Forest ecological disaster collaborative prediction and evaluation method based on multi-task learning

Abstract

The invention discloses a forest ecological disaster collaborative prediction and evaluation method based on multi-task learning, and the method comprises the following steps: S1, collecting multi-modal data of forest ecological disasters, and unifying data features through a time alignment technology and a space alignment technology; s2, a multi-task learning framework is constructed, a dynamic sharing layer extracts sharing features, and a task special module generates input features of fire prediction, pest and disease assessment and climate change modeling; s3, optimizing the shared feature weight by adopting a parrot optimization algorithm; s4, constructing a task dependency graph, distributing inter-task feature weights through a dynamic attention mechanism, and generating a prediction result of each task; s5, a prediction result is verified and stored in a disaster memory bank, and framework parameters are optimized in combination with an incremental learning mechanism; s6, screening features by using a multi-head self-attention mechanism, and optimizing the prediction result of each task; and S7, outputting a prediction and evaluation report. According to the method, multi-modal data processing, multi-task learning and dynamic optimization technologies are combined to realize collaborative prediction of forest disasters.

Description

Technical Field

[0001] The present invention relates to the fields of forest ecological disaster management and artificial intelligence technology, and in particular to a forest ecological disaster collaborative prediction and assessment method based on multi-task learning. Background Art

[0002] Forest ecosystems play a vital role in global ecological balance and economic development. However, with the intensification of climate change and human activities, forest ecological disasters (such as fires, pests and diseases, and the impact of climate change on forests) occur frequently, posing huge challenges to the protection and management of forest resources. At present, the prediction and assessment of forest ecological disasters mainly rely on single disaster monitoring technology and traditional data analysis methods. These methods usually use remote sensing images, meteorological data or ground monitoring data as a single data source, and use empirical models or statistical methods to assess disasters. However, due to the single data source and limited processing capacity, existing technologies cannot fully and accurately capture the complex interactions between multiple disasters, resulting in insufficient prediction accuracy and timeliness.

[0003] Among existing fire prediction technologies, remote sensing images are widely used for monitoring fire points and assessing fire risks. However, due to the limitations of the spatial resolution and temporal update frequency of the images, this method is difficult to meet the needs of real-time monitoring. At the same time, the existing methods of using meteorological data mostly remain at the level of simple trend analysis, ignoring the dynamic impact of key variables such as wind speed and humidity on the development of fires, resulting in low temporal and spatial accuracy of fire predictions. The field of pest and disease assessment usually relies on ground monitoring and manual surveys. Although this method has high local accuracy, it is unable to predict the spread of pests and diseases over a wide area. In addition, traditional statistical models usually assume that data are independent and identically distributed, ignoring the influence of geographical environment and ecological characteristics on the spread of pests and diseases, thereby limiting the comprehensiveness and reliability of the prediction.

[0004] For the assessment of the impact of climate change on forest health, current technologies are mainly based on time series analysis of meteorological data. However, these methods lack the ability to integrate multimodal data (such as remote sensing images and ground monitoring data), and fail to fully consider the multifaceted impacts of climate change on forest vegetation cover, soil moisture, and ecosystem dynamics. At the same time, most existing models use static analysis methods and cannot respond to rapidly changing climate events, such as extreme droughts or heavy rains, in a timely manner, which makes forest health assessments lack timeliness and dynamic adaptability.

[0005] In addition, research on multi-disaster collaborative prediction technology is relatively scarce. There are often complex interactions between fire, pests and diseases, and climate change. For example, climate change may cause forest drying and increase fire risks; vegetation destruction after fire provides favorable conditions for the spread of pests and diseases. However, existing methods usually model these disasters as independent problems and lack the ability to deeply explore and model their collaborative relationships, resulting in poor results in multi-disaster collaborative prediction. At the same time, due to the significant differences in the characteristics and data sources of different disasters, existing methods face huge challenges in the unified processing of multimodal data and the optimization of shared features between tasks, further limiting the accuracy and efficiency of multi-disaster collaborative prediction.

[0006] In the existing technology, the introduction of machine learning and deep learning technology has provided new ideas for the prediction and assessment of forest ecological disasters. For example, convolutional neural networks (CNN) and recurrent neural networks (RNN) have been used to extract features from remote sensing images and time series data and applied to the prediction task of a single disaster. However, these methods usually adopt a static feature sharing mechanism and fail to dynamically adjust the feature allocation weights between different tasks. In addition, when dealing with multi-task learning problems, traditional deep learning methods are prone to negative transfer between tasks, that is, high-priority tasks will interfere with low-priority tasks and reduce the overall prediction performance.

[0007] Therefore, how to provide a collaborative prediction and assessment method for forest ecological disasters based on multi-task learning is an urgent problem that technicians in this field need to solve. Summary of the invention

[0008] One purpose of the present invention is to propose a collaborative prediction and assessment method for forest ecological disasters based on multi-task learning. The present invention combines multimodal data processing technology, multi-task learning framework and dynamic optimization algorithm to achieve collaborative prediction and assessment for tasks such as forest fires, pests and diseases, and climate change modeling. The complex relationship between tasks is captured by dynamic sharing layers and task dependency graphs, and the feature allocation is optimized using the Parrot optimization algorithm and multi-head self-attention mechanism, which significantly improves the accuracy and efficiency of the prediction. In addition, the dynamic adaptability of the model is enhanced through the disaster memory library and incremental learning mechanism, which further improves the system's response speed to new disaster patterns and the stability of long-term predictions, and has the advantages of strong comprehensiveness, high accuracy and excellent dynamic adaptability.

