A forest fire risk assessment method based on interpretable deep learning

Abstract

The application discloses a forest fire risk assessment method based on an interpretable deep learning, relates to the technical field of forestry disaster prevention, and comprises the following steps: collecting multi-source heterogeneous data and preprocessing the multi-source heterogeneous data to generate a multidimensional data cube; extracting a logical relationship for forming a forest fire risk based on the multidimensional data cube, constructing a forest fire risk domain knowledge graph based on the logical relationship of the forest fire risk; taking the multidimensional data cube as a training sample, constructing and training a concept bottleneck interpretable deep learning model, and mapping the multidimensional data cube input into the concept bottleneck interpretable deep learning model to a semantic concept layer defined by the forest fire risk domain knowledge graph. The application generates a multidimensional data cube by collecting multi-source heterogeneous data, constructs a forest fire risk domain knowledge graph, trains a concept bottleneck interpretable deep learning model, maps input data to a semantic concept layer, and outputs a fire risk grade and a concept activation result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of forest disaster prevention, and in particular to a forest fire risk assessment method based on interpretable deep learning. BACKGROUND

[0002] In recent years, with the wide application of deep learning technology in the field of forest fire risk assessment, the fire risk prediction method based on multi-source data fusion has made significant progress. The hybrid architecture based on convolutional neural network and long short-term memory network has become the current mainstream technical solution. By integrating satellite remote sensing data, meteorological observation data and terrain data, using three-dimensional convolutional network to extract spatial features, combining bidirectional LSTM to process time series changes, the end-to-end prediction of forest fire risk level is realized, which shows high prediction accuracy in multiple area tests, confirming the effectiveness of deep learning in processing multi-source heterogeneous forest fire risk data. The existing technology generally adopts a feature-level fusion strategy to automatically extract high-dimensional feature representations of data from different sources through neural networks. This data-driven method can capture complex nonlinear relationships and provide reliable technical support for fire warning.

[0003] However, the existing technology still has obvious limitations in handling model interpretability. The current mainstream method mainly relies on black-box deep neural network architecture, which lacks transparency in the decision-making process and is difficult to provide causal logic explanation for fire risk assessment. It cannot clearly indicate which specific factors dominate the high-risk determination and the interaction between these factors. This lack of explanation not only reduces the decision-making confidence of forestry management departments, but also hinders the continuous optimization of the model. It is particularly worth noting that when the model prediction result is different from the expert experience, the existing technology lacks an effective mechanism to verify and adjust the reasoning process of the model. By constructing a forest fire risk domain knowledge graph and a concept bottleneck model, the prediction ability of deep learning can be combined with the structured expression of domain knowledge. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides a forest fire risk assessment method based on interpretable deep learning, which solves the key problem of the existing black-box deep neural network in the decision-making process of fire risk assessment, which is not transparent and lacks causal logic explanation.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] In a first aspect, the present application provides a forest fire risk assessment method based on interpretable deep learning, which includes collecting multi-source heterogeneous data and preprocessing to generate a multi-dimensional data cube;

[0008] Based on the multi-dimensional data cube, the logical relationship of forming the forest fire danger is extracted, and based on the logical relationship of the forest fire danger, a forest fire danger domain knowledge graph is constructed;

[0009] Based on the multi-dimensional data cube as a training sample, a concept bottleneck interpretable deep learning model is constructed and trained, and the multi-dimensional data cube input into the concept bottleneck interpretable deep learning model is mapped to the semantic concept layer defined by the forest fire danger domain knowledge graph to obtain a fire danger grade result and a semantic concept layer activation result;

[0010] The rules in the forest fire danger domain knowledge graph are converted into a differentiable constraint loss function, and the training process of the concept bottleneck interpretable deep learning model is jointly optimized using the differentiable constraint loss function;

[0011] The concept bottleneck interpretable deep learning model that has been jointly optimized is deployed as a risk assessment service;

[0012] Based on the risk assessment service, a fire danger grade result and a semantic concept layer activation result are received, and an interpretable forest fire danger assessment report is generated.

[0013] As a preferred scheme of the forest fire danger assessment method based on interpretable deep learning, wherein the multi-source heterogeneous data is collected and preprocessed to generate a multi-dimensional data cube, including the following steps,

[0014] Through an API interface, original remote sensing images, meteorological time series data, digital elevation model data and historical forest fire point record vector data are collected from a satellite data center, a meteorological database and a geographic information platform, and spatial data is uniformly converted into a geographic coordinate system and a projection unit;

[0015] According to the target area and the evaluation requirement, a uniform spatial resolution grid is defined, and based on the meteorological time series data, a continuous surface data is generated by using the collaborative kriging interpolation method of elevation and distance;

[0016] The original remote sensing image is cloud detected, and the cloud coverage pixel is repaired by using the time series spectral similarity method to generate cloud-free data;

[0017] Based on the preprocessed original remote sensing image and the digital elevation model data, normalized vegetation index, normalized water index, land surface temperature, slope and slope direction are obtained, and Z-Score standardization processing is performed to generate standardized feature data, and the standardized multi-source feature data is organized according to time, row, column and channel dimensions to generate a multi-dimensional data cube.

