A method, system and computer program product for detecting regional forest structure types

Abstract

The present invention discloses a method, system and computer program product for detecting regional forest structure types, the method comprising: obtaining multi-source remote sensing data of unmanned aerial vehicles in a test area, constructing a sample data set; constructing a multi-task neural network model, obtaining a synchronous detection model for regional vegetation cover type, tree height and forest canopy closure; obtaining satellite multi-spectral remote sensing image data of the area to be detected, inputting the synchronous detection model, obtaining synchronous detection results of regional vegetation cover type, tree height and forest canopy closure; based on the detection results, obtaining regional forest structure type detection results. The technical solution provided by the present invention realizes synchronous detection of vegetation cover type, tree height and forest canopy closure by constructing a multi-task neural network model, obtains regional forest, shrub and grass forest structure type detection results; fine-tunes and improves the network model; performs statistical analysis on the detection results; and constructs a spatiotemporal distribution map of changes in forest structure types in the area to be detected.

Description

Technical Field

[0001] The invention relates to the field of forest image processing, and in particular to a method for detecting regional forest structure types. Background Art

[0002] As global ecological and environmental problems become increasingly serious, scientific management and monitoring of forest, grassland and other ecosystems have become increasingly important. Vegetation cover type, tree height and canopy density are key indicators for describing the structure and function of forest ecosystems. These parameters are not only important for ecological monitoring and resource assessment, but also the basis for ecosystem service analysis and biodiversity conservation.

[0003] Among them, the extraction of vegetation cover types mainly relies on optical remote sensing technology. Affected by weather and light changes, the data quality is often unstable. The measurement of tree height usually uses laser radar (LiDAR) technology. Although this technology can provide high-precision three-dimensional information, its high cost and complex operation limit its application in a large range. The acquisition of canopy density generally relies on ground surveys or empirical models. These methods have limitations in space and time and are difficult to meet the needs of large-scale, continuous monitoring.

[0004] More importantly, traditional techniques often treat the extraction of vegetation cover type, tree height and canopy density as independent processes, lacking a collaborative analysis method. This separate processing method may lead to temporal and spatial inconsistencies between data, affecting the comprehensive assessment of forest, shrub and grass vegetation structure. Therefore, there is an urgent need for a method that can effectively integrate multi-source remote sensing data to simultaneously extract vegetation cover type, tree height and canopy density, so as to comprehensively characterize the forest structure type of forest, shrub and grass. Summary of the invention

[0005] In view of the above-mentioned deficiencies in the current technology, the present invention provides a method for detecting regional forest structure types. By constructing a multi-task neural network model and effectively combining multi-source remote sensing data, it realizes the simultaneous detection of vegetation cover type, tree height and forest density, and thus obtains the regional forest structure type detection result.

[0006] To achieve the above object, the embodiments of the present invention adopt the following technical solutions:

[0007] A method for detecting regional forest structure types comprises the following steps:

[0008] Acquire multi-source remote sensing data of the test area from unmanned aerial vehicles, wherein the multi-source remote sensing data from unmanned aerial vehicles include spatial image data of high spatial resolution remote sensing image data and distance image data based on LiDAR scanning data;

[0009] The vegetation cover type of the spatial image data was calibrated, and the tree height and forest density of the distance image data were calibrated to construct a sample data set;

[0010] According to the sample data set, a multi-task neural network model was constructed and trained based on multi-task learning to obtain a synchronous detection model for regional vegetation cover type, tree height and forest canopy density;

[0011] Acquire satellite multispectral remote sensing image data of the area to be detected;

[0012] The satellite multispectral remote sensing image data of the area to be detected is input into the synchronous detection model of regional vegetation cover type, tree height and forest canopy density to obtain the synchronous detection results of regional vegetation cover type, tree height and forest canopy density;

[0013] Based on the simultaneous detection results of regional vegetation cover type, tree height and forest canopy density, the regional forest structure type detection results are obtained.

[0014] According to one aspect of the present invention, the regional forest structure type detection method further includes: preprocessing the UAV multi-source remote sensing data of the test area and the satellite multi-spectral remote sensing image data of the area to be detected.