[0009] A forest ecological disaster collaborative prediction and assessment method based on multi-task learning according to an embodiment of the present invention includes the following steps:

[0010] S1. Collect multimodal data of forest ecological disasters, including remote sensing images, meteorological data and ground monitoring data, and unify the multimodal data features through time alignment and space alignment technology;

[0011] S2. Construct a multi-task learning framework, including a dynamic shared layer and task-specific modules, where the dynamic shared layer is used to extract shared features from multimodal data features, and the task-specific modules generate fire prediction input features, pest and disease assessment input features, and climate change modeling input features based on the shared features;

[0012] S3. Use the Parrot optimization algorithm to optimize the shared feature weights of the dynamic shared layer, dynamically adjust the allocation ratio of shared features through imitation learning and independent exploration mechanisms, and generate the optimal shared feature allocation strategy;

[0013] S4. Applying the optimal shared feature allocation strategy to a task dependency graph, wherein the task dependency graph is constructed based on the synergistic relationship between fire prediction, pest and disease assessment, and climate change modeling, and allocating shared feature weights between tasks through a dynamic attention mechanism to generate fire prediction results, pest and disease assessment results, and climate change modeling results;

[0014] S5. Verify the fire prediction results, pest and disease assessment results, and climate change modeling results with historical data, store the verified results in the disaster memory bank, and update the parameters of the dynamic shared layer and task-specific modules in combination with the incremental learning mechanism to optimize the multi-task learning framework;

[0015] S6. Based on the optimized multi-task learning framework, the multi-head self-attention mechanism is used to screen shared features and task-specific modules task by task to optimize the results of fire prediction, pest assessment and climate change modeling;

[0016] S7. Output fire risk forecasts, pest and disease control recommendations and ecological health assessment reports to provide support for forest disaster management and emergency decision-making.

[0017] Optionally, the S2 specifically includes:

[0018] S21, initializing a dynamic sharing layer, extracting multimodal data features through a convolutional neural network, wherein the multimodal data features include spatial distribution features of remote sensing images, time series features of meteorological data, and point features of ground monitoring data, and generating preliminary features;

[0019] S22, normalizing the preliminary features to obtain a standardized feature vector;

[0020] S23, integrating the standardized feature vectors by feature fusion technology to generate shared features, where the shared features are used as input for the task-specific module;

[0021] S24, constructing task-specific modules, wherein the task-specific modules include a fire prediction module, a pest assessment module, and a climate change modeling module, each module receiving the shared features as input;

[0022] S25. In the fire prediction module, based on the time series feature analysis, the time trend characteristics and spatial distribution characteristics of the fire occurrence are extracted to generate the fire prediction input features:

[0023] h t =σ(W x x t +W h h t-1 +b h );

[0024] o t =tanh(W o h t +b o );

[0025]

[0026] Among them, h t represents the hidden state of the gated recurrent unit network, σ represents the activation function, and W x , W h and W o represents the weight matrix, b h and b o represents the bias term, T represents the length of the time series, o t represents the output vector of the hidden state at time step t after further processing, tanh represents the hyperbolic tangent function, and F fire represents the fire prediction input features;

[0027] S26. In the pest assessment module, the graph neural network is used to process the shared features to generate pest assessment input features:

[0028]

[0029] in, represents the feature of node i in the l+1th layer of the graph neural network, N(i) represents the neighbor set of node i, represents the feature of node j in the lth layer of the graph neural network, W (l) represents the weight matrix of the lth layer, A ij represents the element of the adjacency matrix, deg(i) represents the degree of node i, deg(j) represents the degree of node j, Aggregate represents the aggregation operation of node features, L represents the total number of layers of the graph neural network, and F pest represents the input features of pest and disease assessment, and n represents the total number of nodes;

[0030] S27. In the climate change modeling module, a multi-head attention mechanism is used to process shared features to generate climate change modeling input features:

[0031] F climate =Concat(head 1 ,head 2 ,…,head m )·W out ;

[0032] Among them, F climate represents the input features of climate change modeling, W out Represents the weight matrix of the output layer, m represents the total number of attention heads, head represents the attention head, and Concat represents the concatenation operation.

[0033] Optionally, the S3 specifically includes:

[0034] S31. Initialize shared feature weight vector The weight vector length is the same as the shared feature vector length output by the dynamic shared layer:

[0035]

[0036] in, represents the weight of the u-th dimension in the initial state, u represents the dimension index of the shared feature, A represents the total number of dimensions of the shared feature, exp represents the exponential function, L u represents the loss value of the u-th dimension in the shared feature in the initial task performance, L v Represents the loss value of the vth dimension in the shared feature in the initial task performance;

[0037] S32. Define a multi-objective optimization function to evaluate the feature allocation strategy of individual parrots. The number of tasks in multi-task learning is known to be B, and the loss value of each task is expressed as R b , where b = 1, 2, ..., B, and two optimization objectives are set. The first optimization objective is to maximize the task coordination:

[0038]

[0039] Among them, S(W (c) ) represents the task coordination degree of the cth parrot individual, and its value range is [0,1]. The larger the value, the better the coordination effect between tasks. tanh represents the hyperbolic tangent function, W (c) represents the shared feature allocation strategy of the cth parrot;

[0040] The second task is to maximize the balance of feature distribution:

[0041]

[0042] Among them, D(W (c)) represents the characteristic distribution balance of the cth parrot individual, represents the weight assigned to the u-th dimension by the c-th parrot, w neam (c) represents the mean value of the weights of all dimensions for the cth parrot; D(W (c) ) is closer to 1, the more balanced the feature distribution is; conversely, the feature distribution is unbalanced;

[0043] S33. Conduct a multi-objective comprehensive evaluation of the shared feature allocation strategy for each parrot individual:

[0044] F(W (c) )=α·S(W (c) )+β·D(W (c) );

[0045] Among them, F(W (c) ) represents the comprehensive evaluation function value of the cth parrot individual, α and β represent the trade-off coefficients;