[0018] As a preferred scheme of the forest fire risk assessment method based on the interpretable deep learning, wherein: based on the multidimensional data cube, logical relationships forming the forest fire risk are extracted, and based on the logical relationships of the forest fire risk, a forest fire risk domain knowledge graph is constructed, including the following steps,

[0019] Based on the multidimensional data cube, the entity type and the relationship type of the knowledge graph are defined, the multivariate time series data in the multidimensional data cube is input into the PC algorithm, the conditional independence between variables is calculated, and an initial data-driven causal graph is obtained;

[0020] The data-driven initial causal graph is reviewed to obtain a set of causal relationships with weights, and based on the multidimensional data cube, a quantitative mapping rule of the level of each entity concept in the knowledge graph is defined

[0021] Using a graph database, the set of causal relationships with weights is converted into a graph structure, the nodes are entities, the edges are relationships, and the confidence score is taken as an attribute of the edge, and a forest fire risk domain knowledge graph is constructed.

[0022] As a preferred scheme of the forest fire risk assessment method based on the interpretable deep learning, wherein: based on the multidimensional data cube as a training sample, a concept bottleneck interpretable deep learning model is constructed and trained, including the following steps,

[0023] The multidimensional data cube is input into a double-branch encoder composed of 3D CNN and LSTM, and the spatio-temporal features are extracted and fused to obtain a high-level fusion feature tensor;

[0024] The high-level fusion feature tensor is input into the concept layer, and the probability vector of the existence of semantic concepts is calculated by the Sigmoid activation function;

[0025] The probability vector of the existence of semantic concepts is input into the self-attention network, the correlation weight between concepts is calculated, and a weighted concept importance vector is obtained;

[0026] The weighted concept importance vector is input into the linear classifier, and the fire risk grade probability distribution is calculated using the Softmax function;

[0027] The real label of the fire risk grade and the real label of the concept are combined to calculate the main task loss, the concept prediction loss and the attention sparsity loss, and the sum is taken to obtain the total loss value of the concept bottleneck interpretable deep learning model;

[0028] Using the back propagation algorithm, the model parameters from the encoder to the classifier are iteratively updated to minimize the total loss value, and a trained concept bottleneck interpretable deep learning model is obtained.

[0029] As a preferred scheme of the forest fire risk assessment method based on the interpretable deep learning provided in the present application, wherein: the multi-dimensional data cube input into the concept bottleneck interpretable deep learning model is mapped to a semantic concept layer defined by a forest fire risk domain knowledge graph, to obtain a fire risk grade result and a semantic concept layer activation result, including the following steps,

[0030] The multi-dimensional data cube is subjected to batch normalization processing to generate a normalized multi-dimensional input tensor, the input tensor is input into a multi-modal feature encoder to obtain a normalized high-level vector, and the normalized high-level vector is input into a concept perception attention module to obtain a spatial attention weight map;

[0031] The concept attention weight map is multiplied element by element with the high-level fusion feature tensor and combined with a residual connection to obtain a weighted attention perception feature tensor;

[0032] The attention perception feature tensor is flattened and input into a concept layer to obtain a semantic concept layer activation result using a Sigmoid activation function;

[0033] The semantic concept layer activation result is obtained using a classifier weight matrix and a semantic concept layer activation result, a contribution degree distribution matrix of the semantic concept layer activation result to fire risk grade prediction, and a concept activation vector is input into a linear classifier to obtain a fire risk grade result.

[0034] As a preferred scheme of the forest fire risk assessment method based on the interpretable deep learning provided in the present application, wherein: rules in a forest fire risk domain knowledge graph are converted into a differentiable constraint loss function, and a training process of a concept bottleneck interpretable deep learning model is jointly optimized using the differentiable constraint loss function, including the following steps,

[0035] Rules connected by logical relationships are extracted from the forest fire risk domain knowledge graph, logical operators in the rules connected by logical relationships are replaced with fuzzy logic operators, and a loss function template is generated;

[0036] A probability vector of the existence of a semantic concept is substituted into the rule template converted by the fuzzy logic to obtain a rule violation degree loss value;

[0037] The rule loss value is weighted and summed according to a confidence score attached to each rule in the forest fire risk domain knowledge graph to obtain a knowledge constraint loss;

[0038] The knowledge constraint loss, the binary cross-entropy loss of the fire risk grade prediction and the probability vector of the existence of the semantic concept are added to form a joint optimization objective function.

[0039] As a preferred scheme of the forest fire risk assessment method based on the interpretable deep learning provided in the present application, wherein: the concept bottleneck interpretable deep learning model subjected to joint optimization is deployed as a risk assessment service, including the following steps,

[0040] The trained concept bottleneck explainable deep learning model generates a lightweight high-performance inference model, and the lightweight high-performance inference model is defined as a RESTful and gRPC service interface, and a standardized input and output are defined.

