Forest carbon reserve estimation method based on artificial intelligence and multi-modal remote sensing data

Abstract

The invention discloses a forest carbon reserve estimation method based on artificial intelligence and multi-modal remote sensing data, and relates to the technical field of forest carbon reserve estimation, and the method integrates optical satellite images, laser radar data and hyperspectral image remote sensing data, improves the comprehensiveness and estimation precision of the data through multi-source data fusion, and improves the estimation precision of the forest carbon reserve. The method can more accurately reflect the coverage condition, the vertical structure and the biological diversity of the forest, overcomes the limitation of a single data source, divides the whole forest region into a plurality of sub-regions, independently constructs a carbon reserve estimation model based on the forest type of each sub-region, refines the carbon reserve estimation of different regions in a region-by-region estimation mode, and improves the estimation accuracy. According to the method, characteristics of different forest types are accurately reflected, a dynamic carbon reserve prediction model based on a differential equation is introduced, historical and real-time remote sensing data are combined, changes of carbon reserves in the time dimension are simulated, and influences of factors such as forest growth, climate changes, natural disasters and felling on the carbon reserves are captured.

Description

Technical Field

[0001] The present invention relates to the technical field of forest carbon stock estimation, in particular to a forest carbon stock estimation method based on artificial intelligence and multimodal remote sensing data. Background Art

[0002] In the study of global carbon cycle and climate change response, forest carbon storage is one of the key indicators of the study. The remote sensing technology used provides the main data support for large-scale forest carbon storage estimation. Different remote sensing technologies can capture forest information from different angles, such as optical satellite images and LiDAR.

[0003] Most of the current forest carbon stock estimation methods use one of the above-mentioned remote sensing technologies as a means of information capture. Among them, optical satellite images are widely used to estimate forest coverage, but they are not sensitive enough to information such as tree height and crown structure. Although LiDAR can provide detailed three-dimensional forest structure data, it can only cover part of the area and is restricted by weather conditions. The single data source collection leads to insufficient accuracy of the estimation results. In addition, different types of forests, such as tropical rainforests, coniferous forests, and broad-leaved forests, also have significant differences in carbon storage and biomass distribution. In order to ensure the accuracy of carbon stock estimation, it is necessary to accurately classify forest types. However, the current forest classification method still uses manual division or predefined classification, which has low classification efficiency and strong subjectivity.

[0004] To address the problem of single data source, most existing estimation methods use a fusion of multiple data sources to solve the problem, such as linear regression and principal component analysis. However, such methods require a linear relationship between different data sources and cannot effectively handle nonlinear correlations between data. In addition, there are significant differences in data formats, temporal resolution, and spatial resolutions of different remote sensing technologies. It is difficult to effectively coordinate the differences during data fusion, which can easily lead to data mismatch. In terms of forest type classification, most existing estimation methods use remote sensing technology and regional empirical models to classify forest types, but they rely too much on optical images. The optical images obtained by remote sensing technology cannot guarantee clarity, which makes it difficult to classify forest types in a fine-grained manner. Some vegetation types are very similar in optical images, and insufficient classification accuracy will directly affect the zoning results of carbon stock estimation. In addition, when regional empirical models are applied to new geographical regions or climatic conditions, their accuracy is not good, and the model must be re-optimized. Cross-regional application is very difficult. Therefore, there is an urgent need for a forest carbon stock estimation method based on artificial intelligence and multimodal remote sensing data to solve such problems. Summary of the invention

[0005] In view of the above existing problems, the present invention is proposed.

[0006] Therefore, the present invention provides a forest carbon stock estimation method based on artificial intelligence and multimodal remote sensing data to solve the problem that the principal component analysis and regression model fusion in the existing solutions can only capture the linear relationship between data sources, ignoring the nonlinear characteristics of forest structure and dynamic changes. The more complex the remote sensing data, the greater the limitations of such methods, and the empirical forest classification method is difficult to adapt to the forest types and climatic conditions in different regions.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0008] The present invention provides a forest carbon storage estimation method based on artificial intelligence and multimodal remote sensing data, which comprises:

[0009] Step S1, multi-source remote sensing data acquisition, collecting remote sensing data, including forest coverage information, vertical structure information, spectral feature information, and surface and canopy information;

[0010] Step S2, remote sensing data feature extraction, remote sensing data fusion, extraction of carbon storage estimation and forest classification related features, the related features include: forest structure, tree species type, canopy height and vegetation index;

[0011] Step S3, forest type classification, based on the features extracted in step S2, using a classification algorithm to classify the forest type;

[0012] Step S4, constructing a carbon storage model, constructing a carbon storage estimation model for different types of forests, and estimating the carbon storage of forests region by region based on the classification of forest types and combining carbon storage estimation models of different regions;

[0013] Step S5, dynamic carbon storage change prediction, uses historical and real-time remote sensing data to build a prediction model, combines forest types and the carbon storage estimation results of step S4, and analyzes the dynamic changes of carbon storage.

