Accurate estimation method and system for mountain forest canopy height
Through multi-source remote sensing data and deep learning technology, a deep Markov model is constructed, which solves the accuracy problem of forest canopy height estimation in mountain forests and achieves more accurate forest parameter prediction.
Patent Information
- Application Number
- CN202411290584.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-14
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-09-14
AI Technical Summary
In mountain forests, it is difficult for the prior art to accurately estimate forest canopy heights, mainly due to the limitations of different types of remote sensing data, amplification errors in complex terrain conditions and uncertainty in forest parameter estimation model.
Using multi-source remote sensing data sets, including satellite-based lidar data, microwave remote sensing data, optical remote sensing data and terrain data, the long-term dependence relationship of sequence data is captured through recurrent neural networks, a deep Markov model is constructed, combined with airborne radar data, Predict the height of forest canopy.
The accuracy of forest canopy height estimation under mountainous conditions is effectively improved, and the remote sensing variable information is fully explored through deep learning methods, overcoming the shortcomings of traditional methods in capturing spatial changes and texture patterns.
Smart Images

Figure CN118795471B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of mountain forest canopy height estimation, and in particular to a method and system for accurately estimating the mountain forest canopy height. Background Art
[0002] Tropical forests are the core of terrestrial ecosystems and play an irreplaceable role in mitigating global / regional climate change and regulating global carbon balance, while mountain forests across vast areas are key to carbon sequestration. As one of the important forest vertical structure parameters, forest canopy height is an important indicator of forest biomass, carbon storage, and forest health. Therefore, it is crucial to accurately estimate the height of large areas of forests, especially in complex mountainous terrains. Accurately quantifying forest canopy height has always been an important task in the national forest resource inventory mission.
[0003] However, measuring forest height in the field is time-consuming and inefficient, and it is impossible to generate large-scale forest canopy heights by relying on manual field measurements. Remote sensing technology has been proven to be an effective method for large-scale forest height inversion. Optical remote sensing images have been used for field measurements to obtain comprehensive forest canopy height products. Unfortunately, passive optical sensors have a strong saturation effect in dense forests. In addition, the obstruction of mountains under complex terrain conditions will also lead to the loss of spectral information, resulting in low estimation accuracy of forest height. Compared with optical remote sensing images, microwave remote sensing has the ability to partially penetrate the forest canopy; therefore, it has also been successfully used to estimate forest height; microwave data types can be divided into PolSAR, InSAR, and PolInSAR. PolSAR can obtain vegetation scattering characteristics, and InSAR can obtain the height of scatterers. PolInSAR has the common characteristics of PolSAR and InSAR, and can reflect the height and scattering characteristics of surface vegetation scatterers. It has great potential in the inversion of regional scale forest canopy height. However, microwave signals usually face different degrees of saturation problems, which is related to the wavelength of electromagnetic waves. Studies have shown that the L-band SAR backscatter signal in the forest is close to saturation when the tree height is 10m. PolInSAR and InSAR have strict requirements for the spatial baseline. This condition is usually difficult to fully meet the requirements in satellite-borne SAR data, such as ALOS / SAOCOM satellites. The most important point is that since SAR data is side-view imaging, signal loss will occur under the cover of the mountain, which is the most important constraint for SAR data in mountainous areas.
[0004] In contrast, laser detection and ranging (LiDAR) can penetrate the forest canopy to obtain vertical structural information. According to the platform, LiDAR systems can be divided into ground-based LiDAR, airborne LiDAR, and satellite-based LiDAR. Airborne LiDAR can quickly obtain forest canopy and internal three-dimensional structural information without being restricted by weather conditions, but it is difficult to apply to large areas due to its high acquisition cost. Satellite-based LiDAR is more suitable for forest height inversion at regional or global scales. However, these satellite-based LiDAR systems have a common problem, that is, they have discrete ground sampling footprints, but they still cannot directly provide surface-scale forest canopy height observations. Therefore, the combination of multi-source remote sensing data is an effective way to improve the accuracy of forest parameter estimation. The combination of multiple data sources can fully consider the height, scattering characteristics, and spectral characteristics of vegetation to more accurately estimate forest parameters.
