0025] In step S140, a multimodal temporal deep learning model is constructed to extract features and fuse the aforementioned multimodal temporal data. In this embodiment, the MMF-KNet network (Multi-Modal Fusion Kernel Network) is preferably used. At the input end, RGB (Red, Green, Blue, red-green-blue three-channel) images at the same time point are fused with canopy height model data in the channel dimension, and a multimodal input tensor containing the time dimension is constructed. In the encoding stage, a Transformer-based network structure is used to extract multi-scale spatial features, and a temporal adapter module is introduced in the encoding process. The features between different time steps are modeled through a multi-head self-attention mechanism to capture the dynamic features of tree species in the phenological change process. Then, the decoder module performs stepwise upsampling and fusion of the fused features to output pixel-level semantic segmentation results.

[0026] In step S150, the multimodal temporal deep learning model trained is used to perform inference calculations on the region to be identified. First, multi-period remote sensing image data and canopy height model data corresponding to the region to be identified are obtained. Multimodal temporal input data is constructed according to the same preprocessing and alignment method as in the training stage. The input data is used to perform forward inference on the model to obtain the probability distribution of tree species corresponding to each pixel. The final tree species classification result is determined according to the principle of maximum probability, thereby outputting pixel-level tree species identification results containing spatial distribution information, realizing high-precision automatic identification of different tree species in fragmented urban green spaces.

[0027] In one feasible implementation, step S110 acquires multiple remote sensing image data of the study area within a preset phenological cycle and simultaneously generates canopy height model data to construct a multimodal dataset containing spectral information, structural information, and time series information, including: S1101. Using an unmanned aerial vehicle (UAV) platform, multiple aerial photography missions are performed within the aforementioned preset phenological cycle at a preset sampling frequency to acquire the aforementioned multi-period remote sensing image data, wherein the aforementioned multi-period remote sensing image data is multi-period RGB image data. S1102. Generate three-dimensional point cloud data based on the above-mentioned multi-period RGB image data, and perform ground point and vegetation point classification processing on the above-mentioned three-dimensional point cloud data. S1103. Construct a digital elevation model based on the classified ground points, and construct a digital surface model based on the above vegetation points. S1104. The canopy height model is obtained by performing pixel-by-pixel difference calculation between the above-mentioned digital surface model and the above-mentioned digital elevation model. S1105. Pair the above canopy height model with the above multi-period RGB image data in chronological order to construct the above multimodal dataset.

[0028] For example, the main purpose of step S110 is to construct a multimodal dataset that simultaneously contains spectral information, structural information, and time series information, thereby providing a high-quality data foundation for subsequent tree species identification. Specifically, in step S1101, an unmanned aerial vehicle (UAV) platform is used to perform multiple aerial photography missions within a preset phenological cycle at a preset sampling frequency to obtain multi-period remote sensing image data of the study area at different time points, preferably multi-period RGB image data. By covering multiple key phenological stages such as budding, leaf expansion, vigorous growth, and decline, the spectral variation characteristics of different tree species in the time dimension can be fully collected.

[0029] In step S1102, three-dimensional reconstruction processing is performed based on the above-mentioned multi-period RGB image data to generate corresponding three-dimensional point cloud data. The three-dimensional point cloud data is then classified into ground points and vegetation points. The ground points are used to represent terrain undulation information, and the vegetation points are used to represent tree canopy structure information, thereby achieving effective separation of terrain and vegetation.

[0030] In step S1103, a digital elevation model is constructed based on the classified ground points to describe the elevation distribution of the bare ground surface; at the same time, a digital surface model is constructed based on the vegetation points to describe the overall height information of the ground surface including vegetation. By constructing the above-mentioned digital elevation model and digital surface model respectively, a foundation is provided for the subsequent extraction of vegetation height features.

[0031] In step S1104, the canopy height model is obtained by performing pixel-by-pixel difference calculation on the above-mentioned digital surface model and the above-mentioned digital elevation model. That is, by calculating the height difference between the digital surface model and the digital elevation model at the same spatial location, the vegetation canopy height at the corresponding location is obtained, thereby realizing the quantitative expression of the vertical structural characteristics of trees.

[0032] In step S1105, the canopy height model is paired with the aforementioned multi-period RGB image data in chronological order, so that each time node corresponds to a set of spectral and structural information, thereby constructing a multimodal dataset containing a time dimension. This multimodal dataset not only reflects the color and texture changes of different tree species at different time stages, but also reflects their spatial structural differences, providing unified and highly consistent input data for the subsequent training and inference of the multimodal temporal deep learning model.