[0014] Furthermore, the remote sensing data is converted into data format, standardized, and noise and outliers are removed;

[0015] The multi-source here refers to multi-source data including optical satellite images, lidar data, and hyperspectral images.

[0016] Furthermore, the remote sensing data in step S1 includes:

[0017] Optical satellite imagery: LiDAR data: Hyperspectral imaging: Among them, H and W represent the height and width of the image, respectively, and C1, C2, and C3 represent the number of channels of optical, LiDAR, and hyperspectral images, respectively;

[0018] Convolutional neural network is used to extract spatial features. For optical satellite images, the convolution operation is expressed as:

[0019] F optical =σ(W conv1 *X optical +b conv1 ),

[0020] The feature extraction of LiDAR data and hyperspectral data are:

[0021] F LiDAR =σ(W conv2 *X LiDAR +b conv2 ),

[0022] F hyperspectral =σ(W conv3 *X hyperspectral +b conv3 ),

[0023] in, are the feature maps extracted from optical, LiDAR and hyperspectral images, respectively. D1, D2 and D3 represent the number of channels of each feature map after feature extraction of optical image, LiDAR data and hyperspectral image, respectively. H′×W′ represents the spatial size of the feature map after convolution operation. W conv1 ,W conv2 ,W conv3 is the convolution kernel weight matrix, b conv1 ,b conv2 ,b conv3 are the respective bias terms, * denotes the convolution operation, and σ(·) is the ReLU activation function.

[0024] Furthermore, remote sensing data fusion is performed in step S2:

[0025] The features of each modality are fused and reduced in dimension using an autoencoder, and the features of each modality are concatenated and expressed as:

[0026]

[0027] Then the autoencoder is used to reduce the dimension and extract the fusion features:

[0028] F encoded =σ(W encoder ·F concat +b encoder ), where F encoded ∈R H′×W′×D represents the fusion features after dimensionality reduction, W encoder represents the weight matrix of the autoencoder, b encoderrepresents the bias term of the autoencoder, D represents the feature dimension after dimensionality reduction, and σ(·) is the ReLU activation function;

[0029] Introduce the attention mechanism and calculate the attention weight:

[0030] Among them, α i,j represents the attention weight of feature position i,j, q i and k j They represent the query vector and key vector respectively, which are obtained from the feature map using linear transformation. N represents the total number of features. Represents a transpose operation;

[0031] Use attention weights to perform weighted summation of features:

[0032] Among them, v j The value vector representing the feature, F attention Represents the fused features after weighting by the attention mechanism.

[0033] Furthermore, in step S3, the forests are divided into different types including coniferous forests, broad-leaved forests and tropical rain forests according to the features extracted after fusion, and the different types of forests are divided into regions according to the classification results.

[0034] Furthermore, the forest type classification in step S3 is performed as follows:

[0035] The fusion feature in step S2 is F attention ∈R H′×W′×D , convert the feature map into a vector and perform global average pooling on each channel:

[0036] Among them, F attention (i, j) represents the eigenvalue of the feature map at position (i, j), F pooled ∈R D represents the feature vector after pooling, retaining the global information of each channel and used as the input of the classifier. H′, W′ represents the spatial size of the feature map, and D represents the number of channels of the feature map.

[0037] Use support vector machine for classification and construct optimization objectives:

[0038] Among them, w represents the normal vector of the classification hyperplane, b represents the bias term of the classification hyperplane, and y i represents the category label of sample i, with a value of {-1,+1}, |w| 2 is the regularization term, Represents a transpose operation;

[0039] The classification decision function of SVM is: in, is the predicted category, i.e., forest type, and sign(·) represents the sign function.

