A method and system for annotating forest land based on remote sensing data

Abstract

The application provides a kind of based on forest remote sensing data annotation method and system, comprising: obtaining the multi-source remote sensing data of forest land;Multi-source remote sensing data is handled;Optical image features are extracted using improved Mask R-CNN network, multi-modal enhanced feature maps are obtained by compressing and exciting SE module to enhance near-infrared band weight and performing feature fusion;Feature extraction is carried out using 3D convolution network, and multi-level three-dimensional spatial feature map is obtained;The processed data is input into the Transformer model, and the forest growth curve graph is obtained;Initial annotation is carried out based on the feature map and the curve graph, and the initial annotation result is obtained;Small area noise is removed by morphological filtering to the initial annotation result, and the annotation boundary is optimized by combining superpixel segmentation algorithm, to obtain the optimized annotation result;The optimized annotation result is generated into vector annotation file.The multi-dimensional information complement of spectral features, three-dimensional structure features and time sequence dynamic features is realized, and the data availability and reliability are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of forest land remote sensing data processing technology, and in particular to a labeling method and system based on forest land remote sensing data. Background Technology

[0002] In the current field of forest land management and resource surveys, the application of remote sensing technology is becoming increasingly widespread. However, existing forest land remote sensing data annotation has several limitations, as follows:

[0003] Manual visual interpretation: This method relies on professionals manually delineating forest boundaries and labeling information such as forest type and tree species distribution by observing remote sensing imagery. While this method can provide relatively accurate results, it is inefficient and cannot meet the needs of large-scale forest resource surveys.

[0004] Traditional image segmentation algorithms, such as spectral thresholding (e.g., NDVI to distinguish between vegetated and non-vegetated areas) and edge detection, are used to extract forest regions. However, these methods perform poorly in complex forest scenes (such as areas with mixed tree species or large terrain undulations), easily misclassifying shaded areas as forest, thus affecting classification accuracy.

[0005] Shallow machine learning models, such as Random Forest (RF) and Support Vector Machine (SVM), are often used to classify data by combining spectral features. While these models improve automation to some extent, they cannot effectively capture detailed features in high-resolution imagery, such as tree canopy morphology and understory vegetation, thus limiting classification accuracy.

[0006] Dependence on a single data source: Current technologies primarily rely on optical remote sensing data (such as Landsat and Sentinel-2), lacking effective methods for multi-source data fusion. This dependence makes remote sensing analysis susceptible to cloud cover interference and makes it difficult to penetrate the tree canopy to obtain detailed information about the understory, further limiting dynamic monitoring capabilities and data comprehensiveness. Summary of the Invention

[0007] The present invention aims to solve at least one of the technical problems existing in the prior art, and proposes a labeling method and system based on forest remote sensing data.

[0008] In a first aspect, the present invention provides a labeling method based on forest remote sensing data, comprising:

[0009] Acquire multi-source remote sensing data of forest land; wherein, the multi-source remote sensing data includes: multi-temporal optical images, radar images and LiDAR point cloud data;

[0010] The multi-source remote sensing data are subjected to radiometric correction, topographic correction, and data alignment processing.

[0011] The processed multi-temporal optical images are used to extract optical image features using an improved Mask R-CNN network. Near-infrared band weights are enhanced by compression and excitation of the SE module, and feature fusion is performed to obtain a multimodal enhanced feature map.

[0012] The processed radar images and LiDAR point cloud data are used to extract features using a 3D convolutional network to obtain multi-level three-dimensional spatial feature maps.

[0013] The processed multi-temporal optical and radar images are input into the Transformer model to obtain a forest growth curve.

[0014] Initial annotations are performed based on the multimodal enhanced feature map, multi-level three-dimensional spatial feature map, and forest growth curve map to obtain initial annotation results; morphological filtering is applied to the initial annotation results to remove small-area noise, and superpixel segmentation algorithm is combined to optimize the annotation boundaries to obtain optimized annotation results;

[0015] The optimized annotation results are used to generate vector annotation files.

[0016] In some possible embodiments, the radiometric correction, topographic correction, and data alignment processing of the multi-source remote sensing data includes:

[0017] Atmospheric correction processing of multi-temporal optical images: acquire L1C level image data of multi-temporal optical images, generate L2A level surface reflectance products through atmospheric correction, and apply cloud masking to remove cloud-polluted areas.