[0033] In one feasible implementation, step S120 performs geometric correction, temporal registration, and spatial alignment on the multimodal dataset to obtain pixel-level aligned multimodal temporal data, including: S1201. Perform geometric fine correction on the above-mentioned multi-phase remote sensing image data based on ground control points to eliminate the spatial positioning error of the above-mentioned multi-phase remote sensing image data. S1202. Using a preset reference image as a reference, a feature point extraction and matching algorithm is used to perform time-series registration of the above-mentioned multi-period remote sensing image data, and an abnormal matching point is removed by a mismatch elimination algorithm. S1203. Based on the registration results, resample the above-mentioned multi-period remote sensing image data for each period to ensure that the images in each period are consistent in spatial location. S1204. Resample the above canopy height model according to a preset resolution so that the above canopy height model and the above multi-period remote sensing image data are consistent in spatial resolution. S1205. Based on the above processing results, the above multi-period remote sensing image data and the above canopy height model are aligned at the pixel level in the time and spatial dimensions to obtain the above multimodal time series data.

[0034] For example, the main purpose of step S120 above is to perform unified spatial and temporal standardization processing on the multimodal dataset to eliminate geometric distortion, spatial offset and resolution differences between data acquired at different times, thereby ensuring the consistency of each modal data in spatial location and temporal dimension during subsequent pixel-level annotation and model modeling.

[0035] Specifically, in step S1201, based on the ground control points deployed in the study area, the above-mentioned multi-period remote sensing image data are subjected to geometric fine correction processing. By establishing the correspondence between image coordinates and real geographic coordinates, the geometric distortion in the image is corrected, thereby eliminating spatial positioning errors caused by factors such as UAV attitude changes, sensor errors and terrain undulations, so that the images of each period have consistent spatial positioning accuracy under a unified geographic reference system.

[0036] In step S1202, using a preset reference image as a reference, time-series registration processing is performed on the aforementioned multi-period remote sensing image data. Stable features in images from different times are extracted through feature point extraction and matching algorithms, and corresponding relationships are established. At the same time, a mismatch elimination algorithm is combined to remove erroneous matching points, thereby improving registration accuracy and robustness, achieving consistent spatial alignment of images from different times, and ensuring that the position of the same ground feature remains consistent in images from different times.

[0037] In step S1203, based on the above registration results, the remote sensing image data of each period are resampled to map the images of different times to a unified spatial grid, thereby ensuring the consistency of pixel positions in the images of each period, so that the same pixel position corresponds to the same geographical area at different time nodes, providing a basis for subsequent temporal feature analysis.

[0038] In step S1204, the canopy height model is resampled according to a preset resolution to ensure that the canopy height model and the multi-period remote sensing image data are consistent in spatial resolution, thereby realizing a one-to-one correspondence between structural information and spectral information in spatial location and ensuring spatial consistency during multimodal data fusion.

[0039] In step S1205, based on the above-mentioned geometric correction, temporal registration, and resampling processing results, the above-mentioned multi-period remote sensing image data and the above-mentioned canopy height model are aligned at the pixel level in both the temporal and spatial dimensions. This ensures that the same spatial location has corresponding spectral and structural information at different time points, thereby obtaining pixel-level aligned multimodal time-series data. This multimodal time-series data can simultaneously reflect the spatial structural characteristics and phenological change characteristics of tree species, providing highly consistent data input for subsequent feature extraction and fusion modeling of multimodal time-series deep learning models.

[0040] In one feasible implementation, step S140 above constructs a multimodal temporal deep learning model, performing feature extraction and fusion modeling on the multimodal temporal data, including: S1401. The above-mentioned multi-period remote sensing image data at each time point are fused with the above-mentioned canopy height model in the channel dimension to form multi-channel input data; S1402. Normalize the above multi-channel input data to construct a multimodal temporal input tensor containing the time dimension; S1403. Input the above multimodal temporal input tensor into the encoder module, and use a Transformer-based backbone network to extract multi-scale spatial features; S1404. Introduce a timing adapter module into the encoder module to perform temporal dimension modeling of the multi-scale spatial features to capture the feature dependencies between different time steps. S1405. The features processed by the above-mentioned timing adapter module are fused and input into the decoder module for upsampling and feature fusion to generate pixel-level semantic segmentation results.

[0041] In one feasible implementation, step S1405 fuses the features processed by the timing adapter module and inputs them into the decoder module for upsampling and feature fusion to generate pixel-level semantic segmentation results, including: S14051. The multi-scale features output by the above timing adapter module are uniformly mapped according to the preset number of channels; S14052. The multi-scale features are upsampled based on the multi-scale feature fusion structure, and the high-level semantic features are fused with the low-level detail features. S14053. In the above feature fusion process, context information enhancement processing is introduced to supplement the above fused features with global information. S14054. Input the enhanced fusion features into the classification layer and output the probability distribution map of tree species corresponding to each pixel. S14055. Determine the pixel-level semantic segmentation result based on the above tree species category probability distribution map.