[0015] According to one aspect of the present invention, the preprocessing of the UAV multi-source remote sensing data of the test area and the satellite multi-spectral remote sensing image data of the area to be detected includes: noise removal, geometric correction, radiation calibration, atmospheric correction and image registration.

[0016] According to one aspect of the present invention, the multi-task neural network model is constructed based on multi-task learning and trained, including:

[0017] Based on the Swin Transformer model, we introduce hierarchical feature expression and a self-attention mechanism based on a shifted window to construct a deep shared feature extraction layer.

[0018] After the shared feature extraction layer, branch networks are constructed for the vegetation cover type detection task, tree height detection task, and forest canopy density detection task respectively. The branch networks adopt self-attention and shared attention strategies.

[0019] The branch network adopts a multi-task loss function, integrates the weighted losses of vegetation cover type detection task, tree height detection task and forest canopy density detection task, introduces adjustable weight coefficients, and dynamically adjusts the weights according to task characteristics and complexity.

[0020] According to one aspect of the present invention, the multi-task neural network model is constructed based on multi-task learning and trained further comprising:

[0021] The branch network for the vegetation cover type detection task uses a classification task decoder;

[0022] The branch network for the tree height detection task uses a tree height estimation decoder;

[0023] The branch network for the forest canopy closure detection task uses a canopy closure estimation decoder.

[0024] According to one aspect of the present invention, the multi-task neural network model is constructed based on multi-task learning and trained further comprising:

[0025] When training the neural network model, the shared feature extraction layer is trained first, and then the task-specific layers are gradually added for fine-tuning;

[0026] The early stopping mechanism and model regularization measures are adopted to prevent overfitting, enhance the generalization ability of the model, and ensure the stability of the training process.

[0027] According to one aspect of the present invention, the multi-task neural network model is constructed based on multi-task learning and trained further comprising:

[0028] Cross-validation, learning rate adjustment, and weight distribution methods are used to optimize the performance of the model.

[0029] According to one aspect of the present invention, the multi-task neural network model is constructed based on multi-task learning and trained further comprising:

[0030] Use multiple indicators, including at least precision, recall, F1 score, and overall accuracy, to comprehensively evaluate model performance;

[0031] Based on the comprehensive evaluation results, the network model is fine-tuned.

[0032] According to one aspect of the present invention, the method for detecting regional forest structure types further comprises the following steps:

[0033] Statistical analysis was conducted on the detection results of regional vegetation cover type, tree height and forest density.

[0034] According to one aspect of the present invention, the method for detecting regional forest structure types further comprises the following steps:

[0035] According to the detection results of regional forest structure types at different time points in the area to be detected, a spatiotemporal distribution map of regional forest structure types is constructed.

[0036] A regional forest structure type detection system, based on the regional forest structure type detection method as described above, comprises:

[0037] The sample construction module is used to obtain multi-source remote sensing data from drones in the test area and construct a sample data set;

[0038] The model building module is used to build and train a multi-task neural network model based on multi-task learning according to the sample data set to obtain a synchronous detection model for regional vegetation cover type, tree height and forest canopy density;

[0039] A data acquisition module, used to acquire satellite multispectral remote sensing image data of the area to be detected;

[0040] The detection module is used to input the satellite multispectral remote sensing image data of the area to be detected into the synchronous detection model of regional vegetation cover type, tree height and forest canopy density, and obtain the synchronous detection results of regional vegetation cover type, tree height and forest canopy density;

[0041] The data integration module is used to obtain the regional forest structure type detection results based on the simultaneous detection results of regional vegetation cover type, tree height and forest canopy density.

[0042] A computer program product comprises a computer program, wherein when the computer program is executed, the steps of the method for detecting regional forest structure types as described above are implemented.

[0043] Advantages of the present invention:

[0044] The present invention provides a method for detecting regional forest structure types. By constructing a multi-task neural network model, the different characteristics of satellite multispectral remote sensing image data are comprehensively extracted, and deep learning is performed to achieve synchronous detection of vegetation cover type, tree height and forest canopy density, thereby obtaining regional forest, shrub and grass forest structure type detection results.

