A dynamic forest health assessment method based on a scientific computing large model

Abstract

The application discloses a kind of dynamic forest health assessment methods based on scientific computing large model, belong to computer model field, including steps: based on multi-modal data constructs forest space-time dataset D;Dynamic forest health assessment network is constructed, including embedding layer, numerical processing layer, mode alignment splicing layer, Transform encoder, text word embedding layer, projection layer and decoder;With D to decision text as expected output, train dynamic forest health assessment network, obtain dynamic forest health assessment model, for generating the predicted decision text of the measured forest class.The application introduces multi-modal fusion and mechanism constraint scientific computing large model, greatly improves the quality and application value of dynamic forest health assessment, by solving multi-modal dynamic modeling and physical consistency problem, provides a new technical path for forest health state assessment, especially in dynamic forest monitoring and risk early warning has wide application prospect.

Description

Technical Field

[0001] This invention relates to the field of specific computer modeling technology, and in particular to a dynamic forest health assessment method based on a large scientific computing model. Background Technology

[0002] With the rapid development of scientific computing and large model technology, large models based on scientific computing have shown great potential in forest health assessment, disaster early warning, and resource management. As a generative assessment model, scientific computing large models, by introducing multimodal feature fusion and mechanism constraint mechanisms, can not only significantly improve the accuracy and timeliness of forest health status assessment, but also ensure the consistency and stability of cross-modal information, thereby promoting the development of forest digital twins and smart forestry. In dynamic forest health assessment, ensuring the fusion quality of remote sensing images, meteorological data and ground sensor data is crucial for accurate monitoring and scientific decision-making. Application requirements often include temporal integrity, physical rationality and interpretability to meet the needs of disaster prevention and sustainable management. However, existing scientific computing large models still have problems in forest health assessment: (1) They lack a unified expression for multi-source data. For example, satellite images, LiDAR, meteorological monitoring and forest surveys have large differences in data dimensions. Existing methods often use independent processing or simple splicing, lacking a unified embedding and expression mechanism, resulting in redundancy or loss of information between modalities. (2) There are shortcomings in multimodal alignment and time-series dynamic modeling. (3) The ability to generate decisions is lacking. Existing methods mostly remain at the level of indicator monitoring or risk classification, and fail to form interpretable and operable decision text outputs. Summary of the Invention

[0003] The purpose of this invention is to provide a dynamic forest health assessment method based on a large scientific computing model that addresses the shortcomings in multimodal data alignment and temporal dynamic modeling.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a dynamic forest health assessment method based on a large scientific computing model, comprising the following steps; S1, construct the forest spatiotemporal dataset D, including S11~S13; S11, the forest is divided into multiple forest compartments, and an observation window [t0,t1] is preset, where t0 and t1 are the start and end times of the observation window, respectively. S12, for one forest compartment s, collect image data I(t,s), numerical data n(t,s), state text ST, and decision text Y at time t. * The image dataset includes three types of images: satellite remote sensing images I sat (t,s), orthophoto I ortho(t,s) and LiDAR canopy height model I lidar (t,s); S13, generate sample D of the forest compartment at time t (s). t =(I(t,s),n(t,s),ST,Y * ), where Y * D t The weak labels are used to construct the forest spatiotemporal dataset D from all samples at time [t0,t1] of the forest compartments; S2 constructs a dynamic forest health assessment network, including an embedding layer, a numerical processing layer, a modal alignment and splicing layer, a Transformer encoder, a text word embedding layer, a projection layer, and a decoder; The embedding layer is used to encode the three types of images separately and introduce temporal dependencies to generate corresponding image feature vectors, which are then concatenated into a concatenated image vector E. img (t,s); The numerical processing layer is used to obtain several numerical features from n(t,s) to form a numerical vector f. num (t,s), and based on f num (t,s) is standardized and MLP mapped to generate a time-by-time standard numerical vector g. num (t,s) and window-level vector ; The modality alignment and splicing layer is used to generate splicing features X. enc , In the formula, For all E within [t0,t1] img (t,s) is a sequence formed by time. For all g in [t0,t1] num (t,s) is a time sequence, X enc The sequence length is L; The Transformer encoder is used to convert X... enc Compressed into a multimodal digest token sequence Z, ∈ K is the sequence length of Z, d dec Let K be the dimension of the decoder's hidden layer, and K ≪ L; The text word embedding layer takes state text as input and outputs the corresponding embedding vector sequence E. state , M1 is E state The sequence length; The projection layer is used for... Generate multimodal summaries Z dec W Z It is a linear mapping matrix; The decoder is used to convert E state As input, Zdec As a condition, generate predictive decision text. ; S3, using the forest spatiotemporal dataset D with Y * To obtain the desired output, a dynamic forest health assessment network is trained to obtain a dynamic forest health assessment model. S4: Obtain all samples of a forest compartment under test in [t0,t1], and obtain the prediction decision text through the dynamic forest health assessment model.