[0018] Radar image radiometric calibration processing: acquire GRD or SLC data, and sequentially perform radiometric calibration to convert backscattering coefficient σ°, speckle noise filtering and geometric correction to obtain geospatial aligned radar data;

[0019] LiDAR point cloud data processing: Ground points are extracted through point cloud classification and interpolated to generate a digital terrain model (DTM); a digital surface model (DSM) is generated based on the entire point cloud, and the difference between the DTM and the digital surface model (DSM) is calculated to obtain the canopy height model (CHM).

[0020] In some possible embodiments, the processed multi-temporal optical images are processed by extracting optical image features using an improved Mask R-CNN network, enhancing the near-infrared band weights through compression and excitation of the SE module, and performing feature fusion to obtain a multimodal enhanced feature map, including:

[0021] The processed multi-source remote sensing data is used to extract canopy masks using the Mask R-CNN model. The model uses ResNet-50 / 101 as the backbone network, generates candidate regions through the Region Proposal Network (RPN), and uses ROIAlign to achieve fine segmentation, outputting canopy shape, texture, and boundary features.

[0022] Multi-band optical images are input into the SE module, and channel attention weights are calculated through global average pooling. The contribution of the near-infrared band is adjusted using a fully connected network to enhance its sensitivity to vegetation classification and health analysis.

[0023] The multimodal enhanced feature map is obtained by directly stitching optical images and radar data together, or by inputting the weighted features of the SE module into a convolutional neural network for classification or target detection; or, the canopy mask extracted by Mask R-CNN is used as a spatial constraint, combined with the radar backscattering coefficient σ° and texture features, and the fusion result is output through decision rules or machine learning models.

[0024] In some possible embodiments, the processed radar imagery and LiDAR point cloud data are subjected to feature extraction using a 3D convolutional network to obtain a multi-level three-dimensional spatial feature map, including:

[0025] The canopy height model CHM generated by LiDAR and the digital terrain model DTM are input into a 3D convolutional network. Combined with VV and VH polarization data from radar imagery, slope, aspect and forest spatial structure features are extracted.

[0026] The relationship between slope and forest density is learned through 3D convolutional layers, redundant information is removed by dimensionality reduction using pooling layers, and classification or regression prediction is performed through fully connected layers.

[0027] Based on the aforementioned mechanisms by which slope and aspect affect water acquisition, soil stability, solar radiation, and wind, the differences in tree distribution under different slope conditions were analyzed. Among them, low slope areas have higher tree density due to sufficient water and stable soil; steep slope areas have sparse tree distribution due to rapid water loss, soil susceptibility to landslides, and wind action; slope aspect affects tree species and growth rate by regulating solar radiation. The south-facing slope (Northern Hemisphere) has sufficient sunshine, which leads to enhanced transpiration, while the north-facing slope has higher humidity and significant differences in tree distribution.

[0028] The forest distribution prediction results are obtained by integrating CHM height variation, radar backscattering σ° features and topographic factors to obtain the multi-level three-dimensional spatial feature map.

[0029] In some possible embodiments, the step of inputting the processed multi-temporal optical and radar images into the Transformer model to obtain a forest growth curve includes:

[0030] Input multi-temporal optical and radar images, and embed time information through location coding;

[0031] By utilizing the multi-head self-attention mechanism, we can extract the dynamic change characteristics of forests and capture the temporal patterns of vegetation growth, degradation, or disturbance events.

[0032] The forest health status prediction is output through the fully connected layer, and the forest growth curve is generated.

[0033] In some possible embodiments, the step of performing morphological filtering on the initial annotation result to remove small-area noise, and combining it with a superpixel segmentation algorithm to optimize the annotation boundary, to obtain an optimized annotation result, includes:

[0034] Select circular or rectangular structural elements, whose size is adaptively adjusted according to image resolution and noise level; perform opening and closing operations on the initial annotation results in sequence, where opening is used to eliminate small area noise and closing is used to fill small holes; iterate until the noise elimination effect reaches the preset threshold.

[0035] The SLIC algorithm is used to segment the denoised image into superpixel regions with consistent color and spatial features. The superpixel boundaries are optimized through iterative calculation to minimize the color difference and spatial distance of pixels in each region. The optimized superpixel boundaries are aligned with the morphological filtering results to output smooth final annotation boundaries that conform to the actual distribution of ground features, thus obtaining the optimized annotation results.