[0046] S34, execute the imitation learning mechanism, and bring the parrot individuals whose comprehensive evaluation function value is less than the set threshold ξ closer to the parrot individuals whose comprehensive evaluation function value is greater than the set threshold ξ:

[0047]

[0048] in, It represents the weight of the u-th dimension of the r-th parrot individual whose comprehensive evaluation function value is less than the set threshold ξ. represents the u-th dimension weight of the r-th parrot individual whose comprehensive evaluation function value is greater than the set threshold ξ, η represents the imitation learning rate, and ← represents the update operation;

[0049] S35, execute the independent exploration mechanism, randomly select and update from the parrot individuals whose comprehensive evaluation function value is equal to the set threshold ξ:

[0050]

[0051] in, represents the u-th dimension weight of the parrot individual whose comprehensive evaluation function value is equal to the set threshold ξ, δ represents the random disturbance factor, and Ω(-1,1) represents the random number sampled in the interval [-1,1];

[0052] S36. After completing imitation learning and independent exploration, recalculate the comprehensive evaluation function value of each individual parrot and sort the parrot group. If the iteration reaches the preset number of iterations or the multi-objective evaluation function converges, end the iteration and generate the optimal shared feature allocation strategy.

[0053] Optionally, the S4 specifically includes:

[0054] S41. Based on fire prediction, pest assessment and climate change modeling, construct a task dependency graph G = (V, E), where V = {T fire ,T pest ,T climate} represents the task node set, T fire represents the fire prediction task, T pest represents the pest assessment task, T climate represents the climate change modeling task, E = {e ij} represents the edge set between tasks, e ij Represents task T i and Task T j synergistic relationship;

[0055] S42. Calculate the collaborative relationship strength between tasks based on historical task data:

[0056]

[0057] Among them, ρ(T i ,T j ) represents task T i and Task T j The strength of the synergistic relationship between i ,R j ) represents task T i and Task T j The covariance between i ) represents task T i The standard deviation, σ(R j ) represents task T j The standard deviation of

[0058] S43. Normalize the collaborative relationship strength between tasks to obtain the standardized task dependency strength

[0059] S44, applying the optimal shared feature allocation strategy to the task dependency graph, and calculating the shared feature weight of each task node:

[0060]

[0061] in, Represents task T i The weighted shared features of shared represents the shared features of the dynamic shared layer, W * represents the optimal shared feature allocation strategy, |V| represents the total number of task nodes;

[0062] S45. Input the weighted shared features into the task-specific module to generate fire prediction results, pest and disease assessment results, and climate change modeling results.

[0063] Optionally, the S5 specifically includes:

[0064] S51, verify the fire prediction results, pest and disease assessment results, and climate change modeling results with historical data task by task, calculate the difference between the prediction results of each task and the historical data, and generate corresponding error data;

[0065] S52. Evaluate the overall error level of fire prediction, pest assessment and climate change modeling tasks based on the error data verified by each task, and mark the accuracy and error distribution of the prediction results of each task;

[0066] S53, storing the verified prediction results and error data in the disaster memory database, and recording the results of each verification;

[0067] S54. Combined with the historical error data in the disaster memory library, the parameters of the dynamic shared layer and the task-specific module are updated through the incremental learning mechanism, the overall error is used to perform gradient updates on the model parameters, and the shared feature allocation weights and the feature processing strategies of the task-specific modules are adjusted:

[0068]

[0069] Among them, Θ new represents the updated parameters, Θ old represents the parameters before updating, τ represents the control factor, represents the gradient calculation, B represents the number of tasks in multi-task learning, represents the validation error of task b, κ represents the regularization coefficient, represents the parameter of the mth historical optimization record in the disaster memory database, and Y represents the number of historical optimization records in the disaster memory database;

[0070] S55, repeat the verification and update process until the overall error level of the multi-task learning framework reaches a preset convergence threshold.

[0071] Optionally, the S6 specifically includes:

[0072] S61. Based on the optimized multi-task learning framework, the shared features and task-specific modules are screened task by task, and the attention weight matrix of each task is calculated through the multi-head self-attention mechanism:

[0073]

[0074] in, represents the attention weight matrix of task b, softmax represents the normalization function, represents the query matrix of task b, represents the key matrix of task b, d b Represents the key matrix dimension of task b;

[0075] S62. Use the attention weight matrix of the task to perform weighted processing on the value matrix to generate a feature vector after task-by-task screening:

[0076]

[0077] in, represents the screening feature vector of task b, represents the value matrix of task b;

[0078] S63. Perform task-by-task optimization on the selected feature vectors, remove redundant feature information, and input them into task-specific modules to further optimize the results of fire prediction, pest and disease assessment, and climate change modeling.

[0079] The beneficial effects of the present invention are:

[0080] First, the present invention realizes the deep fusion of multimodal data, overcoming the defects of the prior art in single data source and one-sided prediction results. By extracting shared features from remote sensing images, meteorological data and ground monitoring data through a dynamic sharing layer, and assigning tasks to these shared features through task-specific modules, the present invention realizes a comprehensive analysis of fire prediction, pest and disease assessment and climate change modeling. In addition, time alignment and spatial alignment technologies effectively solve the differences in time steps and spatial resolutions between different data sources, making the integration of multimodal data features more accurate and providing a reliable data foundation for collaborative prediction.

[0081] Secondly, the present invention innovatively introduces the Parrot Optimization Algorithm to dynamically adjust the shared feature allocation weights of the dynamic sharing layer and generate the optimal shared feature allocation strategy. Compared with the traditional static feature sharing method, the dynamic sharing mechanism can flexibly allocate computing resources according to the priorities and requirements of different tasks, effectively avoiding the negative transfer phenomenon between tasks. By optimizing the feature allocation ratio, the present invention significantly improves the overall efficiency of the multi-task learning framework and enhances the synergistic effectiveness of fire prediction, pest and disease assessment, and climate change modeling tasks.