[0005] Preferably, the numerical data includes meteorological data, ground sensor data, and forest survey data, wherein the meteorological data includes temperature T, humidity RH, and wind speed v. wind The ground sensor data includes forest fuel moisture content (FM) and leaf area index (LAI), while the forest survey data includes tree height (h), diameter at breast height (DBH), crown closure (Canopy(t)), and insect infestation probability (P). inf Insect infestation intensity I, fire line advance rate ; The status text is a textual description of Lin Bans' health status; The decision text Y * The textual description of the management decisions made in forest section S is generated based on historical emergency logs or preset rules. M is Y * Total number of real words in Chinese For Y * The m-th real word in the text, 1≤m≤M.

[0006] Preferably, in the S2 embedding layer, I sat (t,s), I ortho (t,s), I lidar The image feature vectors corresponding to (t,s) are respectively e sat (t,s),e ortho (t,s),e lidar (t,s), the concatenated image vector E of the three elements img (t,s)={e sat (t,s),e ortho (t,s),e lidar (t,s)}, the embedding layer generates image feature vectors including steps Sa1~Sa5; Sa1, acquire an image I img (t,s), I img (t,s)∈{I sat (t,s),I ortho (t,s),I lidar (t,s)}, H, W, and C are respectively I imgThe height, width, and number of channels of (t,s); Sa2, I img Input ViT (t,s) is divided into N by the embedding layer. p Each image block is flattened into a P×P×C block vector, where P is the block size. Let the u-th block vector be p(t,s,u), where 1≤u≤N p ; Sa3 linearly maps p(t,s,u) to a preset dimension d to obtain the mapping vector z(t,s,u). Then, a category Token is added to obtain the Token vector corresponding to each image patch. If C=3, use the ViT embedding layer for direct linear mapping; otherwise, use the following mapping formula. , In the formula, W p Let b be a linear mapping matrix used to map p(t,s,u) to a preset embedding dimension d. p ,π 2D (u) represents the bias term of p(t,s,u) and the two-dimensional position code, respectively. , ; Sa4 encodes z(t,s,u) using ViT to generate the corresponding encoded representation, then uses ViT's Transformer encoder to generate the corresponding encoded features h(t,s,u), and finally generates the pooled sequence v. img (t,s), , where pool(∙) is the mean pooling operation; Sa5, add temporal dependencies according to the following formula to generate the image feature vector e. img (t,s); , In the formula, For time embedding, Trm tmp For the encoding operations of the Transformer encoder, W v b v These are linear mapping and bias, respectively.

[0007] As a preferred option .

[0008] As a preferred option, in S2, based on f num (t,s) is standardized and MLP mapped to generate a time-by-time standard numerical vector g. num (t,s) and window-level vector Specifically, it includes Sb1 to Sb3; Sb1, perform IQR and MLP projection according to the following formula to generate g num (t,s); , , In the formula, Med and IOR represent all f num The median and interquartile range of (t,s), MLP(∙) is the multilayer perceptron operation, and d is the output dimension of the MLP; Sb2, count all f num Mean of (t,s) Maximum value Minimum value Final value Generate vectors , ; Sb3, according to the following formula, performs IQR and MLP projection to generate... ; , , In the formula, , They are all The median and interquartile range.

[0009] Preferably, in S2, the Transformer encoder generates a multimodal digest token sequence Z according to the following formula; , , In the formula, A is the attention score matrix, softmax(∙) is the softmax function, and Q is a preset K-row, d-column query matrix. For the transpose operation, W Q W k W V W O These are the learnable Q-projection matrix, K-projection matrix, V-projection matrix, and output projection matrix, d dec is the dimension of the decoder's hidden layer.