[0041] The interpretive results of the high-frequency query set a hash key cache mechanism, dynamically allocate inference tasks to edge devices according to the device computing power and network state of the request source, package as a Docker image, deploy services in a Kubernetes cluster, and configure automatic scaling and load balancing strategies.

[0042] As a preferred scheme of the forest fire risk assessment method based on the explainable deep learning, the risk assessment service receives the fire risk level result and the semantic concept layer activation result, and generates an explainable forest fire risk assessment report, including the following steps,

[0043] The fire risk level probability distribution, the concept probability vector and the contribution matrix are received from the risk assessment service API, and are parsed into structured data objects;

[0044] Based on the fire risk level probability distribution, a fire risk probability map is generated by mapping the spatial grid points;

[0045] The fire risk probability map is spatially clustered to identify high-risk areas and sort the dominant concept contribution, and a high-risk area is obtained;

[0046] The high-risk area and the concept contribution are combined to generate an explainable forest fire risk assessment report.

[0047] In a second aspect, the present application provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and wherein the computer program is executed by the processor to implement any step of the forest fire risk assessment method based on the explainable deep learning according to the first aspect of the present application.

[0048] In a third aspect, the present application provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement any step of the forest fire risk assessment method based on the explainable deep learning according to the first aspect of the present application.

[0049] The present application has the beneficial effects that: a multi-dimensional data cube is generated by collecting multi-source heterogeneous data, a forest fire risk field knowledge graph is constructed, a concept bottleneck explainable deep learning model is trained, input data is mapped to a semantic concept layer, a fire risk level and a concept activation result are output, knowledge graph rules are further converted into a differentiable constraint loss function for joint optimization, deployed as a risk assessment service and an explainability report is generated, realizing the whole-process explainable forest fire risk assessment from data preprocessing, model training to service deployment, and solving the problem of opaque decision-making process of traditional black box models. BRIEF DESCRIPTION OF DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0051] Fig. 1 The flowchart of the forest fire risk assessment method based on explainable deep learning.

[0052] Fig. 2 The schematic diagram of the forest fire risk field knowledge graph.

[0053] Fig. 3 The training flowchart of the concept bottleneck explainable deep learning model.

[0054] Fig. 4 The flowchart of the multi-dimensional data cube generation. DETAILED DESCRIPTION

[0055] In order to make the above-mentioned purposes, features and advantages of the present application more apparent and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the drawings of the specification.

[0056] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, but the present application can also be implemented in other ways different from those described herein, and those skilled in the art can make similar generalizations without departing from the connotation of the present application, therefore the present application is not limited to the specific embodiments disclosed below.

[0057] Secondly, the term one embodiment or embodiments herein can include specific features, structures or characteristics included in at least one implementation of the present application. In different places in this specification, one embodiment does not refer to the same embodiment, nor is it an independent or alternative embodiment that excludes other embodiments.

[0058] REFERENCE Figs. 1-4For an embodiment of the present application, the embodiment provides a forest fire risk assessment method based on interpretable deep learning, comprising the following steps:

[0059] S1, collecting multi-source heterogeneous data and preprocessing to generate a multi-dimensional data cube.

[0060] S1.1, through the API interface from the satellite data center, meteorological database and geographic information platform, collect original remote sensing image, meteorological time series data, digital elevation model data and historical forest fire point record vector data, and unify the spatial data conversion to geographic coordinate system and projection unit.

[0061] Further, through the API interface of the satellite data center, Sentinel-2 multispectral remote sensing image and MODIS land surface temperature data are obtained, ERA5 reanalysis meteorological time series data and ground meteorological station observation data are collected from the meteorological database, SRTM digital elevation model data is downloaded through the geographic information platform, and MODIS active fire point product historical record vector data is obtained; using the GDAL library, all spatial data are converted to the same coordinate system, the original remote sensing image, digital elevation model data and historical forest fire point record vector data are converted to WGS84 UTM projection coordinate system, and the meteorological station data is matched to the same coordinate system through spatial interpolation, completing the spatial reference alignment of multi-source data.

[0062] S1.2, according to the target area and evaluation requirements, define a uniform spatial resolution grid, based on meteorological time series data, use the collaborative kriging interpolation method of elevation and distance to generate continuous surface data.

[0063] Further, according to the target area range and evaluation accuracy requirement, the spatial resolution grid size is determined, based on the temperature, humidity, wind speed and precipitation observation values in the meteorological time series data, combined with the elevation information in the digital elevation model data, the collaborative kriging interpolation considering elevation and horizontal distance is used to obtain the meteorological element value of each grid point, and the vertical gradient effect and spatial autocorrelation of meteorological elements with elevation change are considered in the interpolation process, to generate continuous surface data.

[0064] S1.3, cloud detection is performed on the original remote sensing image, and the time series spectral similarity method is used to repair the cloud covered pixels to generate cloud-free data.