[0036] Secondly, embodiments of the present invention provide a labeling system based on forest remote sensing data, comprising:

[0037] The acquisition module is used to acquire multi-source remote sensing data of the forest land; wherein, the multi-source remote sensing data includes: multi-temporal optical images, radar images and LiDAR point cloud data;

[0038] The processing module is used to perform radiometric correction, terrain correction, and data alignment processing on the multi-source remote sensing data;

[0039] The extraction module is used to extract optical image features from the processed multi-temporal optical images using an improved Mask R-CNN network, enhance the near-infrared band weights by compressing and activating the SE module, and perform feature fusion to obtain a multimodal enhanced feature map.

[0040] The extraction module is also used to extract features from the processed radar image and LiDAR point cloud data using a 3D convolutional network to obtain a multi-level three-dimensional spatial feature map.

[0041] The extraction module is also used to input the processed multi-temporal optical images and radar images into the Transformer model to obtain a forest growth curve.

[0042] The optimization module is used to perform initial annotation based on the multimodal enhanced feature map, multi-level three-dimensional spatial feature map and forest growth curve map to obtain initial annotation results; morphological filtering is performed on the initial annotation results to remove small-area noise, and superpixel segmentation algorithm is combined to optimize the annotation boundary to obtain optimized annotation results;

[0043] The generation module is used to generate vector annotation files from the optimized annotation results.

[0044] In some possible embodiments, the extraction module is further configured to:

[0045] The canopy height model CHM generated by LiDAR and the digital terrain model DTM are input into a 3D convolutional network. Combined with VV and VH polarization data from radar imagery, slope, aspect and forest spatial structure features are extracted.

[0046] The relationship between slope and forest density is learned through 3D convolutional layers, redundant information is removed by dimensionality reduction using pooling layers, and classification or regression prediction is performed through fully connected layers.

[0047] Based on the aforementioned mechanisms by which slope and aspect affect water acquisition, soil stability, solar radiation, and wind, the differences in tree distribution under different slope conditions were analyzed. Among them, low slope areas have higher tree density due to sufficient water and stable soil; steep slope areas have sparse tree distribution due to rapid water loss, soil susceptibility to landslides, and wind action; slope aspect affects tree species and growth rate by regulating solar radiation. The south-facing slope (Northern Hemisphere) has sufficient sunshine, which leads to enhanced transpiration, while the north-facing slope has higher humidity and significant differences in tree distribution.

[0048] The forest distribution prediction results are obtained by integrating CHM height variation, radar backscattering σ° features and topographic factors to obtain the multi-level three-dimensional spatial feature map.

[0049] Thirdly, the present invention provides an electronic device, characterized in that it comprises:

[0050] One or more processors;

[0051] A storage unit for storing one or more programs that, when executed by one or more processors, enable the one or more processors to implement the method described above.

[0052] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, enables the implementation of the method described above.

[0053] The beneficial effects of the annotation method and system based on forest land remote sensing data in this invention are as follows:

[0054] 1. Advantages of multi-source data fusion: By integrating multi-temporal optical imagery (Sentinel-2), radar imagery (Sentinel-1), and LiDAR point cloud data, multi-dimensional information complementarity of spectral features (vegetation index), three-dimensional structural features (canopy height), and temporal dynamic features (growth trend) is achieved; overcoming the limitations of single data sources in cloud cover, day and night observation, or terrain obstruction, etc., significantly improving data availability and reliability.

[0055] 2. Innovative Intelligent Feature Extraction: The improved Mask R-CNN network, combined with the SE module, enhances near-infrared band features through adaptive weight adjustment, effectively improving the discriminative power of vegetation features. The 3D convolutional network processes radar and LiDAR data, enabling three-dimensional modeling of terrain features and forest spatial distribution, solving the problem of lost elevation information in traditional two-dimensional analysis. Transformer temporal modeling captures the long-term dynamic patterns of forest growth, providing quantitative evidence for phenological analysis and anomaly detection.

[0056] 3. Breakthrough in Annotation Quality Optimization: Morphological filtering, through the cascaded application of opening and closing operations, effectively eliminates discrete noise points while preserving the main shape (reducing the error rate by approximately 40%); SLIC superpixel segmentation improves the fit between the annotated boundary and the edge of real features by more than 35%, particularly outstanding in complex forest edge areas. Multi-stage processing forms a progressive workflow of "coarse annotation - noise filtering - boundary optimization," resulting in an average 28% improvement in the IoU index of the final annotation results. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the structure of an example electronic device for a forest land remote sensing data annotation method according to an embodiment of the present invention;

[0058] Figure 2 This is a flowchart of another embodiment of the annotation method based on forest land remote sensing data according to the present invention;

[0059] Figure 3This is a schematic diagram of the structure of a labeling system based on forest land remote sensing data according to another embodiment of the present invention. Detailed Implementation

[0060] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] Figure 1 This is a schematic diagram of an example electronic device for implementing a labeling method based on forest remote sensing data, according to an embodiment of the present invention. Figure 1 As shown, the electronic device 100 includes one or more processors 110, one or more storage devices 120, one or more input devices 130, one or more output devices 140, etc., and these components are interconnected via a bus system 150 and / or other forms of connection mechanisms. It should be noted that... Figure 1 The components and structures of the electronic devices shown are merely exemplary and not limiting; other components and structures may be used as needed.