[0010] Preferably, the decoder includes L dec The decoding layers are connected sequentially, with the first layer's input being E. state The output of the l-th layer is Y (l) The final layer output is the predicted decision text. The output Y of the (l+1)th decoding layer is obtained according to the following formula. (l +1) ; , , , In the formula, SelfAttn(∙) is the self-attention operation, FFN(∙) is the feedforward network operation, CrossAttn(∙,∙) is the cross-attention operation, and γ l Y is the gating parameter of the l-th layer, σ(∙) is the sigmoid function, and ⊙ is the element-wise multiplication; (la) Y (lb) For Y (l+1) The first intermediate value and the second intermediate value.

[0011] Preferably, in S3, the training specifically includes: S31, Obtain the prediction decision text , , for The m-th predicted word; S32, Constructing the text cross-entropy loss L for a dynamic forest health assessment network NLL ; , , In the formula, For conditional probability, Indicated based on Z dec and E state Predicted The probability, For Y * middle All previous real words; S33, Calculate L NLL And minimize L NLL Adjust the parameters of the dynamic forest health assessment network.

[0012] Compared with the prior art, the advantages of the present invention are as follows: (1) The embedding layer expands the types of images that can be processed in feature extraction and introduces temporal dependencies. First, unlike the embedding layer built into ViT, the embedding layer of this invention is suitable for various numbers of channels. In step Sa3, when the number of channels C=3, the embedding layer built into ViT is directly linearly mapped according to existing technology. When C≠3, the formula of this invention is used for mapping, so that this invention is no longer limited to 3-channel images of RGB images, but is applicable to satellite remote sensing images (the number of channels is usually 8~13) and LiDAR canopy height models (the number of channels is usually 1) mentioned in this invention. Based on this improvement, this invention can integrate image, ground / meteorological, inventory and mechanism-derived features into the same representation space and train on a unified dimension, which can simultaneously characterize structure (LiDAR / orthophoto), function (meteorology / water content / LAI) and risk (FWI, pest PDE), significantly improving the completeness and robustness of health assessment. Second, this invention adds temporal dependencies in Sa5 to generate frame-level image feature vector e. img (t,s), which itself is a time-series sequence that changes over time, ultimately yielding the spliced ​​feature X. enc Its temporal characteristics not only improve the quantitative effect of forest structure and growth status, but also ensure the stability and temporal consistency of the results, providing stronger data support for forest health assessment.

[0013] (2) In the Transformer encoder, X of length L is converted into a variable. enc Compressed into a multimodal summary Z of length K dec K≪L, Z dec It can retain key information across modalities and time periods, while significantly reducing computational complexity and memory overhead. Both increase with Z. dec The length K instead of X enc The length L is scaled to facilitate high-frequency scrolling evaluation.

[0014] (3) A novel decoder was constructed, and a Transformer-based multimodal cross-attention mechanism was introduced as a key step in the evaluation process, which will improve the efficiency of E. state As input, Z dec As a condition, the decoder is guided to generate predictive decision text. The decoder consists of multiple decoding layers, each selectively utilizing Z through cross-attention. dec generate And integrate into This approach not only avoids interference from noise or irrelevant information but also improves the temporal trend modeling of dynamic forest health status. Through this method, long-term predictions of overall forest health levels can be achieved, and key factors such as localized pest spread and fire risk indices can be precisely monitored, significantly improving the accuracy of dynamic forest health assessments. Particularly in the context of the integrated needs of disaster prevention and mitigation and forest management, this technology significantly enhances its practical application value.

[0015] In summary, this invention significantly improves the quality and application value of dynamic forest health assessment by introducing a large-scale scientific computing model with multimodal fusion and mechanistic constraints, providing an innovative solution for forest monitoring and management. By addressing the issues of multimodal dynamic modeling and physical consistency, this method offers a new technical approach for forest health status assessment, with particularly broad application prospects in dynamic forest monitoring and risk early warning. Attached Figure Description

[0016] Figure 1 This is a network structure diagram for the dynamic forest health assessment of this invention. Detailed Implementation

[0017] The present invention will be further described below with reference to the embodiments and accompanying drawings.