[0065] Further, the Scene Classification Layer is applied to the Sentinel-2 multispectral remote sensing image to identify a cloud mask and mark cloud-covered pixels; for the marked cloud-polluted pixels, all co-located pixels in the target region observed on cloudless days in the past 30 days are searched, the historical cloudless pixel value with the highest spectral similarity is selected, and the value of the cloud-polluted pixel is replaced by linear adjustment combined with the change trend of the adjacent effective pixels on the target date; the above repair operation is performed on all cloud mask marked regions to generate a complete cloudless multispectral remote sensing image dataset.

[0066] S1.4, based on the preprocessed original remote sensing image and digital elevation model data, normalized vegetation index, normalized moisture index, land surface temperature, slope and slope direction are obtained, and Z-Score standardization processing is performed to generate standardized feature data; the standardized multisource feature data is organized according to time, row, column and channel dimensions to generate a multidimensional data cube.

[0067] Further, based on the preprocessed Sentinel-2 multispectral remote sensing image, the normalized vegetation index (NDVI) and the normalized moisture index (NDMI) of each pixel, the land surface temperature (LST) is retrieved using the MODIS thermal infrared band; the slope and slope direction are calculated based on the SRTM digital elevation model data; the mean and standard deviation of the normalized vegetation index, the normalized moisture index, the thermal infrared band retrieved land surface temperature, the slope and the slope direction are calculated respectively, and Z-Score standardization processing is performed; the standardized meteorological interpolation data, remote sensing derived features and terrain features are organized according to time dimension, spatial row number, spatial column number and feature channel four dimensions to construct a multidimensional data cube with unified space-time reference, wherein the time dimension is daily scale, the spatial resolution is consistent with the preset grid, and the multidimensional data cube is generated.

[0068] S2, based on the multidimensional data cube, the logical relationship forming forest fire danger is extracted, and based on the logical relationship of forest fire danger, a forest fire danger domain knowledge graph is constructed.

[0069] S2.1, based on the multidimensional data cube, the entity type and the relationship type of the knowledge graph are defined, the multivariate time series data in the multidimensional data cube is input into the PC algorithm, the conditional independence between variables is calculated, and an initial causal graph driven by data is obtained.

[0070] Specifically, the expression is,

[0071] Ind(X,Y|Z) if p-value > a;

[0072] where Ind is conditional independence relationship, Y is historical forest fire point record vector data, X is temperature variable, Z is conditional variable set, and p-value is statistical test.

[0073] S2.2, review the data-driven initial causal graph to obtain a set of causal relationships with weights, and define a quantitative mapping rule for the level of each entity concept in the knowledge graph based on the multi-dimensional data cube.

[0074] Further, based on the initial causal graph, the causal relationships between variables are reviewed and verified by domain experts, and the causal relationship edges that are statistically significant and consistent with the physical mechanism are retained, and each edge is assigned an expert confidence weight. Using the statistical distribution characteristics of each variable in the multi-dimensional data cube, a level division threshold is defined for each entity concept in the knowledge graph (for example, the high temperature concept is divided into three levels: mild high temperature (30-35℃), moderate high temperature (35-40℃), and severe high temperature (>40℃)), and a quantitative mapping rule is established from continuous observation values to discrete concept levels, forming a set of causal relationships with weight annotations and a concept level quantitative rule library.

[0075] S2.3, using a graph database, converting the set of causal relationships with weights into a graph structure, with nodes as entities and edges as relationships, and the confidence score as an edge attribute, to construct a forest fire risk domain knowledge graph.

[0076] Further, a node set is created in the Neo4j graph database, each node corresponding to an entity concept in the forest fire risk domain knowledge graph (such as high temperature, drought, and steep slope), and the node attributes include concept name and level division rule. According to the set of causal relationships with weights, relationship edges are created, with types leadsTo or inhibits, and the expert confidence score as the confidence attribute of the edge; for example, a leadsTo relationship edge is created from the high temperature node to the drought node, with a confidence attribute value of 0.9. All nodes and relationship edges are imported into the graph database to construct the forest fire risk domain knowledge graph.

[0077] S3, based on the multi-dimensional data cube as training samples, construct and train a concept bottleneck interpretable deep learning model.

[0078] S3.1, input the multi-dimensional data cube into a dual-branch encoder composed of 3D CNN and LSTM, extract and fuse the spatio-temporal features to obtain a high-level fusion feature tensor.

[0079] Further, the remote sensing image and terrain features in the multi-dimensional data cube are input into the 3D CNN branch, and local space-time features are extracted by sliding the three-dimensional convolution kernel in the time, spatial height and spatial width dimensions. After multi-level convolution and pooling operations, a spatial feature tensor is output; meanwhile, the meteorological time series data in the multi-dimensional data cube are input into the LSTM branch, and the time evolution law of each grid point of the meteorological variable is captured through the gating mechanism, and a time series feature vector is output. The spatial feature tensor output by the 3D CNN branch and the time series feature vector output by the LSTM branch are spliced in the channel dimension, and feature fusion and dimension reduction are performed through the full connection layer to obtain a high-level fusion feature tensor.