[0005] Common remote sensing estimation of forest canopy height can be achieved by establishing a regression model to obtain a forest canopy height product at the surface scale. Regression model methods can be divided into parametric models and non-parametric models. Common parametric models include multivariate stepwise regression, geographically weighted regression, partial least squares regression, etc. Traditional non-parametric models are mainly machine learning, but these traditional methods of upgrading field measurement data to regional maps cannot fully capture the remote sensing information of mountainous terrain. It is easy to ignore the information of neighboring pixels, including texture patterns that are crucial to describing spatial changes. With the advancement of deep learning, deep learning algorithms have been used in various remote sensing applications, such as time series analysis, image classification, etc. These methods have shown promise in overcoming various geospatial challenges, as demonstrated by numerous researchers, such as recurrent neural networks (RNNs). Although standard RNNs have certain challenges in capturing long-term dependencies, their variants (such as LSTM and GRU) solve this problem, allowing the model to capture longer-term dependencies. Unlike traditional methods, deep learning methods can fully mine remote sensing variable information, thereby providing a more complex representation of forest canopy height distribution. Combining deep learning with multi-source remote sensing data can provide a deeper understanding of forest structural characteristics, with the potential to greatly improve regional forest height mapping efforts.
[0006] In summary, there are three main challenges in using remote sensing data to estimate forest canopy height in tropical mountain forests. The first is that different types of remote sensing data have limitations. The second is that complex terrain conditions amplify the errors of remote sensing data. The third is that there is uncertainty in the estimation model of forest parameters. Summary of the invention
[0007] The main purpose of the present invention is to provide a method and system for accurately estimating the mountain forest canopy height, aiming to achieve accurate estimation of the mountain forest canopy height at a regional scale.
[0008] To achieve the above object, the present invention provides a method for accurately estimating the height of a mountain forest canopy, comprising the following steps:
[0009] S1: Acquire a multi-source remote sensing data set; wherein the multi-source remote sensing data set includes spaceborne lidar data, microwave remote sensing data, optical remote sensing data and terrain data;
[0010] S2: extracting data features of different types of multi-source remote sensing data in the multi-source remote sensing data set to construct a remote sensing data feature set;
[0011] S3: Use recurrent neural networks to capture the long-term dependencies of sequence data, use Markov to calculate the transition relationship between sequence data states, and build a deep Markov model;
[0012] S4: Obtain the forest canopy height model from airborne radar and predict the forest canopy height based on the remote sensing data feature set and deep Markov model.
[0013] Optionally, in step S2, extracting data features from the satellite-borne laser radar data specifically includes:
[0014] S21: Extract canopy height, total canopy cover, plant area index and leaf height diversity index from spaceborne lidar data;
[0015] S22: The Kriging interpolation method in geostatistics is used to interpolate the spot data of canopy height, total canopy coverage, plant area index and leaf height diversity index into surface data as data features.
[0016] Optionally, in step S2, extracting data features from microwave remote sensing data specifically includes:
[0017] S23: Calculate the backscattering coefficient, polarization ratio, radar vegetation index and polarization decomposition parameters as data features based on the fully polarized microwave remote sensing data.
[0018] Optionally, in step S2, extracting data features from the optical remote sensing data specifically includes:
[0019] S24: Extract original single-band variables, band combination variables, vegetation index and other parameters from the optical remote sensing data as data features.
[0020] Optionally, in step S2, extracting data features from terrain data specifically includes:
[0021] S25: Calculate the terrain slope based on the digital elevation model data, and divide the slope into 6 slope levels according to the terrain slope calculation result;
[0022] S26: Convert slope grade into a dummy variable.