[0082] In addition, the present invention captures the collaborative relationship between fire, pests and diseases, and climate change tasks by constructing a task dependency graph and combining it with a dynamic attention mechanism. The task dependency graph analyzes the collaborative strength between tasks based on historical data, and the dynamic attention mechanism reasonably distributes the shared feature weights between tasks, so that highly correlated tasks are fully supported and low-correlation tasks are less disturbed. This innovative design makes up for the shortcomings of task-splitting modeling in the prior art, realizes collaborative prediction between multiple disasters, and makes the prediction results more in line with the complex dynamics of forest ecosystems.

[0083] Furthermore, the present invention uses the disaster memory library and incremental learning mechanism to significantly improve the dynamic adaptability of the model. The disaster memory library stores historical optimization results and verification errors, and quickly adjusts the parameters of the dynamic shared layer and task-specific modules through the incremental learning mechanism, avoiding the high computational cost of global retraining. This design not only enhances the model's ability to respond to new disaster patterns, but also improves the stability and accuracy of long-term dynamic predictions.

[0084] Finally, the present invention uses a multi-head self-attention mechanism to screen shared features and task-specific modules task by task, further optimizing the results of fire prediction, pest and disease assessment, and climate change modeling. By removing redundant features and strengthening important features, the multi-head self-attention mechanism makes the prediction results of each task more accurate and efficient. Ultimately, the system can output fire risk warning information, pest and disease prevention and control recommendations, and ecological health assessment reports, providing a scientific basis for forest disaster management and emergency decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0086] Figure 1 This is an overall flow chart of a collaborative prediction and assessment method for forest ecological disasters based on multi-task learning proposed by the present invention;

[0087] Figure 2 This is a structural schematic diagram of the dynamic sharing layer and task-specific modules of a collaborative prediction and assessment method for forest ecological disasters based on multi-task learning proposed by the present invention. DETAILED DESCRIPTION

[0088] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0089] refer to Figure 1 and Figure 2 , a collaborative prediction and assessment method for forest ecological disasters based on multi-task learning, including the following steps:

[0090] S1. Collect multimodal data of forest ecological disasters, including remote sensing images, meteorological data and ground monitoring data, and unify the multimodal data features through time alignment and space alignment technology;

[0091] S2. Construct a multi-task learning framework, including a dynamic shared layer and task-specific modules, where the dynamic shared layer is used to extract shared features from multimodal data features, and the task-specific modules generate fire prediction input features, pest and disease assessment input features, and climate change modeling input features based on the shared features;

[0092] S3. Use the Parrot optimization algorithm to optimize the shared feature weights of the dynamic shared layer, dynamically adjust the allocation ratio of shared features through imitation learning and independent exploration mechanisms, and generate the optimal shared feature allocation strategy;

[0093] S4. Applying the optimal shared feature allocation strategy to a task dependency graph, wherein the task dependency graph is constructed based on the synergistic relationship between fire prediction, pest and disease assessment, and climate change modeling, and allocating shared feature weights between tasks through a dynamic attention mechanism to generate fire prediction results, pest and disease assessment results, and climate change modeling results;

[0094] S5. Verify the fire prediction results, pest and disease assessment results, and climate change modeling results with historical data, store the verified results in the disaster memory bank, and update the parameters of the dynamic shared layer and task-specific modules in combination with the incremental learning mechanism to optimize the multi-task learning framework;

[0095] S6. Based on the optimized multi-task learning framework, the multi-head self-attention mechanism is used to screen shared features and task-specific modules task by task to optimize the results of fire prediction, pest assessment and climate change modeling;

[0096] S7. Output fire risk forecasts, pest and disease control recommendations and ecological health assessment reports to provide support for forest disaster management and emergency decision-making.

[0097] In this implementation, S2 specifically includes:

[0098] S21, initializing a dynamic sharing layer, extracting multimodal data features through a convolutional neural network, wherein the multimodal data features include spatial distribution features of remote sensing images, time series features of meteorological data, and point features of ground monitoring data, and generating preliminary features;

[0099] S22, normalizing the preliminary features to obtain a standardized feature vector;

[0100] S23, integrating the standardized feature vectors by feature fusion technology to generate shared features, where the shared features are used as input for the task-specific module;

[0101] S24, constructing task-specific modules, wherein the task-specific modules include a fire prediction module, a pest assessment module, and a climate change modeling module, each module receiving the shared features as input;

[0102] S25. In the fire prediction module, based on the time series feature analysis, the time trend characteristics and spatial distribution characteristics of the fire occurrence are extracted to generate the fire prediction input features:

[0103] h t =σ(W x x t +W h h t-1 +b h );

[0104] o t =tanh(W o h t +b o );

[0105]

[0106] Among them, h t represents the hidden state of the gated recurrent unit network, σ represents the activation function, and W x , W h and W o represents the weight matrix, b h and b o represents the bias term, T represents the length of the time series, o t represents the output vector of the hidden state at time step t after further processing, tanh represents the hyperbolic tangent function, and F fire represents the fire prediction input features;

[0107] S26. In the pest assessment module, the graph neural network is used to process the shared features to generate pest assessment input features:

[0108]

[0109] in, represents the feature of node i in the l+1th layer of the graph neural network, N(i) represents the neighbor set of node i, represents the feature of node j in the lth layer of the graph neural network, W (l) represents the weight matrix of the lth layer, A ijrepresents the element of the adjacency matrix, deg(i) represents the degree of node i, deg(j) represents the degree of node j, Aggregate represents the aggregation operation of node features, L represents the total number of layers of the graph neural network, and F pest represents the input features of pest and disease assessment, and n represents the total number of nodes;

[0110] S27. In the climate change modeling module, a multi-head attention mechanism is used to process shared features to generate climate change modeling input features:

[0111] F climate =Concat(head 1 ,head 2 ,…,head m )·W out ;

[0112] Among them, F climate represents the input features of climate change modeling, W out Represents the weight matrix of the output layer, m represents the total number of attention heads, head represents the attention head, and Concat represents the concatenation operation.