[0018] Example 1: See Figure 1 A dynamic forest health assessment method based on a large scientific computing model includes the following steps; S1, construct the forest spatiotemporal dataset D, including S11~S13; S11, the forest is divided into multiple forest compartments, and an observation window [t0,t1] is preset, where t0 and t1 are the start and end times of the observation window, respectively. S12, for one forest compartment s, collect image data I(t,s), numerical data n(t,s), state text ST, and decision text Y at time t. * The image dataset includes three types of images: satellite remote sensing images I sat (t,s), orthophoto I ortho (t,s) and LiDAR canopy height model I lidar (t,s); S13, generate sample D of the forest compartment at time t (s). t =(I(t,s),n(t,s),ST,Y * ), where Y * D t The weak labels are used to construct the forest spatiotemporal dataset D from all samples at time [t0,t1] of the forest compartments; S2 constructs a dynamic forest health assessment network, including an embedding layer, a numerical processing layer, a modal alignment and splicing layer, a Transformer encoder, a text word embedding layer, a projection layer, and a decoder; The embedding layer is used to encode the three types of images separately and introduce temporal dependencies to generate corresponding image feature vectors, which are then concatenated into a concatenated image vector E. img (t,s); The numerical processing layer is used to obtain several numerical features from n(t,s) to form a numerical vector f. num (t,s), and based on f num (t,s) is standardized and MLP mapped to generate a time-by-time standard numerical vector g. num (t,s) and window-level vector ; The modality alignment and splicing layer is used to generate splicing features X. enc , In the formula, For all E within [t0,t1] img (t,s) is a sequence formed by time. For all g in [t0,t1] num (t,s) is a time sequence, X enc The sequence length is L; The Transformer encoder is used to convert X... enc Compressed into a multimodal digest token sequence Z, ∈ K is the sequence length of Z, d dec Let K be the dimension of the decoder's hidden layer, and K ≪ L; The text word embedding layer takes state text as input and outputs the corresponding embedding vector sequence E. state , M1 is E state The sequence length; The projection layer is used for... Generate multimodal summaries Z dec W Z It is a linear mapping matrix; The decoder is used to convert E state As input, Z dec As a condition, generate predictive decision text. ; S3, using the forest spatiotemporal dataset D with Y * To obtain the desired output, a dynamic forest health assessment network is trained to obtain a dynamic forest health assessment model. S4: Obtain all samples of a forest compartment under test in [t0,t1], and obtain the prediction decision text through the dynamic forest health assessment model.

[0019] In this embodiment: The LiDAR Canopy Height Model (CHM) is one of the most commonly used vegetation structure products in forestry remote sensing. Essentially, a LiDAR canopy height model is a two-dimensional raster image where the value of each pixel equals the vertical distance from the top of the vegetation (or any echo) to the ground. It originates from a three-dimensional LiDAR point cloud but has been "reduced" to two dimensions.

[0020] ViT (Vision Transformer) is an innovative model that applies the Transformer architecture to computer vision tasks. It segments images into patches, forms the input through linear mapping and flattening, and extracts features using a Transformer Encoder. It can be summarized into three main parts: an embedding layer, a Transformer Encoder, and an MLP (Multilayer Perceptron).

[0021] Example 2: See Figure 1 More specifically, based on Example 1, the numerical data includes meteorological data, ground sensor data, and forest survey data, wherein the meteorological data includes temperature T, humidity RH, and wind speed v. wind The ground sensor data includes forest fuel moisture content (FM) and leaf area index (LAI), while the forest survey data includes tree height (h), diameter at breast height (DBH), crown closure (Canopy(t)), and insect infestation probability (P). inf Insect infestation intensity I, fire line advance rate ; The status text is a textual description of Lin Bans' health status; The decision text Y * The textual description of the management decisions made in forest section S is generated based on historical emergency logs or preset rules. M is Y * Total number of real words in Chinese For Y * The m-th real word in the text, 1≤m≤M.