[0040] Furthermore, the method of classifying forest types in step S3 also includes:

[0041] The loss function is defined as: Among them, L represents the cross entropy loss value, y i,c represents the true label of sample i, using onehot encoding. Represents the predicted probability that the classifier belongs to category c for sample i, and uses the gradient descent method to update the classifier parameters: Among them, θ represents the parameters of the classifier, including w, b, η represents the learning rate, and controls the update step size. Represents the gradient of the loss function L with respect to the parameter θ. After the training is completed, the new test data feature F is input pooled , use the trained classifier to predict forest type

[0042] In step S3, the feature F extracted in step S2 is used pooled ,The features are vectorized by global average pooling, and then the ,support vector machine is used to build a classification model to classify the forest types. ,During the training process, the cross entropy loss function is used, and the ,parameters of the classifier are optimized by gradient descent.

[0043] Furthermore, the carbon storage estimation model in step S4 is constructed as follows:

[0044] A carbon storage estimation model is designed for each forest type T, and the carbon storage is expressed as:

[0045] C T =β T ·B T , where C T is the total carbon stock of forest type T, β T is the carbon conversion coefficient, which indicates the conversion ratio between forest biomass and carbon storage, B T is the biomass of forest type T;

[0046] Forest biomass B T The estimation formula is: Among them, a T ,b T ,c T is the empirical coefficient related to forest type T, obtained through field measurements, H T is the average height of trees in forest type T, D T is the average DBH of trees in forest type T;

[0047] The biomass estimation formula is adjusted according to the tree species characteristics of different forest types. T ,b T ,c T It has forest type dependency;

[0048] After the forest type classification, the entire forest area is divided into multiple sub-areas according to different types. Assume that the total forest area is divided into N sub-areas, and the forest type of each sub-area i is T i , its carbon storage C i The estimation formula is:

[0049] Among them, C i is the carbon stock of the ith subregion, is the forest type T to which the sub-area belongs i The associated carbon conversion factor, is the forest type T of the sub-area i The biomass per unit area,

[0050] A i is the area of ​​the sub-region;

[0051] Sum the carbon stocks of all sub-regions: Among them, C total is the total carbon storage of the entire forest area, C i is the carbon stock of the ith sub-region; for different types of forests, a carbon stock estimation model is constructed based on the biomass estimation formula. The biomass and carbon conversion coefficient of the forest type are calculated region by region, and the carbon stock is estimated region by region based on the regional area. Then, the carbon stock of all sub-regions is accumulated to obtain the total carbon stock of the entire forest area.

[0052] Furthermore, in step S5, the method of predicting the dynamic carbon storage change is:

[0053] The carbon storage C is modeled using a first-order differential equation i (t) Changes in time t:

[0054] Among them, C i (t) is the carbon storage of the ith sub-region at time t, is the carbon storage growth rate of the ith region, L i (t) is the carbon loss of the ith region at time t, represents the rate of change of carbon storage in the ith region over time t;

[0055] Carbon stock growth rate Depending on the forest type and biomass growth model, the relationship between the growth rate of carbon storage and the growth of forest biomass is expressed as: Among them, r max The maximum growth rate of this forest type, B i (t) is the biomass of the ith region at time t, B max The maximum biomass of the ith region is the maximum forest biomass value that the region can reach. When the biomass approaches the maximum value, the growth rate tends to zero.

[0056] Furthermore, in step S5, the method of predicting the dynamic carbon storage change also includes:

[0057] Construct a biomass growth model to describe the growth of biomass:

[0058] Among them, B i (t) is the biomass of the ith region at time t, B max The maximum biomass of the area, k is the growth rate parameter, t0 is the initial moment of biomass growth, and m is a parameter that affects the shape of the curve;

[0059] Construct a carbon loss model to describe carbon loss:

[0060] L i (t) = L0·e -λ·t , where L i (t) is the carbon loss of the ith region at time t, L0 is the initial value of carbon loss, and λ is the loss decay rate, which indicates the decreasing speed of carbon loss over time;

[0061] Combining the biomass growth model and the carbon loss model, the change of carbon storage over time t is expressed as:

[0062] Among them, C i (0) represents the initial carbon storage of the ith region, represents the net carbon stock increase, is the cumulative carbon stock change from time 0 to t;

[0063] The carbon storage of the entire forest area is the accumulation of carbon storage in each sub-area. For N sub-areas, the change of total carbon storage over time t is: Among them, C total (t) is the carbon storage of the entire forest area at time t, C i (t) is the carbon stock of the ith region at time t;

[0064] The dynamic changes of total carbon storage in the entire forest area over time were obtained by estimating area by area.