[0045] This method can also conduct a comprehensive evaluation of the network model based on multiple indicators, and fine-tune and improve the network model according to the evaluation results; after obtaining the vegetation coverage type, tree height and forest canopy density test results of the area to be tested through the detection model, the test results can be statistically analyzed to further obtain forest-related information and conditions in the area, facilitating subsequent ecological protection, etc.; according to the regional forest structure type detection results, the data is integrated to construct a spatiotemporal distribution map of the changes in forest structure types in the area to be tested, which can provide relevant information for environmental monitoring and decision-making analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0047] Figure 1 This is a flow chart of a method for detecting regional forest structure types according to Embodiment 1 of the present invention;

[0048] Figure 2 A structural diagram of a detection model of a regional forest structure type detection method according to the present invention;

[0049] Figure 3 This is a flow chart of a method for detecting regional forest structure types according to Embodiment 2 of the present invention. DETAILED DESCRIPTION

[0050] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] Embodiment 1

[0052] like Figure 1 As shown, a method for detecting regional forest structure types includes the following steps:

[0053] A method for detecting regional forest structure types comprises the following steps:

[0054] S1: Obtain multi-source remote sensing data from drones in the test area.

[0055] The multi-source remote sensing data of the UAV includes spatial image data of high spatial resolution remote sensing image data and distance image data based on LiDAR scanning data.

[0056] High spatial resolution remote sensing image data refers to remote sensing image data with a small ground sampling distance that can clearly show ground details. This type of data can usually provide rich ground object information and help to carry out fine ground object classification, identification, measurement and monitoring.

[0057] LiDAR scan data is a type of three-dimensional data acquired by a laser scanner (LiDAR scanner). LiDAR mainly measures distance information. The LiDAR scanner calculates the distance to an object by emitting laser pulses and measuring their return time.

[0058] The regional forest structure type detection method also includes: preprocessing the multi-source remote sensing data of unmanned aerial vehicles in the test area.

[0059] The preprocessing of the multi-source remote sensing data of the UAV in the test area includes: noise removal, geometric correction, radiation calibration, atmospheric correction and image registration.

[0060] S2: Calibrate the vegetation cover type of the spatial image data, calibrate the tree height and forest density of the distance image data, and construct a sample data set.

[0061] The vegetation cover types include: trees, shrubs and herbs.

[0062] When constructing a sample dataset, data augmentation is performed on the image data to construct a more extensive sample dataset.

[0063] Data augmentation methods such as random rotation, flipping, and cropping are used to enhance the model's generalization ability to diverse surface features.

[0064] In practical applications, this method tries to establish a large-scale sample database that is diverse and widely distributed in space to support iterative training and parameter optimization of deep learning models. The goal is to increase the training accuracy of the model to more than 85%. For models that do not meet this accuracy standard, detailed inspections will be carried out, and their interpretation signs and labels of change types will be supplemented, training parameters will be adjusted, and iterative training will be repeated to achieve higher accuracy.

[0065] S3: Based on the sample data set, a multi-task neural network model is constructed and trained based on multi-task learning to obtain a simultaneous detection model for regional vegetation cover type, tree height and forest density.

[0066] The multi-task neural network model is constructed based on multi-task learning and trained, including:

[0067] Based on the Swin Transformer model, we introduce hierarchical feature expression and a self-attention mechanism based on a shifted window to build a deep shared feature extraction layer.

[0068] After the shared feature extraction layer, branch networks are constructed for the vegetation cover type detection task, tree height detection task, and forest canopy density detection task respectively. The branch networks adopt self-attention and shared attention strategies.

[0069] The branch network adopts a multi-task loss function, integrates the weighted losses of vegetation cover type detection task, tree height detection task and forest canopy density detection task, introduces adjustable weight coefficients, and dynamically adjusts the weights according to task characteristics and complexity.

[0070] The multi-task neural network model is constructed based on multi-task learning and trained further comprising:

[0071] The branch network for the vegetation cover type detection task uses a classification task decoder;

[0072] The branch network for the tree height detection task uses a tree height estimation decoder;

[0073] The branch network for the forest canopy closure detection task uses a canopy closure estimation decoder.