[0022] In the S2 embedding layer, I sat (t,s), I ortho (t,s), I lidar The image feature vectors corresponding to (t,s) are respectively e sat (t,s),e ortho (t,s),e lidar (t,s), the concatenated image vector E of the three elements img (t,s)={e sat(t,s),e ortho (t,s),e lidar (t,s)}, the embedding layer generates image feature vectors including steps Sa1~Sa5; Sa1, acquire an image I img (t,s), I img (t,s)∈{I sat (t,s),I ortho (t,s),I lidar (t,s)}, H, W, and C are respectively I img The height, width, and number of channels of (t,s); Sa2, I img Input ViT (t,s) is divided into N by the embedding layer. p Each image block is flattened into a P×P×C block vector, where P is the block size. Let the u-th block vector be p(t,s,u), where 1≤u≤N p ; Sa3 linearly maps p(t,s,u) to a preset dimension d to obtain the mapping vector z(t,s,u). Then, a category Token is added to obtain the Token vector corresponding to each image patch. If C=3, use the ViT embedding layer for direct linear mapping; otherwise, use the following mapping formula. , In the formula, W p Let b be a linear mapping matrix used to map p(t,s,u) to a preset embedding dimension d. p ,π 2D (u) represents the bias term of p(t,s,u) and the two-dimensional position code, respectively. , ; Sa4 encodes z(t,s,u) using ViT to generate the corresponding encoded representation, then uses ViT's Transformer encoder to generate the corresponding encoded features h(t,s,u), and finally generates the pooled sequence v. img (t,s), , where pool(∙) is the mean pooling operation; Sa5, add temporal dependencies according to the following formula to generate the image feature vector e. img (t,s); , In the formula, For time embedding, Trm tmp For the encoding operations of the Transformer encoder, W vb v These are linear mapping and bias, respectively.

[0023] .

[0024] In S2, based on f num (t,s) is standardized and MLP mapped to generate a time-by-time standard numerical vector g. num (t,s) and window-level vector Specifically, it includes Sb1 to Sb3; Sb1, perform IQR and MLP projection according to the following formula to generate g num (t,s); , , In the formula, Med and IOR represent all f num The median and interquartile range of (t,s), MLP(∙) is the multilayer perceptron operation, and d is the output dimension of the MLP; Sb2, count all f num Mean of (t,s) Maximum value Minimum value Final value Generate vectors , ; Sb3, according to the following formula, performs IQR and MLP projection to generate... ; , , In the formula, , They are all The median and interquartile range.

[0025] In S2, the Transformer encoder generates a multimodal digest token sequence Z according to the following formula; , , In the formula, A is the attention score matrix, softmax(∙) is the softmax function, and Q is a preset K-row, d-column query matrix. For the transpose operation, W Q W k W V W O These are the learnable Q-projection matrix, K-projection matrix, V-projection matrix, and output projection matrix, d dec is the dimension of the decoder's hidden layer.

[0026] The decoder includes L dec The decoding layers are connected sequentially, with the first layer's input being E. state The output of the l-th layer is Y (l) The final layer output is the predicted decision text. The output Y of the (l+1)th decoding layer is obtained according to the following formula. (l+1) ; , , , In the formula, SelfAttn(∙) is the self-attention operation, FFN(∙) is the feedforward network operation, CrossAttn(∙,∙) is the cross-attention operation, and γ l Y is the gating parameter of the l-th layer, σ(∙) is the sigmoid function, and ⊙ is the element-wise multiplication; (la) Y (lb) For Y (l+1) The first intermediate value and the second intermediate value.

[0027] In S3, the training specifically refers to: S31, Obtain the prediction decision text , , for The m-th predicted word; S32, Constructing the text cross-entropy loss L for a dynamic forest health assessment network NLL ; , , In the formula, For conditional probability, Indicated based on Z dec and E state Predicted The probability, For Y * middle All previous real words; S33, Calculate L NLL And minimize L NLL Adjust the parameters of the dynamic forest health assessment network.

[0028] Example 3: See Figure 1 The numerical data are not limited to those described in Example 1 or Example 2, but may also include other available data related to forest health assessment. Some of the data were obtained in the following manner: (1) Tree height h: , , In the formula, h is the tree height in meters, Hchm is the LiDAR canopy height model in meters, GDD is the accumulated temperature in °C·day, and T is the total temperature. base Base temperature, in °C. The value represents the drone calibration, in meters. T(t,s) is the temperature of the forest compartment at time t, s, and Δt is the time length [t0,t1], where Δt = t1 - t0.

[0029] (2) Diameter at breast height (DBH): , In the formula, the diameter at breast height (DBH) is in centimeters, and C... ortho The orthogonal crown perimeter is expressed in meters; 'a' represents the tree species parameter, and 'b' represents the tree age parameter. is the vegetation index, NIR is the near-infrared reflectance, and Red is the red light reflectance.

[0030] (3) Coronary closure Canopy(t): , In this formula, N represents the total number of forest compartments, and k is the k-th forest compartment among all forest compartments. For the k-th forest compartment, Shadow k (t), A k Let k be the shaded area and forest area at time t, respectively. * is the empirical coefficient, LAI(t) is the leaf area index at time t, and e is the base of the natural logarithm.