[0065] The beneficial effects of the present invention are:

[0066] The present invention integrates optical satellite imagery, lidar data and hyperspectral imagery remote sensing data, making full use of the complementarity of different data sources. Optical images provide spatial resolution information, lidar data captures the vertical structure of the forest, and hyperspectral images reflect the spectral characteristics of different forest types. Through multi-source data fusion, the comprehensiveness and estimation accuracy of the data are greatly improved, which can more accurately reflect the forest coverage, vertical structure and biodiversity, and overcome the limitations of a single data source.

[0067] The present invention introduces an attention mechanism that can adaptively focus on key features related to carbon stock estimation, which not only improves the accuracy of feature extraction, but also avoids the interference of redundant information on the results through data dimensionality reduction, thereby further improving the estimation accuracy.

[0068] The present invention divides the entire forest area into multiple sub-areas, and constructs a carbon stock estimation model based on the forest type of each sub-area. The carbon stock estimation of different regions is refined by estimating area by area, and the characteristics of different forest types are accurately reflected. Specific biomass models and carbon conversion coefficients are constructed for different types of forests, such as coniferous forests, broad-leaved forests and tropical rainforests, respectively, thereby improving the accuracy and regional adaptability of carbon stock estimation.

[0069] The present invention introduces a dynamic carbon storage prediction model based on differential equations, combines historical and real-time remote sensing data, simulates the changes in carbon storage in the time dimension, establishes a logarithmic relationship between the carbon storage growth rate and forest biomass, captures the impact of factors such as forest growth, climate change, natural disasters, and deforestation on carbon storage, and can predict future changing trends. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0071] Figure 1 It is a schematic diagram of the flow chart of the forest carbon stock estimation method of the present invention. DETAILED DESCRIPTION

[0072] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0073] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0074] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0075] Example 1, reference Figure 1 , this embodiment provides a forest carbon stock estimation method based on artificial intelligence and multimodal remote sensing data, comprising the following steps:

[0076] Step S1, multi-source remote sensing data acquisition, collecting remote sensing data, including forest coverage information, vertical structure information, spectral feature information, and surface and canopy information;

[0077] Convert remote sensing data into different formats, standardize it, and remove noise and outliers;

[0078] Here, multi-source refers to multi-source data including optical satellite images, lidar data, and hyperspectral images.

[0079] The remote sensing data in step S1 includes:

[0080] Optical satellite imagery: LiDAR data: Hyperspectral imaging: Among them, H and W represent the height and width of the image, respectively, and C1, C2, and C3 represent the number of channels of optical, LiDAR, and hyperspectral images, respectively;

[0081] Convolutional neural network is used to extract spatial features. For optical satellite images, the convolution operation is expressed as:

[0082] F optical =σ(W conv1 *X optical +b conv1 ),

[0083] The feature extraction of LiDAR data and hyperspectral data are:

[0084] F LiDAR =σ(W conv2 ·X LiDAR +b conv2 ),

[0085] F hyperspectral =σ(W conv3 *X hyperspectral +b conv3 ),

[0086] in, are the feature maps extracted from optical, LiDAR and hyperspectral images, respectively. D1, D2 and D3 represent the number of channels of each feature map after feature extraction of optical image, LiDAR data and hyperspectral image, respectively. H′×W′ represents the spatial size of the feature map after convolution operation. W conv1 ,W conv2 ,W conv3 is the convolution kernel weight matrix, b conv1 ,b conv2 ,b conv3 are their respective bias terms, * denotes the convolution operation, and σ(·) is the ReLU activation function;

[0087] Specifically, the introduction of multi-source remote sensing data provides the model with spatial, structural and spectral information, extracts spatial features through convolutional neural networks, and combines autoencoders for dimensionality reduction and fusion processing to improve the utilization efficiency of multimodal information and avoid information shortage or bias caused by a single data source. The introduction of attention mechanism enables the model to better highlight key features related to carbon storage and enhance the ability to understand complex forest structures.