[0042] For example, the main purpose of step S140 above is to construct a multimodal time-series deep learning model that can simultaneously integrate spectral information, structural information and time series information, thereby achieving refined identification of different tree species in fragmented urban green spaces.

[0043] Specifically, in step S1401, the aforementioned multi-period remote sensing image data at each time point is fused with the aforementioned canopy height model in the channel dimension to form multi-channel input data. This ensures that the input not only includes spectral information reflecting the apparent features of the canopy, such as color and texture, but also height information reflecting tree height and canopy spatial structure, thus providing a more comprehensive feature foundation for subsequent model learning. In step S1402, the aforementioned multi-channel input data is normalized and organized chronologically to form a multimodal temporal input tensor containing the time dimension, eliminating numerical scale differences between different modalities and channels while preserving the dynamic information of tree species changing over time within the phenological cycle. In step S1403, the aforementioned multimodal temporal input tensor is input into the encoder module, where a Transformer-based backbone network is used to extract multi-scale spatial features. A self-attention mechanism is used to establish long-distance dependencies between different spatial locations, thereby obtaining multi-scale semantic representations under different receptive fields.

[0044] In step S1404, a temporal adapter module is introduced into the encoder module to model the multi-scale spatial features in the temporal dimension, thereby capturing the feature dependencies between different time steps. The input feature tensor is represented as follows: in, Indicates batch size, Indicates the length of the time series. Indicates the number of image patches. This represents the feature dimension. To enable the model to model the changes in the same spatial location at different time steps, the feature tensor is first reshaped into... This involves merging the batch dimension and the spatial block dimension, enabling the attention mechanism to focus on the evolutionary relationships between different time steps in the temporal dimension. Based on this, the input features... The query matrix, key matrix, and value matrix are obtained through linear transformation, specifically as follows: in, are learnable parameters, and in, The number of attention heads is denoted as _____. Subsequently, a multi-head self-attention mechanism is used in the time dimension to calculate the correlation between different time steps, and the attention calculation formula is as follows: This formula allows the model to adaptively learn the dependencies between features at different time points and highlight the time nodes that are more critical for identifying the current tree species. Finally, the attention-weighted features are updated through residual connections and layer normalization, with the output format as follows: This results in a feature representation that incorporates dynamic information from the time dimension. Through the above processing, the model can not only utilize the spatial features of a single temporal phase, but also explicitly learn the dynamic evolution patterns of the same spatial location under different phenological stages, thereby enhancing its ability to identify temporal differences among different tree species.

[0045] In step S1405, the features processed by the temporal adapter module are fused and input into the decoder module for upsampling and feature fusion to generate pixel-level semantic segmentation results. Specifically, in step S14051, the multi-scale features output by the temporal adapter module are uniformly mapped according to a preset number of channels to ensure consistency of features at different scales in the channel dimension, facilitating subsequent fusion operations. In step S14052, the multi-scale features are upsampled based on the multi-scale feature fusion structure, and high-level semantic features are fused with low-level detail features to simultaneously consider category discrimination ability and target boundary detail expression ability. In step S14053, context information enhancement processing is introduced during the feature fusion process to supplement the fused features with global information, thereby improving the model's ability to recognize tree crown boundaries, local occlusion, and differences between similar tree species in complex scenes. In step S14054, the enhanced fused features are input into the classification layer, and the probability distribution map of tree species corresponding to each pixel is output. In step S14055, the pixel-level semantic segmentation result is determined based on the above tree species category probability distribution map, thereby obtaining the tree species category to which each pixel in the area to be identified belongs, and realizing high-precision tree species identification in urban fragmented green space scenarios.

[0046] In one feasible implementation, the above-mentioned multimodal temporal deep learning model training process includes: S210. Construct training input data and corresponding label data based on the above training sample dataset; S220. Construct a combined loss function of weighted cross-entropy loss function and Dice loss function to alleviate the class imbalance problem; S230. Set the category weight parameters in the weighted cross-entropy loss function above according to the pixel proportion of each category; S240. Based on the above combined loss function, perform end-to-end training on the above multimodal temporal deep learning model; S250. During training, a preset optimizer is used to update the model parameters, and a preset learning rate adjustment strategy is used to dynamically adjust the learning rate. S260. In the above training process, auxiliary branch outputs are introduced to enhance gradient propagation and improve model convergence stability.

[0047] For example, the training process of the above-mentioned multimodal temporal deep learning model aims to improve the accuracy and stability of pixel-level tree species classification by constructing a reasonable loss function and optimization strategy to achieve effective learning of model parameters.