[0023] Optionally, in step S3, a recurrent neural network is used to capture the long-term dependency of sequence data, specifically including:
[0024] S31: Use a neural network architecture that processes sequence data to capture the long-term dependencies of each time step of the input sequence data, output the hidden state as the input for the next time step, and construct a recurrent neural network;
[0025] Among them, the expression for updating the hidden state of the recurrent neural network is specifically:
[0026] ,
[0027] in, is the input data at time step t, h is the hidden state at time t, and f is the transfer function of the recurrent neural network.
[0028] Optionally, in step S3, the transition relationship between the states of the sequence data is calculated by Markov, specifically including:
[0029] S32: Using the transition probability distribution corresponding to the previous state in the Markov model to control the state transition of the next state;
[0030] In the Markov model, the current state The expression of the probability distribution of is:
[0031] ,
[0032] in, and Respectively indicate the current status The mean and variance of are calculated by the transfer function f.
[0033] Optionally, in step S3, constructing a deep Markov model specifically includes:
[0034] S33: The output of the recurrent neural network is used to drive the calculation of the state transition parameters of the Markov model at each time, and the long-term dependency relationship in the sequence and the transition relationship between states are taken into consideration to construct a deep Markov model;
[0035] Among them, the expression of the deep Markov model is specifically:
[0036] ,
[0037] in, is the output of the recurrent neural network at time step t, and Respectively indicate the current status The mean and variance of the output of the recurrent neural network And the transfer function f is calculated.
[0038] Optionally, in step S4, obtaining a forest canopy height model of an airborne radar and predicting the forest canopy height based on a remote sensing data feature set and a deep Markov model specifically includes:
[0039] S41: Using the airborne radar forest canopy height model as the dependent variable and the remote sensing data features in the remote sensing data feature set as the independent variables, the deep Markov model is used to predict the forest canopy height.
[0040] In addition, in order to achieve the above-mentioned purpose, the present invention also provides a mountain forest canopy height accurate estimation system, comprising:
[0041] An acquisition module, used to acquire a multi-source remote sensing data set; wherein the multi-source remote sensing data set includes space-borne laser radar data, microwave remote sensing data, optical remote sensing data and terrain data;
[0042] An extraction module, used to extract data features of different types of multi-source remote sensing data in the multi-source remote sensing data set and construct a remote sensing data feature set;
[0043] A building module is used to capture the long-term dependencies of sequence data using a recurrent neural network, calculate the transition relationship between sequence data states using Markov, and build a deep Markov model;
[0044] The prediction module is used to obtain the forest canopy height model of the airborne radar and predict the forest canopy height based on the remote sensing data feature set and the deep Markov model.
[0045] The beneficial effects of the present invention are as follows: a method and system for accurately estimating the height of a mountain forest canopy is proposed, the method comprising: obtaining a multi-source remote sensing data set; wherein the multi-source remote sensing data set comprises satellite-borne laser radar data, microwave remote sensing data, optical remote sensing data and terrain data; extracting data features of different types of multi-source remote sensing data in the multi-source remote sensing data set, and constructing a remote sensing data feature set; using a recurrent neural network to capture the long-term dependency of sequence data, and using Markov to calculate the transition relationship between sequence data states, and constructing a deep Markov model; obtaining an airborne radar forest canopy height model, and predicting the forest canopy height based on the remote sensing data feature set and the deep Markov model. The present invention is based on Landsat8 optical remote sensing data, Alos-2 polarimetric SAR data, and satellite-borne laser radar GEDI data, with terrain slope as a dummy variable, and using deep learning Markov regression DMR to estimate the forest canopy height, which can effectively improve the accuracy of forest canopy height estimation under mountainous conditions. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] Figure 1 Schematic diagram of the process of the method for accurately estimating the mountain forest canopy height in this embodiment;
[0047] Figure 2 This is a schematic diagram of the model testing results based only on optical remote sensing variables;
[0048] Figure 3 These are the results of the model test based only on microwave remote sensing variables;
[0049] Figure 4 These are the model verification results based only on the spaceborne lidar GEDI remote sensing variables;
[0050] Figure 5 To test the results of the model based on optical, microwave and GEDI remote sensing variables;
[0051] Figure 6 The model verification results based on optical, microwave and GEDI remote sensing variables and terrain dummy variables in this embodiment;
[0052] Figure 7 Schematic diagram of the principle of the mountain forest canopy height accurate estimation system in the embodiment.