[0088] Step S2, remote sensing data feature extraction, remote sensing data fusion, extraction of carbon storage estimation and forest classification related features, the related features include: forest structure, tree species type, canopy height and vegetation index;

[0089] In step S2, remote sensing data fusion is performed:

[0090] The features of each modality are fused and reduced in dimension using an autoencoder, and the features of each modality are concatenated and expressed as:

[0091]

[0092] Then the autoencoder is used to reduce the dimension and extract the fusion features:

[0093] F encoded =σ(W encoder ·F concat +b encoder ), where F encoded ∈R H′×W′×D represents the fusion features after dimensionality reduction, W encoder represents the weight matrix of the autoencoder, b encoder represents the bias term of the autoencoder, D represents the feature dimension after dimensionality reduction, and σ(·) is the ReLU activation function;

[0094] Introduce the attention mechanism and calculate the attention weight:

[0095] Among them, α i,j represents the attention weight of feature position i,j, q i and k j They represent the query vector and key vector respectively, which are obtained from the feature map using linear transformation. N represents the total number of features. Represents a transpose operation;

[0096] Use attention weights to perform weighted summation of features:

[0097] Among them, v j The value vector representing the feature, F attention Represents the fusion features after weighting by the attention mechanism;

[0098] Specifically, based on the extracted multimodal features, support vector machine (SVM) is used for classification, and global average pooling is used to retain the information in the feature vectorization process. SVM is combined with the gradient descent method to optimize the parameters of the classifier, so that the classifier can still effectively cope with the complexity and diversity of different forest types when dealing with different types of forests. Through forest type classification, the forest area can be divided more finely.

[0099] Step S3, forest type classification, based on the features extracted in step S2, using a classification algorithm to classify the forest type;

[0100] In step S3, the forests are divided into different types including coniferous forests, broad-leaved forests and tropical rain forests according to the features extracted after fusion, and different types of forests are divided into regions according to the classification results;

[0101] The method of forest type classification in step S3 is:

[0102] The fusion feature in step S2 is F attention ∈R H′×W′×D , convert the feature map into a vector and perform global average pooling on each channel:

[0103] Among them, F attention (i, j) represents the eigenvalue of the feature map at position (i, j), F pooled ∈R D represents the feature vector after pooling, retaining the global information of each channel and used as the input of the classifier. H′, W′ represents the spatial size of the feature map, and D represents the number of channels of the feature map.

[0104] Use support vector machine for classification and construct optimization objectives:

[0105] Among them, w represents the normal vector of the classification hyperplane, b represents the bias term of the classification hyperplane, and y i represents the category label of sample i, with a value of {-1,+1}, |w| 2 is the regularization term, Represents a transpose operation;

[0106] The classification decision function of SVM is: in, is the predicted category, i.e., forest type, and sign(·) represents the sign function;

[0107] The method of classifying forest types in step S3 also includes:

[0108] The loss function is defined as: Among them, L represents the cross entropy loss value, y i,c represents the true label of sample i, using onehot encoding. Represents the predicted probability that the classifier belongs to category c for sample i, and uses the gradient descent method to update the classifier parameters: Among them, θ represents the parameters of the classifier, including w, b, η represents the learning rate, and controls the update step size. Represents the gradient of the loss function L with respect to the parameter θ. After the training is completed, the new test data feature F is input pooled , use the trained classifier to predict forest type

[0109] In step S3, the feature F extracted in step S2 is used pooled ,The features are vectorized through global average pooling, and then the classification model is constructed using support vector machine to classify the forest types.,During the training process, the cross entropy loss function is used, and the classifier parameters are optimized by gradient descent;

[0110] Specifically, the carbon storage estimation model estimates carbon storage for different forest types. By dividing the sub-regions, it conducts differentiated modeling for the ecological characteristics, carbon absorption capacity, and tree growth patterns of different types of forests. The region-by-region estimation effectively improves the spatial accuracy of carbon storage estimation, solving the problem of large-scale and diverse forest ecosystems that are difficult to cope with with traditional estimation methods.