[0074] In practical applications, the multi-task neural network model can adopt the multi-task fully convolutional neural network model (MT-FCN).

[0075] The multi-task fully convolutional neural network model (MT-FCN) is a deep learning framework that can handle multiple related visual tasks simultaneously. It usually contains a shared feature extraction part and multiple branch networks designed for specific tasks. By sharing the feature extraction part, MT-FCN can reduce repeated calculations, improve efficiency, and improve the performance of each task through the correlation between tasks.

[0076] In the multi-task fully convolutional neural network model, Swin Transformer can be used as a shared feature extraction layer. The hierarchical design and windowed self-attention mechanism of Swin Transformer enable it to efficiently extract local and global features in the image, providing rich feature representations for subsequent specific tasks.

[0077] Swin Transformer is an advanced deep learning model designed for computer vision tasks. It combines the powerful capabilities of the Transformer model in sequence data processing with the efficient local feature extraction advantages of convolutional neural networks (CNNs) in image recognition. Swin Transformer solves the computational complexity problem in high-resolution image processing while maintaining high performance by introducing hierarchical feature representation and a self-attention mechanism based on shifted windows.

[0078] After the shared feature extraction layer, separate branch networks are created for each task. These branch networks can be built based on the blocks of the Swin Transformer, or combined with other convolutional layers, fully connected layers (fully connected layers are usually not used in FCN, but can be adjusted appropriately according to specific task requirements) or upsampling layers, etc.

[0079] In practical applications, the shared feature extraction layer can perform feature extraction only on the satellite multispectral remote sensing image data of the area to be detected, or it can perform feature extraction on the image data of both the test area and the area to be detected.

[0080] The multi-task neural network model is constructed based on multi-task learning and trained further comprising:

[0081] When training the neural network model, the shared feature extraction layer is trained first, and then the task-specific layers are gradually added for fine-tuning;

[0082] The early stopping mechanism and model regularization measures are adopted to prevent overfitting, enhance the generalization ability of the model, and ensure the stability of the training process.

[0083] Preferably, the building of a multi-task neural network model based on multi-task learning and training also includes:

[0084] Cross-validation, learning rate adjustment, and weight distribution methods are used to optimize the performance of the model.

[0085] S4: Obtain satellite multispectral remote sensing image data of the area to be detected.

[0086] The satellite multispectral remote sensing image data includes space image data and distance image data.

[0087] The regional forest structure type detection method further includes: preprocessing the satellite multi-spectral remote sensing image data of the area to be detected.

[0088] The preprocessing of the satellite multispectral remote sensing image data of the area to be detected includes: noise removal, geometric correction, radiation calibration, atmospheric correction and image registration.

[0089] S5: Input the satellite multispectral remote sensing image data of the area to be detected into the synchronous detection model of regional vegetation cover type, tree height and forest canopy density to obtain the synchronous detection results of regional vegetation cover type, tree height and forest canopy density.

[0090] The output of the detection model of this method includes three parts: one is the vegetation cover type detection result; one is the tree height detection result; and the other is the forest canopy density detection result. These three parts of the output together constitute the model's comprehensive analysis of satellite multispectral remote sensing image data, which can provide rich information for subsequent application scenarios.

[0091] S6: Based on the simultaneous detection results of regional vegetation cover type, tree height and forest density, the regional forest structure type detection results are obtained.

[0092] By integrating the detection results of vegetation cover type, tree height and forest canopy density, the forest structure type of the forest, shrub and grass in the area to be detected can be constructed, including the specific distribution areas of trees, shrubs and herbs.

[0093] like Figure 2 As shown, this is the structure diagram of the detection model in this method.

[0094] The beneficial effect of this embodiment is that: this method constructs a multi-task neural network model, comprehensively extracts the different characteristics of satellite multispectral remote sensing image data, conducts deep learning, realizes simultaneous detection of vegetation cover type, tree height and forest density, integrates these three indicators, and then obtains regional forest, shrub and grass forest structure type detection results.

[0095] Embodiment 2

[0096] like Figure 3As shown, a method for detecting regional forest structure types includes the following steps:

[0097] S1: Obtain multi-source remote sensing data from drones in the test area.