[0048] Specifically, in step S210, training input data and corresponding label data are constructed based on the aforementioned training sample dataset. The input data consists of preprocessed multimodal time-series data, and the label data consists of the corresponding pixel-level tree species annotation results. In step S220, a combined loss function of weighted cross-entropy loss and Dice loss is constructed to alleviate the class imbalance problem while simultaneously considering classification accuracy and region overlap. Its combined form is as follows: in, and , which is the loss weight coefficient, used to adjust the contribution of different loss terms to the overall training objective.

[0049] The weighted cross-entropy loss function is defined as follows: in, Indicates the total number of pixels. Indicates the number of categories. Indicates the first Class weight coefficient, These are real labels (usually in one-hot format). The model predicts that this pixel belongs to the first... The probability of a class. This is achieved by introducing class weights. This can increase the influence of fewer sample classes during the training process.

[0050] In step S230, the category weight parameters in the weighted cross-entropy loss function are set according to the pixel proportion of each category. Specifically, this can be set through the inverse relationship between category frequencies, for example... in, Indicates the first The proportion of pixels of the same type This is a smoothing constant used to prevent excessive weights, thereby enabling adaptive adjustment to class imbalance problems.

[0051] Meanwhile, the Dice loss function is defined as: in, The model represents the first The predicted probability of each pixel. This represents the corresponding real label. To prevent smoothing terms with a denominator of zero, this loss function effectively improves the segmentation ability for small targets or boundary regions by measuring the degree of overlap between the predicted and ground truth regions.

[0052] In step S240, the multimodal temporal deep learning model is trained end-to-end based on the combined loss function. The gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm, and the model weights are updated step by step.

[0053] In step S250, the model parameters are updated using a preset optimizer during training, such as Adam or SGD optimizer, and the learning rate is dynamically adjusted in conjunction with a learning rate scheduling strategy to improve the model convergence speed and avoid getting trapped in local optima.

[0054] In step S260, an auxiliary branch output is introduced during the training process. By adding auxiliary supervision signals in the intermediate layer, shallow features can also be effectively trained, thereby enhancing gradient propagation ability, alleviating the gradient vanishing problem in deep networks, and further improving the overall convergence stability and generalization ability of the model.

[0055] In one feasible implementation, step S150 above performs inference calculations on the region to be identified based on the trained multimodal temporal deep learning model, and outputs pixel-level tree species classification results, including: S1501. Acquire multi-period remote sensing image data of the area to be identified, and simultaneously acquire the corresponding canopy height model data; S1502. The above-mentioned multi-period remote sensing image data and the above-mentioned canopy height model data are preprocessed and aligned to obtain the multimodal time series data to be identified. S1503. Input the above multimodal temporal data into the trained multimodal temporal deep learning model to perform forward inference calculation and obtain the category probability distribution corresponding to each pixel. S1504. Determine the tree species category label for each pixel based on the above category probability distribution; S1505. Output pixel-level tree species classification results containing spatial distribution information of tree species.

[0056] For example, step S150 is mainly used to apply the trained multimodal temporal deep learning model to the region to be identified, thereby achieving pixel-level classification and spatial distribution extraction of tree species. Specifically, in step S1501, multiple remote sensing image data of the region to be identified within a preset phenological cycle are acquired, and the corresponding canopy height model data is acquired simultaneously, so that the input data contains both spectral and structural information, thus maintaining consistency with the data format in the training phase. In step S1502, the aforementioned multiple remote sensing image data and the aforementioned canopy height model data are preprocessed and aligned, including geometric correction, temporal registration, and spatial resolution unification, to obtain pixel-level aligned multimodal temporal data, ensuring consistency between the input data and the model training data in the spatial location and temporal dimensions during the inference phase.

[0057] In step S1503, the aforementioned multimodal temporal data is input into the trained multimodal temporal deep learning model for forward inference computation. The encoder module, temporal adapter module, and decoder module perform layer-by-layer feature extraction and fusion modeling on the input data, outputting the probability distribution of tree species categories corresponding to each pixel. In step S1504, the tree species category label for each pixel is determined based on the aforementioned category probability distribution, thereby transforming the continuous probability output into discrete classification results. In step S1505, the classification results of all pixels are spatially integrated, outputting pixel-level tree species classification results containing spatial distribution information of tree species, thus achieving refined identification of different tree species in fragmented urban green spaces.

[0058] In one feasible implementation, to fully exploit the complementarity between the spectral-texture information contained in RGB temporal images and the three-dimensional structural information reflected by CHM, this invention designs an end-to-end multimodal adaptive fusion network—MMF-KNet. The overall architecture of the model is as follows: Figure 2 As shown, it mainly consists of three parts: data processing, encoder, and decoder. Figure 3 This is a flowchart of the timing adapter's workflow.