[0053] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION
[0054] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.
[0055] The embodiment of the present invention provides a method for accurately estimating the height of a mountain forest canopy. Figure 1 , including the following steps:
[0056] S1: Acquire a multi-source remote sensing data set; wherein the multi-source remote sensing data set includes spaceborne lidar data, microwave remote sensing data, optical remote sensing data and terrain data;
[0057] S2: extracting data features of different types of multi-source remote sensing data in the multi-source remote sensing data set to construct a remote sensing data feature set;
[0058] S3: Use recurrent neural networks to capture the long-term dependencies of sequence data, use Markov to calculate the transition relationship between sequence data states, and build a deep Markov model;
[0059] S4: Obtain the forest canopy height model from airborne radar and predict the forest canopy height based on the remote sensing data feature set and deep Markov model.
[0060] In a preferred embodiment, in step S2, extracting data features from the spaceborne laser radar data specifically includes:
[0061] S21: Extract canopy height, total canopy cover, plant area index and leaf height diversity index from spaceborne lidar data;
[0062] S22: The Kriging interpolation method in geostatistics is used to interpolate the spot data of canopy height, total canopy coverage, plant area index and leaf height diversity index into surface data as data features.
[0063] In a preferred embodiment, in step S2, extracting data features from microwave remote sensing data specifically includes:
[0064] S23: Calculate the backscattering coefficient, polarization ratio, radar vegetation index and polarization decomposition parameters as data features based on the fully polarized microwave remote sensing data.
[0065] In a preferred embodiment, in step S2, extracting data features from the optical remote sensing data specifically includes:
[0066] S24: Extract original single-band variables, band combination variables, vegetation index and other parameters from the optical remote sensing data as data features.
[0067] In a preferred embodiment, in step S2, extracting data features from terrain data specifically includes:
[0068] S25: Calculate the terrain slope based on the digital elevation model data, and divide the slope into 6 slope levels according to the terrain slope calculation result;
[0069] S26: Convert slope grade into a dummy variable.
[0070] In a preferred embodiment, in step S3, a recurrent neural network is used to capture the long-term dependency of sequence data, specifically including:
[0071] S31: Use a neural network architecture that processes sequence data to capture the long-term dependencies of each time step of the input sequence data, output the hidden state as the input for the next time step, and construct a recurrent neural network;
[0072] Among them, the expression for updating the hidden state of the recurrent neural network is specifically:
[0073] ,
[0074] in, is the input data at time step t, h is the hidden state at time t, and f is the transfer function of the recurrent neural network.
[0075] In a preferred embodiment, in step S3, the transition relationship between the sequence data states is calculated by Markov, specifically including:
[0076] S32: Using the transition probability distribution corresponding to the previous state in the Markov model to control the state transition of the next state;
[0077] In the Markov model, the current state The expression of the probability distribution of is:
[0078] ,
[0079] in, and Respectively indicate the current status The mean and variance of are calculated by the transfer function f.
[0080] In a preferred embodiment, in step S3, constructing a deep Markov model specifically includes:
[0081] S33: The output of the recurrent neural network is used to drive the calculation of the state transition parameters of the Markov model at each time, and the long-term dependency relationship in the sequence and the transition relationship between states are taken into consideration to construct a deep Markov model;
[0082] Among them, the expression of the deep Markov model is specifically:
[0083] ,
[0084] in, is the output of the recurrent neural network at time t, and Respectively indicate the current status The mean and variance of the output of the recurrent neural network And the transfer function f is calculated.