[0111] Step S4, constructing a carbon storage model, constructing a carbon storage estimation model for different types of forests, and estimating the carbon storage of forests region by region based on the classification of forest types and combining carbon storage estimation models of different regions;

[0112] The carbon stock estimation model in step S4 is constructed as follows:

[0113] A carbon storage estimation model is designed for each forest type T, and the carbon storage is expressed as:

[0114] C T =β T ·B T , where C T is the total carbon stock of forest type T, β T is the carbon conversion coefficient, which indicates the conversion ratio between forest biomass and carbon storage, B T is the biomass of forest type T;

[0115] Forest biomass B T The estimation formula is: Among them, a T ,b T ,c T is the empirical coefficient related to forest type T, obtained through field measurements, H T is the average height of trees in forest type T, D T is the average DBH of trees in forest type T;

[0116] The biomass estimation formula is adjusted according to the tree species characteristics of different forest types. T ,b T ,c T It has forest type dependency;

[0117] After the forest type classification, the entire forest area is divided into multiple sub-areas according to different types. Assume that the total forest area is divided into N sub-areas, and the forest type of each sub-area i is T i , its carbon storage C i The estimation formula is:

[0118] Among them, C i is the carbon stock of the ith subregion, is the forest type T to which the sub-area belongs i The associated carbon conversion factor, is the forest type T of the sub-area i The biomass per unit area,

[0119] A i is the area of ​​the sub-region;

[0120] Sum the carbon stocks of all sub-regions: Among them, C total is the total carbon storage of the entire forest area, C iis the carbon storage of the ith sub-region; for different types of forests, a carbon storage estimation model is constructed based on the biomass estimation formula. The biomass and carbon conversion coefficient of the forest type are calculated region by region, and the carbon storage is estimated region by region based on the regional area. Then, the carbon storage of all sub-regions is accumulated to obtain the total carbon storage of the entire forest area.

[0121] Step S5, dynamic carbon storage change prediction, using historical and real-time remote sensing data to build a prediction model, combining forest types and the carbon storage estimation results of step S4 to analyze the dynamic changes of carbon storage;

[0122] In step S5, the method of predicting the dynamic carbon storage change is:

[0123] The carbon storage C is modeled using a first-order differential equation i (t) Changes in time t:

[0124] Among them, C i (t) is the carbon storage of the ith sub-region at time t, is the carbon storage growth rate of the ith region, L i (t) is the carbon loss of the ith region at time t, represents the rate of change of carbon storage in the ith region over time t;

[0125] Carbon stock growth rate Depending on the forest type and biomass growth model, the relationship between the growth rate of carbon storage and the growth of forest biomass is expressed as: Among them, r max The maximum growth rate of this forest type, B i (t) is the biomass of the ith region at time t, B max The maximum biomass of the ith region is the maximum forest biomass value that the region can reach. When the biomass approaches the maximum value, the growth rate tends to zero.

[0126] In step S5, the method of predicting the dynamic carbon storage change also includes:

[0127] Construct a biomass growth model to describe the growth of biomass:

[0128] Among them, B i (t) is the biomass of the ith region at time t, B max The maximum biomass of the area, k is the growth rate parameter, t0 is the initial moment of biomass growth, and m is a parameter that affects the shape of the curve;

[0129] Construct a carbon loss model to describe carbon loss:

[0130] L i (t) = L0·e -λ·t , where L i (t) is the carbon loss of the ith region at time t, L0 is the initial value of carbon loss, and λ is the loss decay rate, which indicates the decreasing speed of carbon loss over time;

[0131] Combining the biomass growth model and the carbon loss model, the change of carbon storage over time t is expressed as:

[0132] Among them, C i (0) represents the initial carbon storage of the ith region, represents the net carbon stock increase, is the cumulative carbon stock change from time 0 to t;

[0133] The carbon storage of the entire forest area is the accumulation of carbon storage in each sub-area. For N sub-areas, the change of total carbon storage over time t is: Among them, C total (t) is the carbon storage of the entire forest area at time t, C i (t) is the carbon stock of the ith region at time t;

[0134] The dynamic changes of total carbon storage in the entire forest area over time were obtained by estimating area by area;

[0135] Specifically, the changes in carbon storage over time are modeled using first-order differential equations, and historical and real-time remote sensing data are combined to dynamically capture the growth and loss process of forest carbon storage. The relationship between biomass growth rate and maximum forest biomass is introduced, and the exponential decay model of carbon loss is used to describe the impact of natural disasters, deforestation and degradation factors on carbon storage. This enables carbon storage forecasts to respond to complex natural processes and more accurately reflect the temporal changes in carbon storage.