[0098] The multi-source remote sensing data of the UAV includes spatial image data of high spatial resolution remote sensing image data and distance image data based on LiDAR scanning data.

[0099] High spatial resolution remote sensing image data refers to remote sensing image data with a small ground sampling distance that can clearly show ground details. This type of data can usually provide rich ground object information and help to carry out fine ground object classification, identification, measurement and monitoring.

[0100] LiDAR scan data is a type of three-dimensional data acquired by a laser scanner (LiDAR scanner). LiDAR mainly measures distance information. The LiDAR scanner calculates the distance to an object by emitting laser pulses and measuring their return time.

[0101] The regional forest structure type detection method also includes: preprocessing the multi-source remote sensing data of unmanned aerial vehicles in the test area.

[0102] The preprocessing of the multi-source remote sensing data of the UAV in the test area includes: noise removal, geometric correction, radiation calibration, atmospheric correction and image registration.

[0103] S2: Calibrate the vegetation cover type of the spatial image data, calibrate the tree height and forest density of the distance image data, and construct a sample data set.

[0104] The vegetation cover types include: trees, shrubs and herbs.

[0105] When constructing a sample dataset, data augmentation is performed on the image data to construct a more extensive sample dataset.

[0106] Data augmentation methods such as random rotation, flipping and cropping are used to enhance the model's generalization ability to diverse surface features.

[0107] In practical applications, this method tries to establish a large-scale sample database that is diverse and widely distributed in space to support iterative training and parameter optimization of deep learning models. The goal is to increase the training accuracy of the model to more than 85%. For models that do not meet this accuracy standard, detailed inspections will be carried out, and their interpretation signs and labels of change types will be supplemented, training parameters will be adjusted, and iterative training will be repeated to achieve higher accuracy.

[0108] S3: Based on the sample data set, a multi-task neural network model is constructed and trained based on multi-task learning to obtain a simultaneous detection model for regional vegetation cover type, tree height and forest density.

[0109] The multi-task neural network model is constructed based on multi-task learning and trained, including:

[0110] Based on the Swin Transformer model, we introduce hierarchical feature expression and a self-attention mechanism based on a shifted window to build a deep shared feature extraction layer.

[0111] After the shared feature extraction layer, branch networks are constructed for the vegetation cover type detection task, tree height detection task, and forest canopy density detection task respectively. The branch networks adopt self-attention and shared attention strategies.

[0112] The branch network adopts a multi-task loss function, integrates the weighted losses of vegetation cover type detection task, tree height detection task and forest canopy density detection task, introduces adjustable weight coefficients, and dynamically adjusts the weights according to task characteristics and complexity.

[0113] The multi-task neural network model is constructed based on multi-task learning and trained further comprising:

[0114] The branch network for the vegetation cover type detection task uses a classification task decoder;

[0115] The branch network for the tree height detection task uses a tree height estimation decoder;

[0116] The branch network for the forest canopy closure detection task uses a canopy closure estimation decoder.

[0117] In practical applications, the multi-task neural network model can adopt the multi-task fully convolutional neural network model (MT-FCN).

[0118] The multi-task fully convolutional neural network model (MT-FCN) is a deep learning framework that can handle multiple related visual tasks simultaneously. It usually contains a shared feature extraction part and multiple branch networks designed for specific tasks. By sharing the feature extraction part, MT-FCN can reduce repeated calculations, improve efficiency, and improve the performance of each task through the correlation between tasks.

[0119] In the multi-task fully convolutional neural network model, Swin Transformer can be used as a shared feature extraction layer. The hierarchical design and windowed self-attention mechanism of Swin Transformer enable it to efficiently extract local and global features in the image, providing rich feature representations for subsequent specific tasks.

[0120] Swin Transformer is an advanced deep learning model designed for computer vision tasks. It combines the powerful capabilities of the Transformer model in sequence data processing with the efficient local feature extraction advantages of convolutional neural networks (CNNs) in image recognition. Swin Transformer solves the computational complexity problem in high-resolution image processing while maintaining high performance by introducing hierarchical feature representation and a self-attention mechanism based on shifted windows.