[0085] In a preferred embodiment, in step S4, the forest canopy height model of the airborne radar is obtained, and the forest canopy height is predicted based on the remote sensing data feature set and the deep Markov model, which specifically includes:
[0086] S41: Using the airborne radar forest canopy height model as the dependent variable and the remote sensing data features in the remote sensing data feature set as the independent variables, the deep Markov model is used to predict the forest canopy height.
[0087] In this embodiment, a method for accurately estimating the height of mountain forest canopy is proposed. Based on Landsat8 optical remote sensing data, Alos-2 polarimetric SAR data, and satellite-borne lidar GEDI data, the terrain slope is used as a dummy variable, and deep learning Markov regression DMR is used to estimate the height of the forest canopy, which can effectively improve the accuracy of forest canopy height estimation under mountainous conditions.
[0088] In order to explain the present application more clearly, a specific example of a method for accurately estimating the canopy height of a mountain forest of the present application is provided below.
[0089] The technical solution provided in this specific example is a method for accurately estimating the height of a mountain forest canopy based on multi-source remote sensing data and deep learning. The method is based on multi-source remote sensing data, takes slope grade as a dummy variable, and combines deep learning regression to accurately estimate the height of a mountain forest canopy, including the following steps:
[0090] Step 1: Variable interpolation of the spaceborne lidar GEDI-L2B V2.
[0091] The GEDI-L2B V2 orbital data of the study area can be downloaded for free from the EARTHDATA official website (https: / / search.earthdata.nasa.gov / ). The time span can be extended to increase the spot density. The spot diameter of the data is 25m. The data includes canopy height H, total canopy cover (COV), plant area index (PAI) and leaf height diversity index (FHD). In order to match other data, the Kriging interpolation method in geostatistics is used to interpolate the GEDI spot data into surface data.
[0092] Step 2: ALOS-2 microwave remote sensing variable extraction
[0093] ALOS-2 data comes from the Japan Aerospace Exploration Agency (JAXA). ALOS-2 is an L-band radar satellite launched by JAXA in May 2014. It transmits and receives microwaves with a wavelength of about 23 cm. The data needs to be processed by radiation calibration, polarization azimuth correction, radiation terrain correction, parameter extraction (backscatter coefficient, polarization ratio, radar vegetation index and polarization decomposition parameters), geocoding, etc.
[0094] Step 3: Extraction of Landsat8 Collection2 Level2 optical remote sensing variables.
[0095] According to the administrative division of the study area, the required Landsat8 Collection2 Level2 image data was downloaded from the official website of the Geospatial Data Cloud (http: / / www.gscloud.cn / ). The data has been radiometrically calibrated and topographically corrected. Based on the Landsat8 optical remote sensing data, the original single-band variables, band combination variables, vegetation index and other parameters were extracted.
[0096] Step 4: Setting of terrain dummy variables.
[0097] The terrain slope is calculated based on DEM data, and the slope is divided into 6 levels, and the slope level is set as a dummy variable. Complex terrain conditions are an important factor affecting the quality of remote sensing data. Whether it is SAR, optical or GEDI lidar, the slope data obtained based on DEM divides the slope into 6 levels, namely flat slope, gentle slope, slope, steep slope, abrupt slope and dangerous slope.
[0098]
[0099] In order to verify whether the slope grade division is reasonable, a nonparametric independent sample Kruskal Walis test is performed between the slope grade and the remote sensing variable. After the slope grade is divided, the slope grade is defined as a dummy variable to participate in the construction of the model, which can stratify the data status without reducing the number of samples.
[0100]
[0101] Taking into account the learning ability of the deep learning model, factors such as altitude, slope, direction, and geographical location were added to the forest canopy height estimation model to construct the model.
[0102]
[0103] Step 5: Construction of Deep Markov Model (DMR).
[0104] In the deep Markov model, the Markov model is used to generate latent states in the sequence, which correspond to the feature representation of the sequence data and are used to generate observation data at each time. The hidden state of the RNN is used to calculate the state transition of the Markov model. Specifically, the output of the RNN is used to calculate the state transition parameters of the Markov model at each time. In this way, the combination of RNN and Markov model can take into account the long-term dependencies in the sequence (captured by RNN) and the transition relationship between states (modeled by the Markov model), thereby achieving effective modeling and prediction of sequence data.