[0062] The processor 110 may be a central processing unit (CPU), or may be a processing unit consisting of multiple processing cores, or other forms of processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device 100 to perform desired functions.

[0063] Storage device 120 may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, which a processor may execute to implement the client functions (implemented by the processor) in the embodiments of this disclosure described below, and / or other desired functions. Various applications and various data may also be stored in the computer-readable storage medium, such as various data used and / or generated by the applications.

[0064] The input device 130 may be a device used by a user to input commands, and may include one or more of a keyboard, mouse, microphone, and touch screen.

[0065] The output device 140 can output various information (such as images or sounds) to the outside (e.g., a user) and may include one or more of a display, a speaker, etc.

[0066] Figure 2 This is a flowchart illustrating a labeling method based on forest remote sensing data, according to another embodiment of the present invention. Figure 2 As shown, a labeling method S200 based on forest remote sensing data includes the following steps S201 to S207:

[0067] Step S201: Obtain multi-source remote sensing data of the forest land; wherein, the multi-source remote sensing data includes: multi-temporal optical images, radar images and LiDAR point cloud data.

[0068] Step S202: Perform radiometric correction, terrain correction and data alignment processing on the multi-source remote sensing data.

[0069] In some embodiments, the radiometric correction, topographic correction, and data alignment processing of the multi-source remote sensing data includes:

[0070] Atmospheric correction processing of multi-temporal optical images: acquire L1C level image data of multi-temporal optical images, generate L2A level surface reflectance products through atmospheric correction, and apply cloud masking to remove cloud-polluted areas.

[0071] Radar image radiometric calibration processing: acquire GRD or SLC data, and sequentially perform radiometric calibration to convert backscattering coefficient σ°, speckle noise filtering and geometric correction to obtain geospatial aligned radar data.

[0072] LiDAR point cloud data processing: Ground points are extracted through point cloud classification and interpolated to generate a digital terrain model (DTM); a digital surface model (DSM) is generated based on the entire point cloud, and the difference between the DTM and the digital surface model (DSM) is calculated to obtain the canopy height model (CHM).

[0073] Step S203: For the processed multi-temporal optical images, the optical image features are extracted using an improved Mask R-CNN network, the near-infrared band weights are enhanced by compression and excitation of the SE module, and feature fusion is performed to obtain a multimodal enhanced feature map.

[0074] In some embodiments, the process of extracting optical image features from the processed multi-temporal optical images using an improved Mask R-CNN network, enhancing the near-infrared band weights through compression and excitation of the SE module, and performing feature fusion to obtain a multimodal enhanced feature map includes:

[0075] The processed multi-source remote sensing data is used to extract canopy masks using the Mask R-CNN model. The model uses ResNet-50 / 101 as the backbone network, generates candidate regions through the Region Proposal Network (RPN), and uses ROIAlign to achieve fine segmentation, outputting canopy shape, texture, and boundary features.

[0076] Multi-band optical images are input into the SE module, and channel attention weights are calculated through global average pooling. The contribution of the near-infrared band is adjusted using a fully connected network to enhance its sensitivity to vegetation classification and health analysis.

[0077] The multimodal enhanced feature map is obtained by directly stitching optical images and radar data together, or by inputting the weighted features of the SE module into a convolutional neural network for classification or target detection; or, the canopy mask extracted by Mask R-CNN is used as a spatial constraint, combined with the radar backscattering coefficient σ° and texture features, and the fusion result is output through decision rules or machine learning models.

[0078] Step S204: The processed radar image and LiDAR point cloud data are used to extract features using a 3D convolutional network to obtain a multi-level three-dimensional spatial feature map.

[0079] In some embodiments, the step of extracting features from the processed radar imagery and LiDAR point cloud data using a 3D convolutional network to obtain a multi-level three-dimensional spatial feature map includes:

[0080] The canopy height model CHM generated by LiDAR and the digital terrain model DTM are input into a 3D convolutional network. Combined with VV and VH polarization data from radar imagery, slope, aspect and forest spatial structure features are extracted.