[0121] After the shared feature extraction layer, separate branch networks are created for each task. These branch networks can be built based on the blocks of the Swin Transformer, or combined with other convolutional layers, fully connected layers (fully connected layers are usually not used in FCN, but can be adjusted appropriately according to specific task requirements) or upsampling layers, etc.

[0122] In practical applications, the shared feature extraction layer can perform feature extraction only on the satellite multispectral remote sensing image data of the area to be detected, or it can perform feature extraction on the image data of both the test area and the area to be detected.

[0123] The multi-task neural network model is constructed based on multi-task learning and trained further comprising:

[0124] When training the neural network model, the shared feature extraction layer is trained first, and then the task-specific layers are gradually added for fine-tuning;

[0125] The early stopping mechanism and model regularization measures are adopted to prevent overfitting, enhance the generalization ability of the model, and ensure the stability of the training process.

[0126] The multi-task neural network model is constructed based on multi-task learning and trained further comprising:

[0127] Cross-validation, learning rate adjustment, and weight distribution methods are used to optimize the performance of the model.

[0128] Preferably, the building of a multi-task neural network model based on multi-task learning and training also includes:

[0129] Use multiple indicators, including at least precision, recall, F1 score, and overall accuracy, to comprehensively evaluate model performance;

[0130] Based on the comprehensive evaluation results, the network model is fine-tuned.

[0131] S4: Obtain satellite multispectral remote sensing image data of the area to be detected.

[0132] The satellite multispectral remote sensing image data includes space image data and distance image data.

[0133] Spatial image data can be used to detect vegetation cover types, and distance image data can be used to detect tree height and forest density.

[0134] The regional forest structure type detection method further includes: preprocessing the satellite multi-spectral remote sensing image data of the area to be detected.

[0135] The preprocessing of the satellite multispectral remote sensing image data of the area to be detected includes: noise removal, geometric correction, radiation calibration, atmospheric correction and image registration.

[0136] S5: Input the satellite multispectral remote sensing image data of the area to be detected into the synchronous detection model of regional vegetation cover type, tree height and forest canopy density to obtain the synchronous detection results of regional vegetation cover type, tree height and forest canopy density.

[0137] The output of the detection model of this method includes three parts: one is the vegetation cover type detection result; one is the tree height detection result; and the other is the forest canopy density detection result. These three parts of the output together constitute the model's comprehensive analysis of satellite multispectral remote sensing image data, which can provide rich information for subsequent application scenarios.

[0138] S6: Based on the simultaneous detection results of regional vegetation cover type, tree height and forest density, the regional forest structure type detection results are obtained.

[0139] By integrating the detection results of vegetation cover type, tree height and forest canopy density, the forest structure type of the forest, shrub and grass in the area to be detected can be constructed, including the specific distribution areas of trees, shrubs and herbs.

[0140] like Figure 2 As shown, this is the structure diagram of the detection model in this method.

[0141] S7: Conduct statistical analysis on the detection results of regional vegetation cover type, tree height and forest density.

[0142] After obtaining the detection results of vegetation cover type, tree height and forest canopy density through the neural network model, statistical analysis can also be performed on these data.

[0143] For example, the vegetation coverage types in the area to be detected can be counted, and the main vegetation coverage types, forest, shrub and grass types in the area can be analyzed.

[0144] For example, after obtaining the forest density in different regions, the forest density in each region can be displayed in the form of a heat map, thereby conducting an overall analysis of a large range or region.

[0145] Preferably, the system further comprises:

[0146] S8: Construct a spatiotemporal distribution map of regional forest structure types based on the detection results of regional forest structure types at different time points in the area to be detected.

[0147] The satellite multispectral remote sensing image data of the area to be detected at different time points can be obtained, and the synchronous detection model of regional vegetation cover type, tree height and forest canopy density can be input to obtain the synchronous detection results of vegetation cover type, tree height and forest canopy density at different time points in the area. These three indicators can be integrated to obtain the regional forest structure type detection results at different time points in the area. After that, these data can be integrated to construct a spatiotemporal distribution map showing the distribution of changes in forest structure types in the area and their time process, which can provide relevant information for environmental monitoring and decision analysis.