[0059] Training, validation, and testing data are managed using a custom dataset, MultimodalMultispectralDataset, which supports different modal data formats. During data loading, the temporal RGB data is read and scaled to 512×512 pixels using the rasterio module, normalized to the [0,1] interval, with missing values ​​padded with zeros. The CHM data is clipped to a height of 0-15m and normalized. The RGB and CHM data at each time point are stacked along the channel dimension to form a multimodal temporal tensor of shape (T,4,H,W). This tensor is then uniformly adjusted to a five-dimensional tensor of (B,T,C,H,W) by the data preprocessing module, completing channel normalization and label alignment before being input into the model.

[0060] The encoder consists of a multimodal temporal adapter and a BEiT (Bidirectional Encoder representation from Image Transformers) backbone network, used to extract multi-scale spatial features from fused temporal information. The input data has a shape of (B, T, 4, 512, 512). The wrapper independently calls the BEiT with shared weights at each time point to obtain multi-scale features for each time point. Then, the temporal features of the same scale are stacked and compressed using mean fusion (or attention fusion). BEiT is a visual Transformer with 4 input channels, an image patch size of 16, an embedding dimension of 768, and 12 layers. The outputs of layers 3, 5, 9, and 11 are selected as multi-scale features, with each scale having 768 channels and the spatial size decreasing layer by layer.

[0061] The decoder uses UPerHead as the main segmentation head, supplemented by an FCN auxiliary head. UPerHead unifies the four scale features output from the encoder to 512 channels through 1×1 convolutions, then fuses them step-by-step using a feature pyramid structure, and extracts global contextual information using a pyramid pooling module (PPM). Finally, it outputs a prediction map with 15 classes via 3×3 convolutions, supervised by cross-entropy loss (weight 1.0). The auxiliary head, based on the BEiT layer 2 features, reduces the dimensionality to 256 through two 1×1 convolutions before direct classification, with a loss weight of 0.4, used for stable training and to provide additional gradient signals.

[0062] To capture the dynamic characteristics of autumn phenological evolution, temporal modeling is introduced based on previous fusion. A lightweight multimodal temporal adapter is designed to achieve efficient fusion of multi-temporal features without significantly increasing the number of parameters.

[0063] Specifically, the input is a multi-temporal image sequence. First, image patch embedding is performed, dividing the image at each time step into N blocks and projecting them onto a D-dimensional feature space to obtain a tensor. After the specified layers (layers 3, 6, and 9) of BEiT, the tensor will be reshaped as follows. The module then feeds the data into a temporal adapter: This module applies a multi-head self-attention mechanism along the temporal dimension to capture the dependencies between the same spatial location at different time steps, and then enhances temporal features through residual connections and layer normalization. After all layers have been processed, average pooling is performed along the temporal dimension to output multi-scale feature maps for the decoder. This design decouples temporal modeling from spatial feature extraction, enabling the model to effectively utilize phenological temporal information without significantly increasing computational burden.

[0064] In one feasible implementation, this embodiment uses the riverside green space along Beibinhe Road in Lanzhou, Gansu Province as the research area to verify the multimodal temporal deep learning tree species identification method for fragmented urban green spaces proposed in this invention. The geographical coordinates of the research area are 103°42′12″E, 36°05′11″N, with an average elevation of approximately 1515 meters and a total area of ​​approximately 2.7 square kilometers. It is distributed in a strip along the north bank of the Yellow River valley, representing a typical riverside green space under urban artificial intervention. The climate type of this area is temperate continental monsoon, with an average annual precipitation of approximately 327 mm and an average annual temperature of approximately 10.3℃. During the autumn months of September to November, the diurnal temperature range can reach over 15℃. This significant temperature variation drives obvious differences in the onset time, rate of color change, and leaf fall time of different tree species, thus providing favorable natural conditions for tree species identification based on temporal features.

[0065] Regarding vegetation composition, based on the results of the field quadrat survey in August 2023, a total of 14 major tree species were identified in the study area, covering typical urban green space tree species such as black locust, goldenrain tree, ash, purple-leaf plum, Chinese toon, maple, red ash, willow, crabapple, lilac, golden elm, juniper, arborvitae and cedar. Different tree species showed significant differences in autumn phenological changes, providing discriminative temporal features for model learning.

[0066] For data acquisition, this embodiment uses the DJI Mavic 3M Multispectral UAV as the remote sensing data acquisition platform. This platform integrates an RTK positioning module, which can achieve centimeter-level positioning accuracy under good signal conditions. The flight altitude was set at 60 meters, corresponding to a ground resolution of approximately 2.8 cm / pixel, and was uniformly resampled to a resolution of 0.1 meters. The forward overlap rate was no less than 80%, and the lateral overlap rate was no less than 70% to ensure the accuracy of 3D reconstruction. The flight time was selected to be 13:00 ± 1 hour daily, with the shooting mode triggering the shutter every 2 seconds, and only performed under clear, low-wind conditions to ensure consistent image quality.