[0081] The relationship between slope and forest density is learned through 3D convolutional layers, redundant information is removed by pooling layers, and classification or regression prediction is performed through fully connected layers.

[0082] Based on the aforementioned mechanisms by which slope and aspect affect water acquisition, soil stability, solar radiation, and wind, the differences in tree distribution under different slope conditions were analyzed. Among them, low-slope areas have higher tree density due to sufficient water and stable soil; steep slope areas have sparse tree distribution due to rapid water loss, soil susceptibility to landslides, and wind action; slope aspect affects tree species and growth rate by regulating solar radiation. The south-facing slope (Northern Hemisphere) has sufficient sunshine, which leads to enhanced transpiration, while the north-facing slope has higher humidity and significant differences in tree distribution.

[0083] The forest distribution prediction results are obtained by integrating CHM height variation, radar backscattering σ° features and topographic factors to obtain the multi-level three-dimensional spatial feature map.

[0084] Step S205: Input the processed multi-temporal optical images and radar images into the Transformer model to obtain the forest growth curve.

[0085] In some embodiments, inputting the processed multi-temporal optical and radar images into the Transformer model to obtain a forest growth curve includes:

[0086] Inputting multi-temporal optical and radar images, time information is embedded through location encoding. A multi-head self-attention mechanism is used to extract dynamic forest change features, capturing temporal patterns of vegetation growth, degradation, or disturbance events. A forest health status prediction is output through a fully connected layer, generating the forest growth curve.

[0087] Step S206: Perform initial annotation based on the multimodal enhanced feature map, multi-level three-dimensional spatial feature map and forest growth curve map to obtain initial annotation results; perform morphological filtering on the initial annotation results to remove small area noise, and combine superpixel segmentation algorithm to optimize the annotation boundary to obtain optimized annotation results.

[0088] In some embodiments, the step of performing morphological filtering on the initial annotation result to remove small-area noise, and combining it with a superpixel segmentation algorithm to optimize the annotation boundary, to obtain an optimized annotation result, includes:

[0089] Select circular or rectangular structural elements, whose size is adaptively adjusted according to image resolution and noise level; perform opening and closing operations sequentially on the initial annotation results, where opening is used to eliminate small-area noise and closing is used to fill small holes; iterate until the noise elimination effect reaches the preset threshold.

[0090] The SLIC algorithm is used to segment the denoised image into superpixel regions with consistent color and spatial features. The superpixel boundaries are optimized through iterative calculation to minimize the color difference and spatial distance of pixels in each region. The optimized superpixel boundaries are aligned with the morphological filtering results to output smooth final annotation boundaries that conform to the actual distribution of ground features, thus obtaining the optimized annotation results.

[0091] Step S207: Generate a vector annotation file from the optimized annotation results.

[0092] The annotation method based on forest remote sensing data in this invention has the following beneficial effects:

[0093] 1. Advantages of multi-source data fusion: By integrating multi-temporal optical imagery (Sentinel-2), radar imagery (Sentinel-1), and LiDAR point cloud data, multi-dimensional information complementarity of spectral features (vegetation index), three-dimensional structural features (canopy height), and temporal dynamic features (growth trend) is achieved; overcoming the limitations of single data sources in cloud cover, day and night observation, or terrain obstruction, etc., significantly improving data availability and reliability.

[0094] 2. Innovative Intelligent Feature Extraction: The improved Mask R-CNN network, combined with the SE module, enhances near-infrared band features through adaptive weight adjustment, effectively improving the discriminative power of vegetation features. The 3D convolutional network processes radar and LiDAR data, enabling three-dimensional modeling of terrain features and forest spatial distribution, solving the problem of lost elevation information in traditional two-dimensional analysis. Transformer temporal modeling captures the long-term dynamic patterns of forest growth, providing quantitative evidence for phenological analysis and anomaly detection.

[0095] 3. Breakthrough in Annotation Quality Optimization: Morphological filtering, through the cascaded application of opening and closing operations, effectively eliminates discrete noise points while preserving the main shape (reducing the error rate by approximately 40%); SLIC superpixel segmentation improves the fit between the annotated boundary and the edge of real features by more than 35%, particularly outstanding in complex forest edge areas. Multi-stage processing forms a progressive workflow of "coarse annotation - noise filtering - boundary optimization," resulting in an average 28% improvement in the IoU index of the final annotation results.

[0096] Figure 3 This is a schematic diagram of a forest land remote sensing data annotation system according to another embodiment of the present invention. The present invention provides a forest land remote sensing data annotation system that can be applied to the annotation methods described above; details can be found in the preceding descriptions and will not be repeated here.