[0148] The beneficial effects of this embodiment are: this method can also conduct a comprehensive evaluation of the network model based on multiple indicators, and fine-tune and improve the network model according to the evaluation results; after obtaining the vegetation coverage type, tree height and forest density detection results of the area to be detected through the detection model, the detection results are statistically analyzed to further obtain forest-related information and conditions in the area, which is convenient for subsequent ecological protection, etc.; according to the regional forest structure type detection results, the data is integrated to construct a spatiotemporal distribution map of the changes in forest structure types in the area to be detected, which can provide relevant information for environmental monitoring and decision-making analysis.

[0149] Embodiment 3

[0150] A regional forest structure type detection system, based on the regional forest structure type detection method as described in Embodiment 1 or 2, comprises:

[0151] The sample construction module is used to obtain multi-source remote sensing data from drones in the test area and construct a sample data set;

[0152] The model building module is used to build and train a multi-task neural network model based on multi-task learning according to the sample data set to obtain a synchronous detection model for regional vegetation cover type, tree height and forest canopy density;

[0153] A data acquisition module, used to acquire satellite multispectral remote sensing image data of the area to be detected;

[0154] The detection module is used to input the satellite multispectral remote sensing image data of the area to be detected into the synchronous detection model of regional vegetation cover type, tree height and forest canopy density, and obtain the synchronous detection results of regional vegetation cover type, tree height and forest canopy density;

[0155] The data integration module is used to obtain the regional forest structure type detection results based on the simultaneous detection results of regional vegetation cover type, tree height and forest canopy density.

[0156] Preferably, the system further comprises:

[0157] A preprocessing module is used to preprocess the multi-source remote sensing data of UAVs in the test area and the satellite multi-spectral remote sensing image data of the area to be detected;

[0158] The statistical analysis module is used to perform statistical analysis on the detection results of regional vegetation cover type, tree height and forest canopy density.

[0159] Preferably, the system further comprises:

[0160] The distribution map construction module is used to construct a spatiotemporal distribution map of regional forest structure types based on the regional forest structure type detection results at different time points in the area to be detected.

[0161] Embodiment 4

[0162] A computer program, when executed, implements the steps of the method for detecting regional forest structure types as described in embodiment one or two.

[0163] Embodiment 5

[0164] A readable storage medium stores a computer program as described in Example 4, and when the computer program is executed, the steps of the regional forest structure type detection method as described in Example 1 or 2 are implemented.

[0165] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with the art within the technical scope disclosed in the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for detecting regional forest structure types, characterized in that: The following steps are involved: Acquire multi-source remote sensing data of the test area from unmanned aerial vehicles, wherein the multi-source remote sensing data from unmanned aerial vehicles include spatial image data of high spatial resolution remote sensing image data and distance image data based on LiDAR scanning data; The vegetation cover type of the spatial image data was calibrated, and the tree height and forest density of the distance image data were calibrated to construct a sample data set; According to the sample data set, a multi-task neural network model was constructed and trained based on multi-task learning to obtain a synchronous detection model for regional vegetation cover type, tree height and forest canopy density; Acquire satellite multispectral remote sensing image data of the area to be detected; The satellite multispectral remote sensing image data of the area to be detected is input into the synchronous detection model of regional vegetation cover type, tree height and forest canopy density to obtain the synchronous detection results of regional vegetation cover type, tree height and forest canopy density; Based on the simultaneous detection results of regional vegetation cover type, tree height and forest density, the regional forest structure type detection results are obtained; The multi-task neural network model is constructed based on multi-task learning and trained, including: Based on the Swin Transformer model, we introduce hierarchical feature expression and a self-attention mechanism based on a shifted window to construct a deep shared feature extraction layer. After the shared feature extraction layer, branch-based networks are constructed for the vegetation cover type detection task, tree height detection task, and forest canopy density detection task, respectively. The branch networks use self-attention and shared attention strategies. The branch network adopts a multi-task loss function, integrates the weighted losses of vegetation cover type detection task, tree height detection task and forest canopy density detection task, introduces an adjustable weight coefficient, and dynamically adjusts the weight according to the task characteristics and complexity; The branch network for the vegetation cover type detection task uses a classification task decoder; The branch network for the tree height detection task uses a tree height estimation decoder; The branch network for the forest canopy closure detection task uses a canopy closure estimation decoder; Wherein, constructing a multi-task neural network model based on multi-task learning and training it also includes: When training the neural network model, the shared feature extraction layer is trained first, and then the task-specific layers are gradually added for fine-tuning; Use early stopping mechanism and model regularization measures to prevent overfitting, enhance the generalization ability of the model, and ensure the stability of the training process; Use cross-validation, learning rate adjustment, and weight distribution methods to optimize model performance; Use multiple indicators, including precision, recall, F1 score, and overall accuracy, to comprehensively evaluate model performance; Based on the comprehensive evaluation results, the network model is fine-tuned.