[0080] S3.2, the high-level fusion feature tensor is input into the concept layer, and the probability vector of the existence of semantic concepts is calculated through the Sigmoid activation function.

[0081] Specifically, the expression is,

[0082] C hat =σ(W c ·F fused +b);

[0083] Wherein, C hat is the probability vector of the existence of semantic concepts, W c is the weight matrix of the concept layer, F fused is the high-level fusion feature tensor, and b is the bias vector of the concept layer.

[0084] S3.3, the probability vector of the existence of semantic concepts is input into the self-attention network, and the correlation weight between concepts is calculated to obtain a weighted concept importance vector.

[0085] Specifically, the expression is,

[0086]

[0087] Wherein, C attended is the weighted concept importance vector, Q is the query matrix, V is the value, d k is the dimension of the key vector, and K T is the transpose of the key matrix.

[0088] S3.4, the weighted concept importance vector is input into the linear classifier, and the fire danger grade probability distribution is calculated using the Softmax function.

[0089] Specifically, the expression is,

[0090] U hat =Softmax(W y ·C attended +b);

[0091] wherein U hat is the fire risk level probability distribution, W y is the fire risk level corresponding to each column of the weight matrix.

[0092] S3.5, combined with the real label of the fire risk level and the real label of the concept, calculate the main task loss, the concept prediction loss and the attention sparsity loss, and sum to obtain the total loss value of the concept bottleneck explainable deep learning model.

[0093] Specifically, the main task loss expression is,

[0094]

[0095] wherein, is the main task loss, N y is the number of fire risk level categories, y ij is the real fire risk level label of sample i, is the fire risk level probability distribution, B is the batch size, j is the fire risk level category index, and i is the sample index.

[0096] Specifically, the concept prediction loss expression is,

[0097]

[0098] wherein, is the concept prediction loss, N c is the number of semantic concepts, c ik is the real label of concept k of sample i, is the concept probability, and k is the concept index.

[0099] Specifically, the attention sparsity loss expression is,

[0100]

[0101] wherein, is the attention sparsity loss, A ikm is the attention weight of concept k of sample i to target concept m, and m is the target concept.

[0102] Specifically, the total loss expression is,

[0103]

[0104] wherein, is the total loss, a is the concept prediction loss weight, and b is the attention sparsity loss weight;

[0105] S3.6, using the back propagation algorithm, the total loss value is minimized as the goal, the model parameters from the encoder to the classifier are iteratively updated, and the trained concept bottleneck interpretable deep learning model is obtained.

[0106] Further, the gradient of the total loss value of the concept bottleneck interpretable deep learning model to all parameters of the 3D CNN encoder, the LSTM encoder, the concept layer, the self-attention network and the linear classifier is updated using the Adam optimizer, the convolution kernel weight of the 3D CNN, the gate unit weight of the LSTM, the full connection weight of the concept layer, the query key matrix of the self-attention network and the weight matrix of the linear classifier are updated according to the gradient direction and the learning rate, the main task loss, the concept prediction loss and the attention sparsity loss are optimized synchronously in each iteration, the gradient is transmitted from the classifier to the encoder layer by layer through the back propagation chain rule, the early stopping mechanism is set, when the total loss value on the validation set does not decrease continuously for multiple epochs, the training is terminated, the parameter state of each layer at this time is saved, and the trained concept bottleneck interpretable deep learning model is obtained.

[0107] S4, the multi-dimensional data cube input into the concept bottleneck interpretable deep learning model is mapped to the semantic concept layer defined by the forest fire danger knowledge graph, and the fire danger grade result and the semantic concept layer activation result are obtained.

[0108] S4.1, the multi-dimensional data cube is batch normalized to generate a normalized multi-dimensional input tensor, the input tensor is input into the multi-modal feature encoder to obtain a normalized high-level vector, and the normalized high-level vector is input into the concept perception attention module to obtain a spatial attention weight map.

[0109] Further, the mean and variance of each feature channel in the multi-dimensional data cube are obtained, the batch normalization processing is performed to generate a normalized multi-dimensional input tensor with zero mean and unit variance, the normalized multi-dimensional input tensor is input into the trained multi-modal feature encoder, and the normalized high-level feature vector is calculated through forward propagation; the normalized high-level feature vector is input into the concept perception attention module, the number of feature channels is compressed to the same as the number of concepts through the convolution layer, and then the spatial attention weight map is generated through the Sigmoid activation function.

[0110] S4.2, the concept attention weight map is multiplied element by element with the high-level fusion feature tensor and combined with the residual connection to obtain a weighted attention perception feature tensor.

[0111] Further, the concept attention weight map is replicated and expanded in the channel dimension to the same number of channels as the high-level fusion feature tensor, achieving dimension matching. An element-wise multiplication operation is performed on the concept attention weight map and the high-level fusion feature tensor to obtain an attention-weighted feature tensor. The original high-level fusion feature tensor and the attention-weighted feature tensor are added through a residual connection, preserving the original feature information while enhancing the key region feature response. The output attention-aware feature tensor not only retains global context information but also highlights local features important for semantic concept prediction.