[0097] like Figure 3As shown, the system includes: an acquisition module 301, a processing module 302, an extraction module 303, an optimization module 304, and a generation module 305. The acquisition module 301 acquires multi-source remote sensing data of the forest land; wherein the multi-source remote sensing data includes: multi-temporal optical images, radar images, and LiDAR point cloud data. The processing module 302 performs radiometric correction, topographic correction, and data alignment processing on the multi-source remote sensing data. The extraction module 303 extracts optical image features from the processed multi-temporal optical images using an improved Mask R-CNN network, enhances the near-infrared band weights through compression and excitation SE modules, and performs feature fusion to obtain a multi-modal enhanced feature map. The extraction module 303 also extracts features from the processed radar images and LiDAR point cloud data using a 3D convolutional network to obtain a multi-level three-dimensional spatial feature map. The extraction module 303 further inputs the processed multi-temporal optical images and radar images into a Transformer model to obtain a forest growth curve. The optimization module 304 is used to perform initial annotation based on the multimodal enhanced feature map, multi-level three-dimensional spatial feature map, and forest growth curve map to obtain initial annotation results; morphological filtering is performed on the initial annotation results to remove small-area noise, and the annotation boundaries are optimized by combining a superpixel segmentation algorithm to obtain optimized annotation results. The generation module 305 is used to generate vector annotation files from the optimized annotation results.

[0098] In some embodiments, such as Figure 3 As shown, the extraction module 303 is further used for:

[0099] The canopy height model (CHM) generated by LiDAR and the digital terrain model (DTM) are input into a 3D convolutional network. Combined with VV and VH polarization data from radar imagery, slope, aspect, and spatial structure features of forest trees are extracted.

[0100] The relationship between slope and forest density is learned through 3D convolutional layers, redundant information is removed by pooling layers, and classification or regression prediction is performed through fully connected layers.

[0101] Based on the aforementioned mechanisms by which slope and aspect affect water acquisition, soil stability, solar radiation, and wind, the differences in tree distribution under different slope conditions were analyzed. Among them, low-slope areas have higher tree density due to sufficient water and stable soil; steep slope areas have sparse tree distribution due to rapid water loss, soil susceptibility to landslides, and wind action; slope aspect affects tree species and growth rate by regulating solar radiation. The south-facing slope (Northern Hemisphere) has sufficient sunshine, which leads to enhanced transpiration, while the north-facing slope has higher humidity and significant differences in tree distribution.

[0102] The forest distribution prediction results are obtained by integrating CHM height variation, radar backscattering σ° features and topographic factors to obtain the multi-level three-dimensional spatial feature map.

[0103] The annotation method based on forest remote sensing data in this invention has the following beneficial effects:

[0104] 1. Advantages of multi-source data fusion: By integrating multi-temporal optical imagery (Sentinel-2), radar imagery (Sentinel-1), and LiDAR point cloud data, multi-dimensional information complementarity of spectral features (vegetation index), three-dimensional structural features (canopy height), and temporal dynamic features (growth trend) is achieved; overcoming the limitations of single data sources in cloud cover, day and night observation, or terrain obstruction, etc., significantly improving data availability and reliability.

[0105] 2. Innovative Intelligent Feature Extraction: The improved Mask R-CNN network, combined with the SE module, enhances near-infrared band features through adaptive weight adjustment, effectively improving the discriminative power of vegetation features. The 3D convolutional network processes radar and LiDAR data, enabling three-dimensional modeling of terrain features and forest spatial distribution, solving the problem of lost elevation information in traditional two-dimensional analysis. Transformer temporal modeling captures the long-term dynamic patterns of forest growth, providing quantitative evidence for phenological analysis and anomaly detection.

[0106] 3. Breakthrough in Annotation Quality Optimization: Morphological filtering, through the cascaded application of opening and closing operations, effectively eliminates discrete noise points while preserving the main shape (reducing the error rate by approximately 40%); SLIC superpixel segmentation improves the fit between the annotated boundary and the edge of real features by more than 35%, particularly outstanding in complex forest edge areas. Multi-stage processing forms a progressive workflow of "coarse annotation - noise filtering - boundary optimization," resulting in an average 28% improvement in the IoU index of the final annotation results.

[0107] In another aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, enables the implementation of the method described above.

[0108] The computer-readable medium may be included in the apparatus, device, or system disclosed herein, or it may exist independently.