[0067] Based on the data collection results described above, a multimodal dataset was constructed according to the method of this invention, and a canopy height model (CHM) was generated. Geometric correction, temporal registration, and spatial alignment were then performed to obtain pixel-level aligned multimodal temporal data. On this basis, a training sample dataset was constructed using manually labeled data, and a multimodal temporal deep learning model was built for training and inference.

[0068] In terms of model structure, this embodiment uses BEiT, a Transformer-based network, as the encoder backbone network and introduces a time-series adapter module for time-dimensional modeling. Comparative experiments, as shown in Table 1, demonstrate that among the three backbone networks—ResNet50, Swin Transformer, and BEiT—the BEiT model performs best, achieving a mIoU (mean Intersection over Union) of 72.3%, significantly outperforming other models, while maintaining a superior level of computational efficiency.

[0069] Table 1 Performance Comparison of Different Backbone Networks Regarding multimodal fusion strategies, as shown in Table 2, comparative experiments on early-stage fusion, mid-stage fusion, and late-stage fusion revealed that the early-stage fusion strategy, which directly splices RGB and CHM along the channel dimension, achieved the best results, with an overall classification accuracy of 77.87%, significantly outperforming other fusion methods.

[0070] Table 2 Performance Comparison of Different Fusion Strategies In terms of time series modeling, such as Figure 4 As shown, through ablation experiments with different time series lengths (Len2 to Len16), the results show that the model performance first increases and then decreases with the increase of time series length. The best performance is achieved under the Len10 condition, with an F-score (F1 score, harmonic mean of precision and recall) of 76.7% and an IoU of 63.26%. This indicates that appropriate time series information can significantly improve the model's recognition ability, but excessively long time series will introduce redundant information and reduce model performance.

[0071] In modal combination ablation experiments, such as Figure 5 As shown, a comparative analysis was conducted on single RGB mode, RGB+CHM dual mode, RGB+temporal mode, and RGB+CHM+temporal tri-mode. The results show that the tri-mode fusion scheme has the best performance, with an mIoU of 66.3%, indicating that structural information and temporal information have a significant supplementary effect on spectral information.

[0072] like Figure 6As shown, at the species level, the IoU of almost all tree species was significantly improved, especially in categories that performed poorly in the baseline. For example, the IoU of Platycladus orientalis improved from 0 to 20.1%, that of Ulmus parvifolia from 5.5% to 67.9%, and that of Syringa vulgaris from 0.03% to 63.9%. The fact that some tree species had an IoU of 0 in the baseline task may be due to missed detections caused by the multimodal data ablation experiments being conducted under the same conditions (Batch size=1). We have avoided this situation by adjusting the training strategy during model design and optimization.

[0073] like Figure 7 As shown, confusion matrix analysis reveals the improvement mechanism of CHM. The addition of height information effectively reduced misclassification: in the baseline, cedar was easily misclassified as background (14.5%) and black locust (15.3%). After adding CHM, the misclassification rates dropped to 4.9% and nearly 0, respectively, and the classification accuracy improved from 49.2% to 83.4%. For tree species with similar colors but different heights, CHM provides an independent discriminative dimension. Taking ash and red ash as examples, in the baseline, the two were severely confused (ash accuracy 28.6%, red ash 0%). After adding CHM, the accuracy of ash improved to 81.5%, and red ash to 69.9%, with a significant reduction in the mutual misclassification rate. In addition, golden elm was easily misclassified as willow (20.6%) and maple (16.0%) in the baseline. After adding CHM, the misclassification rate dropped to 6.0% and 2.5% respectively, and the accuracy improved from 3.6% to 86.0%. This shows that the structural information of CHM helps the model to more accurately define the canopy boundary and reduce confusion caused by spectral similarity or boundary ambiguity.

[0074] In this typical urban fragmented green space scenario, the method of the present invention achieves high-precision identification of multiple tree species by fusing spectral information, structural information and time series information, combined with a Transformer-based deep learning model. Its comprehensive mIoU reaches 66.3%, verifying the effectiveness, stability and practical value of the method of the present invention in complex urban ecological environments.