[0112] S4.3, the attention-aware feature tensor is flattened and input into the concept layer to obtain a semantic concept layer activation result using a Sigmoid activation function.

[0113] Specifically, the expression is,

[0114] C het = σ(W c · Flatten(F attended )+b);

[0115] where C het is the semantic concept layer activation result.

[0116] S4.4, using the classifier weight matrix and the semantic concept layer activation result, a contribution distribution matrix of the semantic concept layer activation result to fire risk grade prediction is obtained, and the concept activation vector is input into a linear classifier to obtain a fire risk grade result.

[0117] D = Softmax(W e · C het +b);

[0118] where D is the fire risk grade result, and W e is the weight matrix of the linear classifier.

[0119] It should be noted that the formula extracts the weight matrix of the linear classifier, multiplies the absolute value of the weight matrix with the semantic concept layer activation result element by element to obtain the contribution distribution matrix, inputs the semantic concept layer activation result into the linear classifier to calculate the fire risk grade probability distribution, and completes the mapping from the concept space to the risk decision.

[0120] S5, convert the rules in the forest fire risk domain knowledge graph into a differentiable constraint loss function, and use the differentiable constraint loss function to jointly optimize the training process of the concept bottleneck explainable deep learning model.

[0121] S5.1, extract rules connected by logical relationships from the forest fire risk domain knowledge graph, replace the logical operators in the rules connected by logical relationships with fuzzy logic operators, and generate a loss function template.

[0122] Further, by traversing all causal relationship edges with confidence weights in the forest fire risk domain knowledge graph, logical rules connected by leadsTo and inhibits relationships are extracted, the Boolean logical operators in the rules are replaced by corresponding differentiable fuzzy logic operators, where logical AND is replaced by element-wise multiplication operation, logical OR is replaced by probability OR operation, logical implication is replaced by ReLU difference operation, and a loss function template is constructed based on the replaced fuzzy logic expression, the confidence weights in the original rules are retained as adjustment coefficients of the loss terms, a set of rule-constrained loss function templates that can be embedded in the neural network training process is generated, and each template corresponds to a causal relationship rule in the knowledge graph.

[0123] S5.2, the probability vector of the existence of semantic concepts is substituted into the converted rule template of fuzzy logic to obtain a rule violation degree loss value.

[0124] Further, each concept probability value in the probability vector of the existence of semantic concepts is substituted into the converted rule template of fuzzy logic to obtain the violation degree of each rule. Based on leadsTo type rules (such as high temperature → drought), the difference between the premise concept probability and the conclusion concept probability is obtained by ReLU (ReLU (c high temperature

[0125] -c drought) as the violation degree, and for inhibits type rules (such as precipitation → drought), the violation degree is obtained; the final loss value of each rule is the violation degree multiplied by the confidence weight of the rule in the knowledge graph, and a rule-constrained loss value set is generated.

[0126] S5.3, according to the confidence score attached to each rule in the forest fire risk domain knowledge graph, the rule loss value is weighted and summed to obtain a knowledge constraint loss.

[0127] Further, the confidence score corresponding to each rule in the forest fire risk domain knowledge graph is obtained as a weight coefficient, and each loss term in the rule violation degree loss value set is weighted; all weighted rule violation degree loss values are added and summed to obtain a total knowledge constraint loss value; wherein the confidence score is derived from the rule confidence weight given by expert review in the knowledge graph construction process, ensuring that high-confidence rules occupy a larger proportion in the total knowledge constraint loss, and obtaining the knowledge constraint loss.

[0128] S5.4, the binary cross-entropy loss of the knowledge constraint loss, the fire risk level prediction and the probability vector of the existence of semantic concepts is added to form a joint optimization objective function.

[0129] Further, the knowledge constraint loss, the cross-entropy loss of fire risk grade prediction, and the binary cross-entropy loss of the probability vector of the existence of semantic concepts are linearly combined according to a preset weight coefficient to form a joint optimization objective function. The objective function optimizes the task prediction accuracy, the concept prediction accuracy, and the knowledge rule compliance of the model at the same time, and ensures that the concept bottleneck explainable deep learning model updates parameters under the dual constraints of data driving and knowledge driving.

[0130] S6. Deploying the concept bottleneck explainable deep learning model optimized jointly as a risk assessment service

[0131] S6.1. Generating a lightweight high-performance inference model for the trained concept bottleneck explainable deep learning model, defining standardized input and output for the RESTful and gRPC service interfaces.