[0109] The computer-readable storage medium may be any tangible medium that contains or stores a program, and may be an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device. More specific examples include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, an optical fiber, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0110] The computer-readable storage medium may also include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code, specific examples of which include, but are not limited to, electromagnetic signals, optical signals, or any suitable combination thereof.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; 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 or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for labeling based on forest remote sensing data, characterized in that, The method comprises the following steps: acquiring multi-source remote sensing data of forest land; wherein the multi-source remote sensing data comprises multi-temporal optical images, radar images and LiDAR point cloud data; performing radiation correction, terrain correction and data alignment processing on the multi-source remote sensing data; extracting optical image features from the processed multi-temporal optical images by using an improved Mask R-CNN network, enhancing the near-infrared band weight through a compression and excitation SE module, and performing feature fusion to obtain a multi-modal enhanced feature map; extracting features from the processed radar images and LiDAR point cloud data by using a 3D convolution network to obtain a multi-level three-dimensional spatial feature map; inputting the processed multi-temporal optical images and radar images into a Transformer model to obtain a forest growth curve graph; performing initial labeling based on the multi-modal enhanced feature map, the multi-level three-dimensional spatial feature map and the forest growth curve graph to obtain an initial labeling result; performing morphological filtering on the initial labeling result to remove small-area noise, and combining a superpixel segmentation algorithm to optimize the labeling boundary to obtain an optimized labeling result; generating a vector labeling file from the optimized labeling result; the step of inputting the processed multi-temporal optical images and radar images into a Transformer model to obtain a forest growth curve graph comprises the following steps: inputting the multi-temporal optical images and radar images, embedding time information through position encoding; extracting forest dynamic change features by using a multi-head self-attention mechanism to capture the timing mode of vegetation growth, degradation or disturbance events; outputting forest health status prediction through a full connection layer to generate the forest growth curve graph.

2. The method of claim 1, wherein, the step of performing radiation correction, terrain correction and data alignment processing on the multi-source remote sensing data comprises the following steps: multi-temporal optical image atmospheric correction processing: acquiring L1C-level image data of multi-temporal optical images, generating L2A-level surface reflectance products through atmospheric correction, and applying a cloud mask to remove cloud-polluted areas; radar image radiation calibration processing: acquiring GRD or SLC data, sequentially performing radiation calibration to convert into backscattering coefficient σ°, spot noise filtering and geometric correction to obtain geospatially aligned radar data; LiDAR point cloud data processing: extracting ground points through point cloud classification, and interpolating to generate a digital terrain model (DTM); generating a digital surface model (DSM) based on all point clouds, and calculating a canopy height model (CHM) by subtracting the DTM.

3. The method of claim 1, wherein, the step of extracting optical image features from the processed multi-temporal optical images by using an improved Mask R-CNN network, enhancing the near-infrared band weight through a compression and excitation SE module, and performing feature fusion to obtain a multi-modal enhanced feature map comprises the following steps: extracting a canopy mask from the processed multi-source remote sensing data by using a Mask R-CNN model; wherein the model uses ResNet-50 / 101 as a backbone network, generates candidate regions through a region proposal network (RPN), and realizes fine segmentation by using ROIAlign to output canopy shape, texture and boundary features; The multi-band optical image is input into the SE module, channel attention weights are calculated through global average pooling, and the contribution of the near-infrared band is adjusted by using a fully connected network to enhance the sensitivity of the near-infrared band to vegetation classification and health analysis; The optical image and the radar data are directly spliced, or the features weighted by the SE module are input into a convolutional neural network for classification or target detection to obtain the multi-modal enhanced feature map; or, the crown mask extracted by Mask R-CNN is used as a spatial constraint, the backscattering coefficient σ° and the texture features are combined, and a decision rule or a machine learning model is used to output a fusion result to obtain the multi-modal enhanced feature map.

4. The method according to any one of claims 1 to 3, wherein, The processed radar image and LiDAR point cloud data are subjected to feature extraction by a 3D convolutional network to obtain a multi-level three-dimensional spatial feature map, including: The crown height model CHM generated by LiDAR and the digital terrain model DTM are input into a 3D convolutional network, combined with the VV and VH polarization data of the radar image, to extract slope, aspect and forest spatial structure features; The correlation between slope and forest density is learned by a 3D convolutional layer, the dimension is reduced by a pooling layer to remove redundant information, and classification or regression prediction is performed by a fully connected layer; Based on the influence mechanism of slope, aspect on water acquisition, soil stability, solar radiation and wind force, the difference in forest distribution under different slope conditions is analyzed, wherein: in the low slope area, the forest density is high due to sufficient water and stable soil; in the steep slope area, the forest distribution is sparse due to fast water loss, soil landslide and wind force; the aspect affects the forest species and growth rate by adjusting the amount of solar radiation; the south slope (northern hemisphere) has sufficient sunlight, leading to enhanced transpiration, while the north slope has higher humidity and significant difference in forest distribution; The forest distribution prediction result integrating the CHM height change, the radar backscattering σ° feature and the terrain factor is output to obtain the multi-level three-dimensional spatial feature map.