[0105] It should be noted that for recurrent neural networks (RNNs):
[0106] RNN is a neural network architecture specifically designed to process sequence data. It receives input at each time step and outputs a hidden state. This hidden state is used as input at the next time step, and so on. This recursive structure allows RNN to capture long-term dependencies in the sequence, taking into account previous information in the sequence when predicting the state at the next moment.
[0107] The hidden state of RNN (denoted as ) is updated as follows:
[0108] ,
[0109] in, is the input data at time step t, h is the hidden state at time t, and f is the transfer function of the RNN.
[0110] It should be noted that for the Markov model:
[0111] Markov process is a type of random process. Its original model, Markov chain, was proposed by Russian mathematician AA Markov in 1907. The Markov model describes a state sequence in which the probability distribution of each state depends only on the previous state. Given the current state (now), its future evolution (future) does not depend on its past evolution (past). In the Markov model, the transition between states is controlled by the probability distribution, the transition probability between states depends only on the previous state, the transition probability between states is controlled by the transition function f, and the current state The probability distribution of is:
[0112] ,
[0113] in, and Respectively indicate the current status The mean and variance of are calculated by the transfer function f.
[0114] It should be noted that for the combination of RNN and Markov model:
[0115] The deep Markov model combines RNN and Markov model. The calculation of its hidden state and state transition parameters is obtained by RNN calculation, which is specifically expressed as:
[0116] ,
[0117] in, is the output of the RNN at time step t, and Respectively indicate the current status The mean and variance of And the transfer function f is calculated.
[0118] In this way, by combining RNN and Markov model, the model can take into account the long-term dependencies in the sequence and the transition relationships between states at the same time, thereby achieving effective modeling and prediction of sequence data. Finally, the deep Markov model (DMR) is used as the regression model, the CHM of high-precision airborne LiDAR is used as the dependent variable, and the multi-source remote sensing parameters of steps 1 to 4 are used as independent variables. A regression model between the dependent variable and the independent variable is constructed based on deep Markov. In the process of model building, the data is divided into training samples and verification samples. The training samples are used to fit the deep Markov model, and the verification samples are used to test the accuracy of the model. Finally, the forest canopy height CHM is predicted based on the verification samples. The test results of the existing model are shown in the figure. Figure 2-5 As shown, the test results of the model provided in this embodiment are as follows Figure 6 As shown, the coefficient of determination R 2 It is 0.772.
[0119] like Figure 7 As shown, the embodiment of the present invention also provides a mountain forest canopy height accurate estimation system, which includes:
[0120] An acquisition module 10 is used to acquire a multi-source remote sensing data set; wherein the multi-source remote sensing data set includes space-borne laser radar data, microwave remote sensing data, optical remote sensing data and terrain data;
[0121] An extraction module 20, used to extract data features of different types of multi-source remote sensing data in the multi-source remote sensing data set, and construct a remote sensing data feature set;
[0122] A construction module 30 is used to capture the long-term dependency of sequence data by using a recurrent neural network, calculate the transition relationship between the states of the sequence data by Markov, and construct a deep Markov model;
[0123] The prediction module 40 is used to obtain the forest canopy height model of the airborne radar and predict the forest canopy height based on the remote sensing data feature set and the deep Markov model.
[0124] Other embodiments or specific implementations of the mountain forest canopy height accurate estimation system of the present invention can refer to the above-mentioned method embodiments, which will not be described in detail here.
[0125] It is understood that, in the description of this specification, the description with reference to the terms "one embodiment", "another embodiment", "other embodiments", or "first to Nth embodiments" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.
[0126] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.
[0127] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.