[0113] In this implementation, S3 specifically includes:

[0114] S31. Initialize shared feature weight vector The weight vector length is the same as the shared feature vector length output by the dynamic shared layer:

[0115]

[0116] in, represents the weight of the u-th dimension in the initial state, u represents the dimension index of the shared feature, A represents the total number of dimensions of the shared feature, exp represents the exponential function, L u represents the loss value of the u-th dimension in the shared feature in the initial task performance, L v Represents the loss value of the vth dimension in the shared feature in the initial task performance;

[0117] S32. Define a multi-objective optimization function to evaluate the feature allocation strategy of individual parrots. The number of tasks in multi-task learning is known to be B, and the loss value of each task is expressed as R b , where b = 1, 2, ..., B, and two optimization objectives are set. The first optimization objective is to maximize the task coordination:

[0118]

[0119] Among them, S(W (c)) represents the task coordination degree of the cth parrot individual, and its value range is [0,1]. The larger the value, the better the coordination effect between tasks. tanh represents the hyperbolic tangent function, W (c) represents the shared feature allocation strategy of the cth parrot;

[0120] The second task is to maximize the balance of feature distribution:

[0121]

[0122] Among them, D(W (c) ) represents the characteristic distribution balance of the cth parrot individual, represents the weight assigned to the u-th dimension by the c-th parrot, w neam (c) represents the mean value of the weights of all dimensions for the cth parrot; D(W (c) ) is closer to 1, the more balanced the feature distribution is; conversely, the feature distribution is unbalanced;

[0123] S33. Conduct a multi-objective comprehensive evaluation of the shared feature allocation strategy for each parrot individual:

[0124] F(W (c) )=α·S(W (c) )+β·D(W (c) );

[0125] Among them, F(W (c) ) represents the comprehensive evaluation function value of the cth parrot individual, α and β represent the trade-off coefficients;

[0126] S34, execute the imitation learning mechanism, and bring the parrot individuals whose comprehensive evaluation function value is less than the set threshold ξ closer to the parrot individuals whose comprehensive evaluation function value is greater than the set threshold ξ:

[0127]

[0128] in, It represents the u-th dimension weight of the r-th parrot individual whose comprehensive evaluation function value is less than the set threshold ξ. represents the u-th dimension weight of the r-th parrot individual whose comprehensive evaluation function value is greater than the set threshold ξ, η represents the imitation learning rate, and ← represents the update operation;

[0129] S35, execute the independent exploration mechanism, randomly select and update from the parrot individuals whose comprehensive evaluation function value is equal to the set threshold ξ:

[0130]

[0131] in, represents the u-th dimension weight of the parrot individual whose comprehensive evaluation function value is equal to the set threshold ξ, δ represents the random disturbance factor, and Ω(-1,1) represents the random number sampled in the interval [-1,1];

[0132] S36. After completing imitation learning and independent exploration, recalculate the comprehensive evaluation function value of each individual parrot and sort the parrot group. If the iteration reaches the preset number of iterations or the multi-objective evaluation function converges, end the iteration and generate the optimal shared feature allocation strategy.

[0133] In this implementation, S4 specifically includes:

[0134] S41. Based on fire prediction, pest assessment and climate change modeling, construct a task dependency graph G = (V, E), where V = {T fire ,T pest ,T climate} represents the task node set, T fire represents the fire prediction task, T pest represents the pest assessment task, T climate represents the climate change modeling task, E = {e ij} represents the edge set between tasks, e ij Represents task T i and Task T j synergistic relationship;

[0135] S42. Calculate the collaborative relationship strength between tasks based on historical task data:

[0136]

[0137] Among them, ρ(T i ,T j ) represents task T i and Task T j The strength of the synergistic relationship between i ,R j ) represents task T i and Task T j The covariance between i ) represents task T i The standard deviation, σ(R j ) represents task T j The standard deviation of

[0138] S43. Normalize the collaborative relationship strength between tasks to obtain the standardized task dependency strength

[0139] S44, applying the optimal shared feature allocation strategy to the task dependency graph, and calculating the shared feature weight of each task node:

[0140]

[0141] in, Represents task T i The weighted shared features of shared represents the shared features of the dynamic shared layer, W * represents the optimal shared feature allocation strategy, |V| represents the total number of task nodes;

[0142] S45. Input the weighted shared features into the task-specific module to generate fire prediction results, pest and disease assessment results, and climate change modeling results.

[0143] In this implementation manner, S5 specifically includes:

[0144] S51, verify the fire prediction results, pest and disease assessment results, and climate change modeling results with historical data task by task, calculate the difference between the prediction results of each task and the historical data, and generate corresponding error data;

[0145] S52. Evaluate the overall error level of fire prediction, pest assessment and climate change modeling tasks based on the error data verified by each task, and mark the accuracy and error distribution of the prediction results of each task;

[0146] S53, storing the verified prediction results and error data in the disaster memory database, and recording the results of each verification;

[0147] S54. Combined with the historical error data in the disaster memory library, the parameters of the dynamic shared layer and the task-specific module are updated through the incremental learning mechanism, the overall error is used to perform gradient updates on the model parameters, and the shared feature allocation weights and the feature processing strategies of the task-specific modules are adjusted:

[0148]

[0149] Among them, Θ new represents the updated parameters, Θ old represents the parameters before updating, τ represents the control factor, represents the gradient calculation, B represents the number of tasks in multi-task learning, represents the validation error of task b, κ represents the regularization coefficient, represents the parameter of the mth historical optimization record in the disaster memory database, and Y represents the number of historical optimization records in the disaster memory database;

[0150] S55, repeat the verification and update process until the overall error level of the multi-task learning framework reaches a preset convergence threshold.