[0031] (4) Probability of pest infection P inf : , , In the formula, σ(∙) is the Sigmoid function, NDWI is the normalized differential water index, NIR is the near-infrared reflectance, SWIR represents the short-wave infrared reflectance, and w1 and w2 are the weights of the multiplication terms, respectively. This is an indicator function, where RH represents humidity, with a value between 0 and 1.

[0032] (5) Pest Intensity I: , In the formula, I represents the pest intensity, I + Let represent the pest intensity at the next moment, Lapl(I) be the Laplace operator, D(T) be the temperature-dependent diffusion coefficient, β be the propagation coefficient, ρ be the tree density per unit (trees / ha), μ be the control efficiency, and Spray be the spraying intensity.

[0033] (6) Fireline advance rate : , In the formula, v wind Wind speed (m / s), ∇F fire gradient, FM forest fuel moisture content (0–1), and slope. The empirical coefficient for FM. Slope is the empirical coefficient.

[0034] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A dynamic forest health assessment method based on a large-scale scientific computing model, characterized in that, Includes the following steps; S1, construct the forest spatiotemporal dataset D, including S11~S13; S11, the forest is divided into multiple forest compartments, and an observation window [t0,t1] is preset, where t0 and t1 are the start and end times of the observation window, respectively. S12, for one forest compartment s, collect image data I(t,s), numerical data n(t,s), state text ST, and decision text Y at time t. * The image dataset includes three types of images: satellite remote sensing images I sat (t,s), orthophoto I ortho (t,s) and LiDAR canopy height model I lidar (t,s); S13, generate sample D of the forest compartment at time t (s). t =(I(t,s),n(t,s),ST,Y * ), where Y * D t The weak labels are used to construct the forest spatiotemporal dataset D from all samples at time [t0,t1] of the forest compartments; S2 constructs a dynamic forest health assessment network, including an embedding layer, a numerical processing layer, a modal alignment and splicing layer, a Transformer encoder, a text word embedding layer, a projection layer, and a decoder; The embedding layer is used to encode the three types of images separately and introduce temporal dependencies to generate corresponding image feature vectors, which are then concatenated into a concatenated image vector E. img (t,s); The numerical processing layer is used to obtain several numerical features from n(t,s) to form a numerical vector f. num (t,s), and based on f num (t,s) is standardized and MLP mapped to generate a time-by-time standard numerical vector g. num (t,s) and window-level vector ; The modality alignment and splicing layer is used to generate splicing feature X. enc , In the formula, For all E within [t0,t1] img (t,s) is a sequence formed by time. For all g in [t0,t1] num (t,s) is a time sequence, X enc The sequence length is L; The Transformer encoder is used to convert X... enc Compressed into a multimodal digest token sequence Z, Z∈ K is the sequence length of Z, d dec For the decoder hidden layer dimension, ; The text word embedding layer takes state text as input and outputs the corresponding embedding vector sequence E. state , M1 is E state The sequence length; The projection layer is used for... Generate multimodal summaries Z dec W Z It is a linear mapping matrix; The decoder is used to convert E state As input, Z dec As a condition, generate predictive decision text. ; S3, using the forest spatiotemporal dataset D with Y * To obtain the desired output, a dynamic forest health assessment network is trained to obtain a dynamic forest health assessment model. S4: Obtain all samples of a forest compartment under test in [t0,t1], and obtain the prediction decision text through the dynamic forest health assessment model.

2. The dynamic forest health assessment method based on a large-scale scientific computing model according to claim 1, characterized in that, The numerical data includes meteorological data, ground sensor data, and forest survey data, wherein the meteorological data includes temperature (T), humidity (RH), and wind speed (v). wind The ground sensor data includes forest fuel moisture content (FM) and leaf area index (LAI), while the forest survey data includes tree height (h), diameter at breast height (DBH), crown closure (Canopy(t)), and insect infestation probability (P). inf Insect infestation intensity I, fire line advance rate ; The status text is a textual description of the health status of Linban s; The decision text Y * The textual description of the management decisions made in forest section S is generated based on historical emergency logs or preset rules. M is Y * Total number of real words in Chinese For Y * The m-th real word in the text, 1≤m≤M.