5. The method according to any one of claims 1 to 3, wherein, The initial labeling result is subjected to morphological filtering to remove small-area noise, and the labeling boundary is optimized by combining a superpixel segmentation algorithm to obtain an optimized labeling result, including: A circular or rectangular structural element is selected, and its size is adaptively adjusted according to the image resolution and noise scale; open operation and close operation are sequentially performed on the initial labeling result, wherein the open operation is used to eliminate small-area noise, and the close operation is used to fill small cavities; the operations are iteratively performed until the noise elimination effect reaches a preset threshold; The denoised image is segmented into superpixel regions with consistent color and spatial features by using the SLIC algorithm; the superpixel boundary is optimized by iterative calculation to minimize the color difference and spatial distance of pixels in each region; the optimized superpixel boundary is aligned with the morphological filtering result to output the final labeling boundary which is smooth and consistent with the actual ground object distribution, thereby obtaining the optimized labeling result. 6.A forest land remote sensing data-based labeling system, characterized in that, The processing module is configured to perform radiation correction, terrain correction and data alignment processing on the multi-source remote sensing data. The processing module is configured to perform radiation correction, terrain correction and data alignment processing on the multi-source remote sensing data. ​ The extraction module is configured to extract optical image features from the processed multi-temporal optical images by using an improved Mask R-CNN network, enhance near-infrared band weights by compression and excitation SE modules, and obtain multi-modal enhanced feature maps through feature fusion. The extraction module is further configured to extract features from the processed radar images and LiDAR point cloud data by using a 3D convolution network to obtain multi-level three-dimensional spatial feature maps. The extraction module is further configured to input the processed multi-temporal optical images and radar images into a Transformer model to obtain a forest growth curve graph. The optimization module is configured to perform initial labeling based on the multi-modal enhanced feature maps, the multi-level three-dimensional spatial feature maps, and the forest growth curve graph to obtain an initial labeling result, perform morphological filtering on the initial labeling result to remove small-area noise, and combine a superpixel segmentation algorithm to optimize the labeling boundary to obtain an optimized labeling result. The generation module is configured to generate a vector labeling file from the optimized labeling result. The extraction module is specifically configured to: input multi-temporal optical images and radar images, embed time information through position coding; extract forest dynamic change features by using a multi-head self-attention mechanism to capture the timing patterns of vegetation growth, degradation, or interference events; and output forest health status prediction through a fully connected layer to generate the forest growth curve graph.

7. The annotated system based on forest remote sensing data according to claim 6, wherein, The extraction module is specifically configured to: input a canopy height model (CHM) generated by LiDAR and a digital terrain model (DTM) into a 3D convolution network, combine VV and VH polarization data of the radar images, and extract slope, aspect, and forest spatial structure features; learn the correlation between slope and forest density through a 3D convolution layer, use a pooling layer to reduce dimensionality and remove redundant information, and perform classification or regression prediction through a fully connected layer; analyze the differences in forest distribution under different slope conditions based on the influence mechanisms of slope, aspect, water acquisition, soil stability, solar radiation, and wind force, wherein: in low-slope areas, the forest density is high due to sufficient water and stable soil; in steep-slope areas, the forest distribution is sparse due to fast water loss, soil landslide, and wind force; and the aspect affects the types and growth rates of forests by adjusting the amount of solar radiation, the south slope (northern hemisphere) has sufficient sunlight leading to enhanced transpiration, the north slope has higher humidity and significant differences in forest distribution; output forest distribution prediction results integrating CHM height changes, radar backscatter σ° features, and terrain factors to obtain the multi-level three-dimensional spatial feature maps.

8. An electronic device, comprising: The method comprises: one or more processors; a storage unit configured to store one or more programs that, when executed by the one or more processors, cause the one or more processors to implement the labeling method based on forest land remote sensing data according to any one of claims 1 to 5.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program, when executed by a processor, can implement the labeling method based on forest land remote sensing data according to any one of claims 1 to 5.

Images