[0151] In this implementation manner, S6 specifically includes:

[0152] S61. Based on the optimized multi-task learning framework, the shared features and task-specific modules are screened task by task, and the attention weight matrix of each task is calculated through the multi-head self-attention mechanism:

[0153]

[0154] in, represents the attention weight matrix of task b, softmax represents the normalization function, represents the query matrix of task b, represents the key matrix of task b, d b Represents the key matrix dimension of task b;

[0155] S62. Use the attention weight matrix of the task to perform weighted processing on the value matrix to generate a feature vector after task-by-task screening:

[0156]

[0157] in, represents the screening feature vector of task b, represents the value matrix of task b;

[0158] S63. Perform task-by-task optimization on the selected feature vectors, remove redundant feature information, and input them into task-specific modules to further optimize the results of fire prediction, pest and disease assessment, and climate change modeling.

[0159] Embodiment 1:

[0160] In order to verify the feasibility of the present invention in implementation, the present invention is applied to the management of a large forest ecosystem in a northern temperate zone. The region is located at high latitudes with a complex ecological environment and faces the combined threat of multiple disasters such as forest fires, pests and diseases, and climate change all year round. The research team selected a typical sub-region of the region for the experiment, which covers an area of ​​about 2,000 square kilometers. The forest type is mainly coniferous forest, and some broad-leaved forests are also distributed. Historical data show that the region has an average of 17 fires per year in the past five years, the area of ​​pests and diseases has spread to more than 150 square kilometers, and the climate fluctuates significantly, and extreme drought events occur frequently, posing a significant threat to the regional ecosystem.

[0161] In the experiment, multimodal data were first collected, including remote sensing images, meteorological data, and ground monitoring data. The remote sensing image data comes from the Sentinel-2 satellite, with a time span of 2018 to 2023 and a spatial resolution of 10 meters. The content covers ecological characteristics such as vegetation cover, fire point distribution, and soil moisture. Meteorological data are provided by regional meteorological stations, including hourly temperature, humidity, wind speed, and precipitation. The ground monitoring data comes from 200 monitoring points deployed in the sub-area, mainly recording the spatial distribution and severity of forest pests and diseases. These data are processed by time alignment and space alignment technology and unified into a daily time step and a spatial resolution of 1 km.

[0162] The multi-task learning framework of the present invention is applied to first extract shared features of multimodal data through a dynamic sharing layer. In the experiment, remote sensing images mainly provide spatial distribution features, such as vegetation health index and fire point distribution; meteorological data contributes time series features, such as wind speed changes and precipitation trends; and ground monitoring data provides point features, such as the distribution range and density of pests and diseases. The dynamic sharing layer fuses these features to generate shared features for fire prediction, pest and disease assessment, and climate change modeling.

[0163] In the fire prediction task, the model predicted high-risk areas where fires may occur in the next week through time series feature analysis. The results showed that the model's prediction accuracy for the time of fire occurrence reached 87.4%, and the overlap rate of spatial distribution prediction reached 90.2%. In the pest assessment task, the model successfully predicted the spread and severity of pests and diseases by processing shared features through graph neural networks, and the overlap rate with the actual ground monitoring data reached 83.6%. In the climate change modeling task, the multi-head self-attention mechanism was used to predict the trend of vegetation coverage changes in the next quarter. The results showed that the model's prediction accuracy for areas with reduced vegetation coverage reached 85.7%.

[0164] To further verify the dynamic adaptability of the present invention, the research team divided the experiment into two phases. The first phase used data from 2018 to 2022 to train the model, and the second phase used real-time data from 2023 to perform incremental learning on the model. The experiment showed that the incremental learning mechanism significantly improved the model's adaptability to the new disaster pattern in 2023, with the accuracy of fire prediction increased from 81.2% to 89.5%, the accuracy of pest and disease assessment increased from 77.8% to 85.1%, and the accuracy of climate change modeling increased from 80.4% to 87.3%.

[0165] The experimental results clearly demonstrate the superiority of the present invention in the collaborative prediction and assessment of forest ecological disasters. The deep fusion of multimodal data significantly improves the prediction ability of the model, the division of labor and cooperation between the dynamic sharing layer and task-specific modules improves the overall efficiency of multi-task learning, and the Parrot optimization algorithm and incremental learning mechanism enhance the dynamic adaptability of the model.

[0166] Table 1 Comparison of experimental results of forest ecological disaster prediction and assessment

[0167]

[0168]

[0169] It can be seen from Table 1 above that the performance of the present invention in the collaborative prediction and assessment of forest ecological disasters is significantly better than that of traditional models, especially in terms of prediction ability after incremental optimization. In the fire prediction task, the model successfully predicted the time and spatial distribution of fire occurrence by combining remote sensing images and meteorological data. The time prediction accuracy of the initial model was 81.2%, which was increased to 89.5% after incremental optimization, and the spatial prediction accuracy was increased from 83.7% to 91.8%. The overlap rate of actual observation data also reached more than 90%, and the error reduction was 8.3% and 8.1% respectively. These results show that through dynamic feature sharing and optimization mechanism, the model can effectively capture the dynamic characteristics of fire and improve the accuracy and reliability of prediction.