[0075] In summary, this invention addresses the problems of insufficient data dimensions, limited fusion capabilities, and weak model interpretability in existing urban tree species identification methods. It systematically innovates from multiple levels, including data construction, model design, training optimization, and mechanism analysis, significantly improving the overall accuracy and applicability of tree species identification. Firstly, at the data level, this invention constructs a high-frequency, multimodal UAV time-series dataset covering the complete autumn phenological evolution process. This dataset continuously characterizes the dynamic differences of different tree species in the onset, rate, and leaf fall of color change, effectively compensating for the shortcomings of traditional single-phase or sparse time-series data in reflecting phenological characteristics, thus improving data expressiveness from the source. Secondly, at the model level, the proposed multimodal time-series fusion model, through the construction of a "spectral-structural-temporal" collaborative modeling mechanism, achieves deep fusion of RGB imagery, canopy height models, and phenological time-series information in both spatial and temporal dimensions. This significantly enhances the ability to distinguish tree species with significant structural differences and spectral similarities, greatly improving overall identification accuracy and model stability. At the training level, this invention effectively alleviates the class imbalance problem by introducing a combined loss function that combines weighted cross-entropy and Dice, significantly improving the identification ability of rare tree species and enabling the model to maintain good generalization performance even in scenarios with complex class distributions. At the mechanistic level, this invention, through systematic ablation experiments and information content analysis, reveals the role mechanism and synergistic relationship of different modal features in tree species identification, clarifies the influence of key phenological stages and temporal density on model performance, enhances the interpretability of the model, and provides a theoretical basis for subsequent data collection and model optimization. At the application level, this invention can achieve pixel-level high-precision tree species identification in real-world scenarios with fragmented urban green spaces and complex canopy overlaps, significantly reducing the cost of manual surveys, improving monitoring efficiency, and providing reliable technical support for applications such as urban green space resource management, biodiversity assessment, and carbon sink calculation. In summary, this invention outperforms existing technologies in terms of identification accuracy, dynamic monitoring capabilities, model robustness, and practical application value.

[0076] Secondly, this invention also proposes a tree species identification system for fragmented urban green spaces, such as... Figure 8 As shown, the tree species identification method for fragmented urban green spaces as described in any one of the first aspects of the claim includes: The construction unit 21 is used to acquire remote sensing image data of the study area in multiple periods within a preset phenological cycle, and simultaneously generate canopy height model data to construct a multimodal dataset containing spectral information, structural information and time series information. Processing unit 22 is used to perform geometric correction, temporal registration and spatial alignment processing on the above multimodal dataset to obtain pixel-level aligned multimodal temporal data; Labeling unit 23 is used to perform pixel-level labeling based on aligned multimodal time-series data to construct a training sample dataset; Modeling unit 24 is used to construct a multimodal temporal deep learning model and perform feature extraction and fusion modeling on the above multimodal temporal data; Classification unit 25 is used to perform inference calculations on the region to be identified based on the trained multimodal temporal deep learning model, and output pixel-level tree species classification results.

[0077] In one feasible implementation, a tree species identification system for fragmented urban green spaces can also perform any step of the method proposed in the first aspect.

[0078] Thirdly, the present invention also proposes an electronic device 300, such as... Figure 9 As shown, it includes a memory 310, a processor 320, and a computer program 311 stored in the memory 310 and executable on the processor. When the processor 320 executes the computer program 311, it implements the steps of the tree species identification method for urban fragmented green space as described in any of the first aspects.

[0079] Fourthly, the present invention also proposes a computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the tree species identification method for fragmented urban green spaces as described in any one of the first aspects.

[0080] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0081] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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 this application.

Claims

1. A multimodal temporal deep learning method for tree species identification in fragmented urban green spaces, characterized in that, include: Acquire remote sensing image data of the study area within a preset phenological cycle and simultaneously generate canopy height model data to construct a multimodal dataset containing spectral, structural, and time-series information; Geometric correction, temporal registration, and spatial alignment are performed on the multimodal dataset to obtain pixel-level aligned multimodal temporal data; A training sample dataset is constructed by performing pixel-level annotations on the aligned multimodal time-series data; A multimodal temporal deep learning model is constructed to perform feature extraction and fusion modeling on the multimodal temporal data; The trained multimodal temporal deep learning model performs inference calculations on the region to be identified and outputs pixel-level tree species classification results.

2. The method according to claim 1, characterized in that, The process involves acquiring multiple remote sensing image data of the study area within a preset phenological cycle and simultaneously generating canopy height model data to construct a multimodal dataset containing spectral, structural, and time-series information, including: The drone platform is used to perform multiple aerial photography missions within the preset phenological cycle at a preset sampling frequency to acquire the multi-period remote sensing image data, wherein the multi-period remote sensing image data is multi-period RGB image data; Three-dimensional point cloud data is generated based on the multi-period RGB image data, and the three-dimensional point cloud data is classified into ground points and vegetation points. A digital elevation model is constructed based on the classified ground points, and a digital surface model is constructed based on the vegetation points. The canopy height model is obtained by performing pixel-by-pixel difference calculation between the digital surface model and the digital elevation model; The canopy height model is paired with the multi-period RGB image data of each period in chronological order to construct the multimodal dataset.