[0132] Further, the concept bottleneck explainable deep learning model trained is pruned and quantized, redundant nodes in the concept layer with a contribution lower than a threshold are removed, the model weights are converted from FP32 precision to INT8 precision, the model pruned and quantized is graph optimized and kernel automatically tuned using the TensorRT framework, a lightweight high-performance inference engine is generated, the inference engine is encapsulated as a RESTful API service based on the FastAPI framework, a high-performance remote calling interface is provided through the gRPC protocol, and standardized input and output protocols are defined, wherein the input is JSON or message conforming to the multi-dimensional data cube format, and the output includes structured data of fire risk grade probability distribution, concept activation vector, and contribution matrix.

[0133] S6.2. Setting a hash key cache mechanism for the explanatory results of high-frequency queries, dynamically allocating inference tasks to edge devices according to the device computing power and network status of the request source, packaging as a Docker image, deploying services in a Kubernetes cluster, and configuring automatic scaling and load balancing strategies.

[0134] Further, a Redis cache based on the hash key of the request parameters is established for the fire risk grade results and concept explanation data of high-frequency requests, the cache expiration time is set to be dynamically adjusted, the terminal type and network delay of the request source are detected through a device detection middleware, the resource monitoring data of the edge node are combined, lightweight query tasks are allocated to ordinary computing nodes, the lightweight high-performance inference model and dependent libraries are packaged as multi-architecture Docker images, the images are deployed to a Kubernetes cluster, the Horizontal Pod Autoscaler is configured to automatically scale the number of replicas based on QPS and CPU / GPU utilization, and a weighted round-robin load balancing strategy is configured through an Ingress controller to ensure high availability and low latency response of the service.

[0135] S7. Based on the risk assessment service, receive the fire risk level result and the semantic concept layer activation result, and generate an interpretable forest fire risk assessment report.

[0136] S7.1, receive the fire risk level probability distribution, the concept probability vector, and the contribution matrix from the risk assessment service API, and parse them into structured data objects.

[0137] Further, the JSON format response data is obtained from the risk assessment service API through HTTP / REST protocol, which includes the fire risk level probability distribution array, the concept probability vector dictionary, and the contribution matrix nested dictionary; the original response data is converted into structured data objects in memory using the JSON parser deserialization tool, the fire risk level probability distribution is mapped to an ordered key-value pair set, the concept probability vector is converted to a dictionary of float values with concept names as keys, and the contribution matrix is parsed into a double nested dictionary structure, completing the standardized conversion of API response data to internal data structures.

[0138] S7.2, based on the fire risk level probability distribution, generate a fire risk probability map by mapping the spatial grid points;

[0139] Further, the spatial grid coordinate information in the multi-dimensional data cube is extracted, the highest probability value in the fire risk level probability distribution of each grid point is taken as the fire risk intensity value at that location, and the Matplotlib library is used to create a two-dimensional grid map with grid point longitude and latitude as coordinate axes and fire risk intensity value as color mapping value. Overlay OpenStreetMap or GoogleMaps as the base map, use bilinear interpolation algorithm to smooth the discrete grid points, generate continuous fire risk probability heat map; the output result is an interactive HTML map with geographic coordinate reference in PNG or WebMercator projection, where the color gradient and legend correspond to different fire risk level probability intervals.

[0140] S7.3, spatial clustering of fire risk probability map, identifying high-risk areas and sorting dominant concept contribution, getting high-risk areas.

[0141] Further, the DBSCAN spatial clustering algorithm is applied to the fire risk probability map, taking the fire risk intensity value as the density threshold parameter, to identify continuous high-risk area polygons; for all grid points within each high-risk area polygon, extract the corresponding concept contribution matrix and the average contribution of each semantic concept (such as high temperature, drought); sort the concept list in descending order of contribution average, and finally output the high-risk area list with geographical boundaries, each area associated with the dominant concept and its contribution ranking, forming a structured risk assessment result.

[0142] S7.4, combine the high-risk area with the concept contribution degree to generate an explainable forest fire risk assessment report.

[0143] Further, based on the spatial boundary information of the high-risk area and the dominant concept contribution degree sorting result, a Jinja2 template engine is used to fill in a predefined report template to integrate the geographical location of the high-risk area and the corresponding contribution degree value into a structured text description; a Matplotlib is used to generate a high-risk area heat map and a concept contribution degree bar chart, and a WeasyLib is used to render the graphic and text content to generate an explainable forest fire risk assessment report.

[0144] The embodiment also provides a computer device suitable for the case of the forest fire risk assessment method based on explainable deep learning, including a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions to realize the forest fire risk assessment method based on explainable deep learning proposed in the above embodiment.

[0145] The computer device can be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with external terminals in a wired or wireless manner. The wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device. In addition, the input device can be an external keyboard, touchpad or mouse, etc.

[0146] The embodiment also provides a storage medium on which a computer program is stored, the program being executed by a processor to implement the forest fire risk assessment method based on the interpretable deep learning proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, a magnetic disk, or an optical disk.