[0170] In the task of pest assessment, the model used features extracted from ground monitoring data and remote sensing images to accurately assess the spread and severity of pests and diseases. The initial model predicted the spread of pests and diseases with an accuracy of 77.8%, which increased to 85.1% after incremental optimization; the prediction accuracy of severity increased from 78.4% to 85.6%. The overlap rates of actual observation data reached 83.6% and 84.2%, respectively. This shows that the model performs well in the deep fusion and feature optimization of multimodal data, and can effectively model the complex dynamics of pests and diseases.

[0171] In the climate change modeling task, the model successfully predicted the changing trend of vegetation cover by jointly processing remote sensing images and meteorological data. The accuracy of the initial model was 80.4%, which was increased to 87.3% after incremental optimization, and the overlap rate of actual observation data was 85.7%. The error reduction reached 6.9%, fully demonstrating the advantages of the invention in dynamic adaptability and long-term trend prediction.

[0172] Overall, the experimental results show that the present invention is significantly better than the initial model in the three major tasks of fire prediction, pest assessment and climate change modeling. The incremental learning mechanism effectively enhances the dynamic adaptability of the model, and the multimodal data fusion and dynamic feature sharing mechanism further improve the comprehensiveness and accuracy of the prediction. In addition, the high overlap rate between the prediction results of each task and the actual observation data shows that the present invention has strong practicality and scientific reliability, and provides efficient technical support for forest disaster management and emergency decision-making.

[0173] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A collaborative prediction and assessment method for forest ecological disasters based on multi-task learning, characterized in that: The steps include: S1. Collect multimodal data of forest ecological disasters, including remote sensing images, meteorological data and ground monitoring data, and unify the multimodal data features through time alignment and space alignment technology; S2. Construct a multi-task learning framework, including a dynamic shared layer and task-specific modules, where the dynamic shared layer is used to extract shared features from multimodal data features, and the task-specific modules generate fire prediction input features, pest and disease assessment input features, and climate change modeling input features based on the shared features; S3. Use the Parrot optimization algorithm to optimize the shared feature weights of the dynamic shared layer, dynamically adjust the allocation ratio of shared features through imitation learning and independent exploration mechanisms, and generate the optimal shared feature allocation strategy; S4. Applying the optimal shared feature allocation strategy to a task dependency graph, wherein the task dependency graph is constructed based on the synergistic relationship between fire prediction, pest and disease assessment, and climate change modeling, and allocating shared feature weights between tasks through a dynamic attention mechanism to generate fire prediction results, pest and disease assessment results, and climate change modeling results; S5. Verify the fire prediction results, pest and disease assessment results, and climate change modeling results with historical data, store the verified results in the disaster memory bank, and update the parameters of the dynamic shared layer and task-specific modules in combination with the incremental learning mechanism to optimize the multi-task learning framework; S6. Based on the optimized multi-task learning framework, the multi-head self-attention mechanism is used to screen shared features and task-specific modules task by task to optimize the results of fire prediction, pest assessment and climate change modeling; S7. Output fire risk forecasts, pest and disease control recommendations and ecological health assessment reports to provide support for forest disaster management and emergency decision-making.

2. According to claim 1, a method for collaborative prediction and assessment of forest ecological disasters based on multi-task learning is characterized in that: The S2 specifically includes: S21, initializing a dynamic sharing layer, extracting multimodal data features through a convolutional neural network, wherein the multimodal data features include spatial distribution features of remote sensing images, time series features of meteorological data, and point features of ground monitoring data, and generating preliminary features; S22, normalizing the preliminary features to obtain a standardized feature vector; S23, integrating the standardized feature vectors by feature fusion technology to generate shared features, where the shared features are used as input for the task-specific module; S24, constructing task-specific modules, wherein the task-specific modules include a fire prediction module, a pest assessment module, and a climate change modeling module, each module receiving the shared features as input; S25. In the fire prediction module, based on the time series feature analysis, the time trend characteristics and spatial distribution characteristics of the fire occurrence are extracted to generate the fire prediction input features: h t =σ(W x x t +W h h t-1 +b h ); o t =tanh(W o h t +b o ); Among them, h t represents the hidden state of the gated recurrent unit network, σ represents the activation function, and W x , W h and W o represents the weight matrix, b h and b o represents the bias term, T represents the length of the time series, o t represents the output vector of the hidden state at time step t after further processing, tanh represents the hyperbolic tangent function, and F fire represents the fire prediction input features; S26. In the pest assessment module, the graph neural network is used to process the shared features to generate pest assessment input features: in, represents the feature of node i in the l+1th layer of the graph neural network, N(i) represents the neighbor set of node i, represents the feature of node j in the lth layer of the graph neural network, W (l) represents the weight matrix of the lth layer, A ij represents the element of the adjacency matrix, deg(i) represents the degree of node i, deg(j) represents the degree of node j, Aggregate represents the aggregation operation of node features, L represents the total number of layers of the graph neural network, and F pest represents the input features of pest and disease assessment, and n represents the total number of nodes; S27. In the climate change modeling module, a multi-head attention mechanism is used to process shared features to generate climate change modeling input features: F climate =Concat(head1,head2,…,head m )·W out ; Among them, F climate represents the input features of climate change modeling, W out Represents the weight matrix of the output layer, m represents the total number of attention heads, head represents the attention head, and Concat represents the concatenation operation.