3. The method according to claim 1, characterized in that, The process of performing geometric correction, temporal registration, and spatial alignment on the multimodal dataset to obtain pixel-level aligned multimodal temporal data includes: Geometric fine correction is performed on the multi-phase remote sensing image data based on ground control points to eliminate spatial positioning errors in the multi-phase remote sensing image data; Using a preset reference image as a reference, a feature point extraction and matching algorithm is used to perform time-series registration of the multi-period remote sensing image data in each period, and an abnormal matching point is removed by a mismatch elimination algorithm. Based on the registration results, the remote sensing image data of each period are resampled to ensure that the images of each period are consistent in spatial location. The canopy height model is resampled according to a preset resolution so that the canopy height model and the multi-period remote sensing image data are consistent in spatial resolution; Based on the above processing results, the multi-period remote sensing image data and the canopy height model are aligned at the pixel level in both the time and spatial dimensions to obtain the multimodal time series data.

4. The method according to claim 1, characterized in that, The construction of the multimodal temporal deep learning model, which involves feature extraction and fusion modeling of the multimodal temporal data, includes: The multi-period remote sensing image data at each time point is fused with the canopy height model in the channel dimension to form multi-channel input data; The multi-channel input data is normalized to construct a multimodal temporal input tensor containing the time dimension; The multimodal temporal input tensor input encoder module uses a Transformer-based backbone network to extract multi-scale spatial features. A time-adaptor module is introduced into the encoder module to model the multi-scale spatial features in the time dimension in order to capture the feature dependencies between different time steps; The features processed by the timing adapter module are fused and input into the decoder module for upsampling and feature fusion to generate pixel-level semantic segmentation results.

5. The method according to claim 4, characterized in that, The process of fusing the features processed by the timing adapter module and inputting them into the decoder module for upsampling and feature fusion to generate pixel-level semantic segmentation results includes: The multi-scale features output by the timing adapter module are uniformly mapped according to a preset number of channels; The multi-scale features are upsampled based on the multi-scale feature fusion structure, and high-level semantic features are fused with low-level detail features; Contextual information enhancement processing is introduced during the feature fusion process to supplement the fused features with global information; The enhanced fusion features are input into the classification layer, and the probability distribution map of tree species corresponding to each pixel is output. The pixel-level semantic segmentation result is determined based on the probability distribution map of the tree species categories.

6. The method according to claim 1, characterized in that, The training process of the multimodal temporal deep learning model includes: Based on the training sample dataset, construct training input data and corresponding label data; A combined loss function of weighted cross-entropy loss and Dice loss is constructed to alleviate the class imbalance problem; The category weight parameters in the weighted cross-entropy loss function are set according to the pixel proportion of each category. The multimodal temporal deep learning model is trained end-to-end based on the combined loss function. During training, a preset optimizer is used to update the model parameters, and a preset learning rate adjustment strategy is used to dynamically adjust the learning rate. An auxiliary branch output is introduced during the training process to enhance gradient propagation and improve model convergence stability.

7. The method according to claim 1, characterized in that, The trained multimodal temporal deep learning model performs inference calculations on the region to be identified and outputs pixel-level tree species classification results, including: Acquire multi-period remote sensing image data of the area to be identified, and simultaneously acquire the corresponding canopy height model data; The multi-period remote sensing image data and the canopy height model data are preprocessed and aligned to obtain multimodal time-series data to be identified; The multimodal temporal data is input into the trained multimodal temporal deep learning model for forward inference calculation to obtain the category probability distribution corresponding to each pixel; The tree species category label for each pixel is determined based on the aforementioned category probability distribution; Output pixel-level tree species classification results containing spatial distribution information of tree species.

8. A tree species identification system for fragmented urban green spaces, used to implement the tree species identification method for fragmented urban green spaces as described in any one of claims 1 to 7, comprising: The building unit is used to acquire remote sensing image data of the study area in multiple periods within a preset phenological cycle, and simultaneously generate canopy height model data to construct a multimodal dataset containing spectral information, structural information and time series information. The processing unit is used to perform geometric correction, temporal registration and spatial alignment processing on the multimodal dataset to obtain pixel-level aligned multimodal temporal data; The annotation unit is used to perform pixel-level annotations on aligned multimodal time-series data to construct a training sample dataset; The modeling unit is used to construct a multimodal temporal deep learning model and to perform feature extraction and fusion modeling on the multimodal temporal data. The classification unit is used to perform inference calculations on the region to be identified based on the trained multimodal temporal deep learning model, and output pixel-level tree species classification results.

9. An electronic device, comprising: The memory and processor are characterized in that the processor is used to implement the steps of the tree species identification method for fragmented urban green spaces as described in any one of claims 1-7 when executing a computer program stored in the memory.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the tree species identification method for fragmented urban green spaces as described in any one of claims 1-7.

Images