3. The dynamic forest health assessment method based on a large-scale scientific computing model according to claim 1, characterized in that, In the S2 embedding layer, I sat (t,s), I ortho (t,s), I lidar The image feature vectors corresponding to (t,s) are respectively e sat (t,s), e ortho (t,s), e lidar (t,s), the concatenated image vector E of the three elements img (t,s)={e sat (t,s), e ortho (t,s), e lidar (t,s)}, the embedding layer generates image feature vectors including steps Sa1~Sa5; Sa1, acquire an image I img (t,s), I img (t,s)∈{I sat (t,s),I ortho (t,s),I lidar (t,s)}, H, W, and C are respectively I img The height, width, and number of channels of (t,s); Sa2, I img Input ViT (t,s) is divided into N by the embedding layer. p Each image block is flattened into a P×P×C block vector, where P is the block size. Let the u-th block vector be p(t,s,u), where 1≤u≤N p ; Sa3 linearly maps p(t,s,u) to a preset dimension d to obtain the mapping vector z(t,s,u), and then adds a category token to obtain the token vector corresponding to each image patch; If C=3, use the ViT embedding layer for direct linear mapping; otherwise, use the following mapping formula. ， In the formula, W p Let b be a linear mapping matrix used to map p(t,s,u) to a preset embedding dimension d. p ,π 2D (u) represents the bias term of p(t,s,u) and the two-dimensional position code, respectively. , ; Sa4 encodes z(t,s,u) using ViT to generate the corresponding encoded representation, then uses ViT's Transformer encoder to generate the corresponding encoded features h(t,s,u), and finally generates the pooled sequence v. img (t,s), , where Pool(∙) is the mean pooling operation; Sa5, add temporal dependencies according to the following formula to generate the image feature vector e. img (t,s); ， In the formula, For time embedding, Trm tmp For the encoding operations of the Transformer encoder, W v b v These are linear mapping and bias, respectively.

4. The dynamic forest health assessment method based on a large-scale scientific computing model according to claim 2, characterized in that, 。 5. The dynamic forest health assessment method based on a large-scale scientific computing model according to claim 1, characterized in that, In S2, based on f num (t,s) is standardized and MLP mapped to generate a time-by-time standard numerical vector g. num (t,s) and window-level vector Specifically, it includes Sb1 to Sb3; Sb1, perform IQR and MLP projection according to the following formula to generate g num (t,s); ， ， In the formula, Med and IQR represent all f num The median and interquartile range of (t,s), MLP(∙) is the multilayer perceptron operation, and d is the output dimension of the MLP; Sb2, count all f num Mean of (t,s) Maximum value Minimum value Final value Generate statistical vectors ; Sb3, according to the following formula, performs IQR and MLP projection to generate... ; ， ， In the formula, , They are all The median and interquartile range.

6. The dynamic forest health assessment method based on a large-scale scientific computing model according to claim 1, characterized in that, In S2, the Transformer encoder generates a multimodal digest token sequence Z according to the following formula; ， ， In the formula, A is the attention score matrix, softmax(∙) is the softmax function, and Q is a preset K-row, d-column query matrix. For the transpose operation, W Q W k W V W O These are the learnable Q-projection matrix, K-projection matrix, V-projection matrix, and output projection matrix, d dec is the dimension of the decoder's hidden layer.

7. The dynamic forest health assessment method based on a large-scale scientific computing model according to claim 1, characterized in that, The decoder includes L dec The decoding layers are connected sequentially, with the first layer's input being E. state The output of the l-th layer is Y (l) The final layer output is the prediction decision text. The output Y of the (l+1)th decoding layer is obtained according to the following formula. (l+1) ; ， ， ， In the formula, SelfAttn(∙) is the self-attention operation, FFN(∙) is the feedforward network operation, CrossAttn(∙,∙) is the cross-attention operation, and γ l Y is the gating parameter of the l-th layer, σ(∙) is the sigmoid function, and ⊙ is the element-wise multiplication; (la) Y (lb) For Y (l+1) The first intermediate value and the second intermediate value.

8. The dynamic forest health assessment method based on a large-scale scientific computing model according to claim 2, characterized in that, In S3, the training specifically refers to: S31, Obtain the prediction decision text , , for The m-th predicted word; S32, Constructing the text cross-entropy loss L for a dynamic forest health assessment network NLL ; ， ， In the formula, For conditional probability, Indicated based on Z dec and E state Predicted The probability, For Y * middle All previous real words; S33, Calculate L NLL And minimize L NLL Adjust the parameters of the dynamic forest health assessment network.

Images