3. The method for collaborative prediction and assessment of forest ecological disasters based on multi-task learning according to claim 1 is characterized in that: The S3 specifically includes: S31. Initialize shared feature weight vector The weight vector length is the same as the shared feature vector length output by the dynamic shared layer: in, represents the weight of the u-th dimension in the initial state, u represents the dimension index of the shared feature, A represents the total number of dimensions of the shared feature, exp represents the exponential function, and L u represents the loss value of the u-th dimension in the shared feature in the initial task performance, L v Represents the loss value of the vth dimension in the shared feature in the initial task performance; S32. Define a multi-objective optimization function to evaluate the feature allocation strategy of individual parrots. The number of tasks in multi-task learning is known to be B, and the loss value of each task is expressed as R b , where b = 1, 2, ..., B, and two optimization objectives are set. The first optimization objective is to maximize the task coordination: Among them, S(W (c) ) represents the task coordination degree of the cth parrot individual, and its value range is [0,1]. The larger the value, the better the coordination effect between tasks. tanh represents the hyperbolic tangent function, W (c) represents the shared feature allocation strategy of the cth parrot; The second task is to maximize the balance of feature distribution: Among them, D(W (c) ) represents the characteristic distribution balance of the cth parrot individual, represents the weight assigned by the cth parrot to the uth dimension, w neam (c) represents the mean value of the weights of all dimensions for the cth parrot; D(W (c) ) is closer to 1, indicating that the feature distribution is more balanced; conversely, the feature distribution is unbalanced; S33. Conduct a multi-objective comprehensive evaluation of the shared feature allocation strategy for each parrot individual: F(W (c) )=α·S(W (c) )+β·D(W (c) ); Among them, F(W (c) ) represents the comprehensive evaluation function value of the cth parrot individual, α and β represent the trade-off coefficients; S34, execute the imitation learning mechanism, and bring the parrot individuals whose comprehensive evaluation function value is less than the set threshold ξ closer to the parrot individuals whose comprehensive evaluation function value is greater than the set threshold ξ: in, It represents the weight of the u-th dimension of the r-th parrot individual whose comprehensive evaluation function value is less than the set threshold ξ. represents the u-th dimension weight of the r-th parrot individual whose comprehensive evaluation function value is greater than the set threshold ξ, η represents the imitation learning rate, and ← represents the update operation; S35, execute the independent exploration mechanism, randomly select and update from the parrot individuals whose comprehensive evaluation function value is equal to the set threshold ξ: in, represents the u-th dimension weight of the parrot individual whose comprehensive evaluation function value is equal to the set threshold ξ, δ represents the random disturbance factor, and Ω(-1,1) represents the random number sampled in the interval [-1,1]; S36. After completing imitation learning and independent exploration, recalculate the comprehensive evaluation function value of each individual parrot and sort the parrot group. If the iteration reaches the preset number of iterations or the multi-objective evaluation function converges, end the iteration and generate the optimal shared feature allocation strategy.

4. The method for collaborative prediction and assessment of forest ecological disasters based on multi-task learning according to claim 1 is characterized in that: The S4 specifically includes: S41. Based on fire prediction, pest assessment and climate change modeling, construct a task dependency graph G = (V, E), where V = {T fire ,T pest ,T climate } represents the task node set, T fire represents the fire prediction task, T pest represents the pest assessment task, T climate represents the climate change modeling task, E = {e ij } represents the edge set between tasks, e ij Represents task T i and Task T j synergistic relationship; S42. Calculate the collaborative relationship strength between tasks based on historical task data: Among them, ρ(T i ,T j ) represents task T i and Task T j The strength of the synergistic relationship between i ,R j ) represents task T i and Task T j The covariance between σ(R i ) represents task T i The standard deviation, σ(R j ) represents task T j The standard deviation of S43. Normalize the collaborative relationship strength between tasks to obtain the standardized task dependency strength S44, applying the optimal shared feature allocation strategy to the task dependency graph, and calculating the shared feature weight of each task node: in, Represents task T i The weighted shared features of shared represents the shared features of the dynamic shared layer, W * represents the optimal shared feature allocation strategy, |V| represents the total number of task nodes; S45. Input the weighted shared features into the task-specific module to generate fire prediction results, pest and disease assessment results, and climate change modeling results.

5. The method for collaborative prediction and assessment of forest ecological disasters based on multi-task learning according to claim 1 is characterized in that: The S5 specifically includes: S51, verify the fire prediction results, pest and disease assessment results, and climate change modeling results with historical data task by task, calculate the difference between the prediction results of each task and the historical data, and generate corresponding error data; S52. Evaluate the overall error level of fire prediction, pest assessment and climate change modeling tasks based on the error data verified by each task, and mark the accuracy and error distribution of the prediction results of each task; S53, storing the verified prediction results and error data in the disaster memory database, and recording the results of each verification; S54. Combined with the historical error data in the disaster memory library, the parameters of the dynamic shared layer and the task-specific module are updated through the incremental learning mechanism, the overall error is used to perform gradient updates on the model parameters, and the shared feature allocation weights and the feature processing strategies of the task-specific modules are adjusted: Among them, Θ new represents the updated parameters, Θ old represents the parameters before updating, τ represents the control factor, represents the gradient calculation, B represents the number of tasks in multi-task learning, represents the validation error of task b, κ represents the regularization coefficient, represents the parameter of the mth historical optimization record in the disaster memory database, and Y represents the number of historical optimization records in the disaster memory database; S55, repeat the verification and update process until the overall error level of the multi-task learning framework reaches a preset convergence threshold.

6. The method for collaborative prediction and assessment of forest ecological disasters based on multi-task learning according to claim 1 is characterized in that: The S6 specifically includes: S61. Based on the optimized multi-task learning framework, the shared features and task-specific modules are screened task by task, and the attention weight matrix of each task is calculated through the multi-head self-attention mechanism: in, represents the attention weight matrix of task b, softmax represents the normalization function, represents the query matrix of task b, represents the key matrix of task b, d b Represents the key matrix dimension of task b; S62. Use the attention weight matrix of the task to perform weighted processing on the value matrix to generate a feature vector after task-by-task screening: in, represents the screening feature vector of task b, represents the value matrix of task b; S63. Perform task-by-task optimization on the selected feature vectors, remove redundant feature information, and input them into task-specific modules to further optimize the results of fire prediction, pest and disease assessment, and climate change modeling.

Images