[0136] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A forest carbon stock estimation method based on artificial intelligence and multimodal remote sensing data, characterized by: include, Step S1, multi-source remote sensing data acquisition, collecting remote sensing data, including forest coverage information, vertical structure information, spectral feature information, and surface and canopy information; Step S2, remote sensing data feature extraction, remote sensing data fusion, extraction of carbon storage estimation and forest classification related features, the related features include: forest structure, tree species type, canopy height and vegetation index; Step S3, forest type classification, based on the features extracted in step S2, using a classification algorithm to classify the forest type; Step S4, constructing a carbon storage model, constructing a carbon storage estimation model for different types of forests, and estimating the carbon storage of forests region by region based on the classification of forest types and combining carbon storage estimation models of different regions; Step S5, dynamic carbon storage change prediction, uses historical and real-time remote sensing data to build a prediction model, combines forest types and the carbon storage estimation results of step S4, and analyzes the dynamic changes of carbon storage.

2. The method for estimating forest carbon stocks based on artificial intelligence and multimodal remote sensing data according to claim 1, characterized in that: Convert remote sensing data into different formats, standardize it, and remove noise and outliers; The multi-source here refers to multi-source data including optical satellite images, lidar data, and hyperspectral images.

3. The method for estimating forest carbon stocks based on artificial intelligence and multimodal remote sensing data according to claim 2, characterized in that: The remote sensing data in step S1 includes: Optical satellite imagery: LiDAR data: Hyperspectral imaging: Among them, H and W represent the height and width of the image, respectively, and C1, C2, and C3 represent the number of channels of optical, LiDAR, and hyperspectral images, respectively; Convolutional neural network is used to extract spatial features. For optical satellite images, the convolution operation is expressed as: F optical =σ(W conv1 *X optical +b conv1 ), The feature extraction of LiDAR data and hyperspectral data are: F LiDAR =σ(W conv2 *X LiDAR +b conv2 ), F hyperspectral =σ(W conv3 *X hyperspectral +b conv3 ), in, are the feature maps extracted from optical, LiDAR and hyperspectral images, respectively. D1, D2 and D3 represent the number of channels of each feature map after feature extraction of optical image, LiDAR data and hyperspectral image, respectively. H′×W′ represents the spatial size of the feature map after convolution operation. W conv1 ,W conv2 ,W conv3 is the convolution kernel weight matrix, b conv1 ,b conv2 ,b conv3 are the respective bias terms, * denotes the convolution operation, and σ(·) is the ReLU activation function.

4. The method for estimating forest carbon stocks based on artificial intelligence and multimodal remote sensing data according to claim 3, characterized in that: In step S2, remote sensing data fusion is performed: The features of each modality are fused and reduced in dimension using an autoencoder, and the features of each modality are concatenated and expressed as: Then the autoencoder is used to reduce the dimension and extract the fusion features: F encoded =σ(W encoder ·F concat +b encoder ), where F encoded ∈R H′×W′×D represents the fusion features after dimensionality reduction, W encoder represents the weight matrix of the autoencoder, b encoder represents the bias term of the autoencoder, D represents the feature dimension after dimensionality reduction, and σ(·) is the ReLU activation function; Introduce the attention mechanism and calculate the attention weight: Among them, α i,j represents the attention weight of feature position i,j, q i and k j They represent the query vector and key vector respectively, which are obtained from the feature map using linear transformation, N represents the total number of features, and T represents the transposition operation; Use attention weights to perform weighted summation of features: Among them, v j The value vector representing the feature, F attention Represents the fused features after weighting by the attention mechanism.

5. The method for estimating forest carbon stocks based on artificial intelligence and multimodal remote sensing data according to claim 4, characterized in that: In step S3, the forests are divided into different types including coniferous forests, broad-leaved forests and tropical rain forests according to the features extracted after fusion, and the different types of forests are divided into regions according to the classification results.

6. The method for estimating forest carbon stocks based on artificial intelligence and multimodal remote sensing data according to claim 5, characterized in that: The method of forest type classification in step S3 is: The fusion feature in step S2 is F attention ∈R H′×W′×D , convert the feature map into a vector and perform global average pooling on each channel: Among them, F attention (i, j) represents the eigenvalue of the feature map at position (i, j), F pooled ∈R D represents the feature vector after pooling, retaining the global information of each channel and used as the input of the classifier. H′, W′ represents the spatial size of the feature map, and D represents the number of channels of the feature map. Use support vector machine for classification and construct optimization objectives: Among them, w represents the normal vector of the classification hyperplane, b represents the bias term of the classification hyperplane, and y i represents the category label of sample i, with a value of {-1,+1}, |w| 2 is the regularization term, Represents a transpose operation; The classification decision function of SVM is: in, is the predicted category, i.e., forest type, and sign(·) represents the sign function.