2. The method for detecting regional forest structure types according to claim 1, characterized in that: The regional forest structure type detection method also includes: preprocessing the unmanned aerial vehicle multi-source remote sensing data of the test area and the satellite multi-spectral remote sensing image data of the area to be detected.

3. The method for detecting regional forest structure types according to claim 2, characterized in that: The preprocessing of the UAV multi-source remote sensing data of the test area and the satellite multi-spectral remote sensing image data of the area to be detected includes: noise removal, geometric correction, radiation calibration, atmospheric correction and image registration.

4. The method for detecting regional forest structure types according to claim 1, characterized in that: The regional forest structure type detection method also includes the following steps: Statistical analysis was conducted on the detection results of regional vegetation cover type, tree height and forest density.

5. The method for detecting regional forest structure types according to claim 1, characterized in that: The regional forest structure type detection method also includes the following steps: According to the detection results of regional forest structure types at different time points in the area to be detected, a spatiotemporal distribution map of regional forest structure types is constructed.

6. A regional forest structure type detection system, characterized in that: The method for detecting regional forest structure types according to any one of claims 1 to 5 comprises: The sample construction module is used to obtain multi-source remote sensing data from drones in the test area and construct a sample data set; The model building module is used to build and train a multi-task neural network model based on multi-task learning according to the sample data set to obtain a synchronous detection model for regional vegetation cover type, tree height and forest canopy density; A data acquisition module, used to acquire satellite multispectral remote sensing image data of the area to be detected; The detection module is used to input the satellite multispectral remote sensing image data of the area to be detected into the synchronous detection model of regional vegetation cover type, tree height and forest canopy density, and obtain the synchronous detection results of regional vegetation cover type, tree height and forest canopy density; The data integration module is used to obtain the detection results of regional forest structure types based on the simultaneous detection results of regional vegetation cover types, tree heights and forest canopy density; The multi-task neural network model is constructed based on multi-task learning and trained, including: Based on the Swin Transformer model, we introduce hierarchical feature expression and a self-attention mechanism based on a shifted window to construct a deep shared feature extraction layer. After the shared feature extraction layer, branch-based networks are constructed for the vegetation cover type detection task, tree height detection task, and forest canopy density detection task, respectively. The branch networks use self-attention and shared attention strategies. The branch network adopts a multi-task loss function, integrates the weighted losses of vegetation cover type detection task, tree height detection task and forest canopy density detection task, introduces an adjustable weight coefficient, and dynamically adjusts the weight according to the task characteristics and complexity; The branch network for the vegetation cover type detection task uses a classification task decoder; The branch network for the tree height detection task uses a tree height estimation decoder; The branch network for the forest canopy closure detection task uses a canopy closure estimation decoder; Wherein, constructing a multi-task neural network model based on multi-task learning and training it also includes: When training the neural network model, the shared feature extraction layer is trained first, and then the task-specific layers are gradually added for fine-tuning; Use early stopping mechanism and model regularization measures to prevent overfitting, enhance the generalization ability of the model, and ensure the stability of the training process; Use cross-validation, learning rate adjustment, and weight distribution methods to optimize model performance; Use multiple indicators, including precision, recall, F1 score, and overall accuracy, to comprehensively evaluate model performance; Based on the comprehensive evaluation results, the network model is fine-tuned.

7. A computer program product, characterized in that It comprises a computer program, which, when executed, implements the steps of the regional forest structure type detection method as described in any one of claims 1 to 5.

Images