Claims
1. A method for accurately estimating the height of a mountain forest canopy, characterized in that: The following steps are involved: S1: Acquire a multi-source remote sensing data set; wherein the multi-source remote sensing data set includes spaceborne lidar data, microwave remote sensing data, optical remote sensing data and terrain data; S2: extracting data features of different types of multi-source remote sensing data in the multi-source remote sensing data set to construct a remote sensing data feature set; In step S2, extracting data features from the satellite-borne laser radar data specifically includes: S21: Extract canopy height, total canopy cover, plant area index and leaf height diversity index from spaceborne lidar data; S22: The Kriging interpolation method in geostatistics was used to interpolate the spot data of canopy height, total canopy coverage, plant area index and leaf height diversity index into surface data as data features; In step S2, extracting data features from microwave remote sensing data specifically includes: S23: Calculate backscattering coefficient, polarization ratio, radar vegetation index and polarization decomposition parameters as data features based on full polarization microwave remote sensing data; In step S2, extracting data features from the optical remote sensing data specifically includes: S24: extracting original single-band variables, band combination variables, and vegetation index parameters from optical remote sensing data as data features; In step S2, extracting data features from terrain data specifically includes: S25: Calculate the terrain slope based on the digital elevation model data, and divide the slope into 6 slope levels according to the terrain slope calculation result; S26: Convert slope grade into a dummy variable; S3: Use recurrent neural networks to capture the long-term dependencies of sequence data, use Markov to calculate the transition relationship between sequence data states, and build a deep Markov model; In step S3, a recurrent neural network is used to capture the long-term dependency of sequence data, which specifically includes: S31: Use a neural network architecture that processes sequence data to capture the long-term dependencies of each time step of the input sequence data, output the hidden state as the input for the next time step, and construct a recurrent neural network; Among them, the expression for updating the hidden state of the recurrent neural network is specifically: ; in, is the input data at time step t, is the hidden state output by the recurrent neural network at time step t, and f is the transfer function of the recurrent neural network; In step S3, the transition relationship between the states of the sequence data is calculated by Markov, which specifically includes: S32: Using the transition probability distribution corresponding to the previous state in the Markov model to control the state transition of the next state; In the Markov model, the current state The expression of the probability distribution of is: ; in, and Respectively indicate the current status The mean and variance of are calculated by the transfer function f of the recurrent neural network; In step S3, a deep Markov model is constructed, which specifically includes: S33: The output of the recurrent neural network is used to drive the calculation of the state transition parameters of the Markov model at each time, and the long-term dependency relationship in the sequence and the transition relationship between states are taken into consideration to construct a deep Markov model; Among them, the expression of the deep Markov model is specifically: ; in, is the hidden state output by the recurrent neural network at time step t, and Respectively indicate the current status The mean and variance of are calculated by the transfer function f of the recurrent neural network; S4: Obtain the forest canopy height model from airborne radar and predict the forest canopy height based on the remote sensing data feature set and deep Markov model.
2. The method for accurately estimating the mountain forest canopy height according to claim 1, characterized in that: In step S4, the forest canopy height model of the airborne radar is obtained, and the forest canopy height is predicted based on the remote sensing data feature set and the deep Markov model, which specifically includes: S41: Using the airborne radar forest canopy height model as the dependent variable and the remote sensing data features in the remote sensing data feature set as the independent variables, the deep Markov model is used to predict the forest canopy height.
3. A mountain forest canopy height accurate estimation system, characterized in that: The method for accurately estimating the mountain forest canopy height as claimed in any one of claims 1 to 2 comprises: An acquisition module, used to acquire a multi-source remote sensing data set; wherein the multi-source remote sensing data set includes space-borne laser radar data, microwave remote sensing data, optical remote sensing data and terrain data; An extraction module, used to extract data features of different types of multi-source remote sensing data in the multi-source remote sensing data set and construct a remote sensing data feature set; A building module is used to capture the long-term dependencies of sequence data using a recurrent neural network, calculate the transition relationship between sequence data states using Markov, and build a deep Markov model; The prediction module is used to obtain the forest canopy height model of the airborne radar and predict the forest canopy height based on the remote sensing data feature set and the deep Markov model.