7. The method for estimating forest carbon stocks based on artificial intelligence and multimodal remote sensing data according to claim 6, characterized in that: The method of classifying forest types in step S3 also includes: The loss function is defined as: Among them, L represents the cross entropy loss value, y i,c represents the true label of sample i, using onehot encoding. It represents the predicted probability that the sample i belongs to category c, and the classifier parameters are updated using the gradient descent method: Among them, θ represents the parameters of the classifier, including w, b, η represents the learning rate, which controls the update step size, Represents the gradient of the loss function L with respect to the parameter θ. After the training is completed, the new test data feature F is input pooled , use the trained classifier to predict forest type 8. The method for estimating forest carbon stocks based on artificial intelligence and multimodal remote sensing data according to claim 7, characterized in that: The carbon stock estimation model in step S4 is constructed as follows: A carbon storage estimation model is designed for each forest type T, and the carbon storage is expressed as: C T =β T ·B T , where C T is the total carbon stock of forest type T, β T is the carbon conversion coefficient, which indicates the conversion ratio between forest biomass and carbon storage, B T is the biomass of forest type T; Forest biomass B T The estimation formula is: Among them, a T ,b T ,c T is the empirical coefficient related to forest type T, obtained through field measurements, H T is the average height of trees in forest type T, D T is the average DBH of trees in forest type T; After the forest type classification, the entire forest area is divided into multiple sub-areas according to different types. Assume that the total forest area is divided into N sub-areas, and the forest type of each sub-area i is T i , its carbon storage C i The estimation formula is: Among them, C i is the carbon stock of the ith subregion, is the forest type T to which the sub-area belongs i The associated carbon conversion factor, is the forest type T of the sub-area i Biomass per unit area, A i is the area of ​​the sub-region; Sum the carbon stocks of all sub-regions: Among them, C total is the total carbon storage of the entire forest area, C i is the carbon stock of the ith sub-region.

9. The method for estimating forest carbon stocks based on artificial intelligence and multimodal remote sensing data according to claim 8, characterized in that: In step S5, the method of predicting the dynamic carbon storage change is: The carbon storage C is modeled using a first-order differential equation i (t) Changes in time t: Among them, C i (t) is the carbon storage of the ith sub-region at time t, is the carbon storage growth rate of the ith region, L i (t) is the carbon loss of the ith region at time t, represents the rate of change of carbon storage in the ith region over time t; Carbon stock growth rate Depending on the forest type and biomass growth model, the relationship between the growth rate of carbon storage and the growth of forest biomass is expressed as: Among them, r max The maximum growth rate of this forest type, B i (t) is the biomass of the ith region at time t, B max The maximum biomass of the ith region is the maximum forest biomass value that can be achieved in the region.

10. The method for estimating forest carbon stocks based on artificial intelligence and multimodal remote sensing data according to claim 9, characterized in that: In step S5, the method of predicting the dynamic carbon storage change also includes: Construct a biomass growth model to describe the growth of biomass: Among them, B i (t) is the biomass of the ith region at time t, B max The maximum biomass of the area, k is the growth rate parameter, t0 is the initial moment of biomass growth, and m is a parameter that affects the shape of the curve; Construct a carbon loss model to describe carbon loss: L i (t) = L0·e -λ·t , where L i (t) is the carbon loss of the ith region at time t, L0 is the initial value of carbon loss, and λ is the loss decay rate, which indicates the decreasing speed of carbon loss over time; Combining the biomass growth model and the carbon loss model, the change of carbon storage over time t is expressed as: Among them, C i (0) represents the initial carbon storage of the ith region, represents the net carbon stock increase, is the cumulative carbon stock change from time 0 to t; The carbon storage of the entire forest area is the accumulation of carbon storage in each sub-area. For N sub-areas, the change of total carbon storage over time t is: Among them, C total (t) is the carbon storage of the entire forest area at time t, C i (t) is the carbon stock of the ith region at time t.

Images