[0147] To sum up, the present application realizes the whole-process interpretable forest fire risk assessment from data preprocessing, model training to service deployment by: generating a multi-dimensional data cube by collecting multi-source heterogeneous data, constructing a forest fire risk domain knowledge graph, and training a concept bottleneck interpretable deep learning model, mapping input data to a semantic concept layer, outputting a fire risk grade and a concept activation result, further converting knowledge graph rules into a differentiable constraint loss function for joint optimization, deploying a risk assessment service and generating an interpretability report.

[0148] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application but not limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present application, and all of them should be covered in the scope of the claims of the present application.

Claims

1. A forest fire danger assessment method based on interpretable deep learning, characterized in that: Comprising, Collecting multi-source heterogeneous data and preprocessing to generate a multi-dimensional data cube, which contains original remote sensing images, meteorological time series data, digital elevation model data and historical forest fire point record vector data; Based on the multi-dimensional data cube, the logical relationship of forming forest fire danger is extracted by causal algorithm, and based on the logical relationship of forest fire danger, a forest fire danger domain knowledge graph is constructed; Based on the multi-dimensional data cube as a training sample, a concept bottleneck interpretable deep learning model is constructed and trained, including the following steps: Input the multi-dimensional data cube into the double-branch encoder composed of 3D CNN and LSTM, extract and fuse the spatio-temporal features to obtain a high-level fusion feature tensor; The high-level fusion feature tensor is input into the concept layer, and the probability vector of the existence of semantic concepts is calculated by the Sigmoid activation function; The probability vector of the existence of semantic concepts is input into the self-attention network, and the correlation weight between concepts is calculated to obtain a weighted concept importance vector; The weighted concept importance vector is input into the linear classifier, and the fire danger grade probability distribution is calculated using the Softmax function; Combine the real labels of fire danger grade and concepts to calculate the main task loss, concept prediction loss and attention sparsity loss, and sum to obtain the total loss value of the concept bottleneck interpretable deep learning model; Using the back propagation algorithm, the model parameters from the encoder to the classifier are iteratively updated to minimize the total loss value, and a trained concept bottleneck interpretable deep learning model is obtained; Batch normalization is performed on the multi-dimensional data cube to generate a normalized multi-dimensional input tensor, which is input into the multi-modal feature encoder to obtain a normalized high-level vector, which is input into the concept perception attention module to obtain a spatial attention weight map. Multiply the concept attention weight map and the high-level fusion feature tensor element by element and combine the residual connection to obtain the weighted attention perception feature tensor; After flattening the attention perception feature tensor, input it into the concept layer to obtain the semantic concept layer activation result using the Sigmoid activation function; Using the classifier weight matrix and the semantic concept layer activation result, the contribution distribution matrix of the semantic concept layer activation result to the fire danger grade prediction is obtained, and the concept activation vector is input into the linear classifier to obtain the fire danger grade result; Convert the rules in the forest fire danger domain knowledge graph into a differentiable constraint loss function, and use the differentiable constraint loss function to jointly optimize the training process of the concept bottleneck interpretable deep learning model; Deploy the concept bottleneck interpretable deep learning model that has been jointly optimized as a risk assessment service; Based on the risk assessment service, receive the fire danger grade result and the semantic concept layer activation result to generate an interpretable forest fire danger assessment report. 2.The forest fire danger assessment method based on interpretable deep learning according to claim 1, wherein: Collecting multi-source heterogeneous data and preprocessing to generate a multi-dimensional data cube, Including the following steps, Collect original remote sensing images, meteorological time series data, digital elevation model data and historical forest fire point record vector data from satellite data center, meteorological database and geographic information platform through API interface, and perform spatial data unified conversion to geographic coordinate system and projection unit; According to the target area and evaluation requirements, define a unified spatial resolution grid, and generate continuous surface data based on the elevation and distance collaborative kriging interpolation method based on meteorological time series data; Perform cloud detection on the original remote sensing image, and repair the cloud-covered pixels using the time series spectral similarity method to generate cloud-free data; Based on the preprocessed original remote sensing image and digital elevation model data, obtain normalized vegetation index, normalized water index, land surface temperature, slope and slope direction derived features, and perform Z-Score standardization processing to generate standardized feature data. Organize the standardized multi-source feature data according to time, row, column and channel dimensions to generate a multi-dimensional data cube. 3.The forest fire danger assessment method based on interpretable deep learning according to claim 2, wherein: Based on the multi-dimensional data cube, extract the logical relationship of forming forest fire danger, and based on the logical relationship of forest fire danger, construct a forest fire danger domain knowledge graph, including the following steps, Based on the multi-dimensional data cube, define the entity type and relationship type of the knowledge graph, input the multi-variable time series data in the multi-dimensional data cube into the PC algorithm to calculate the conditional independence between variables, and obtain the initial causal graph driven by data; Review the data-driven initial causal graph to obtain a set of causal relationships with weights, and define a quantitative mapping rule for each entity concept in the knowledge graph based on the multi-dimensional data cube; Use a graph database to convert the set of causal relationships with weights into a graph structure, with nodes as entities and edges as relationships, and set the confidence score as an attribute of the edge, to construct a forest fire danger domain knowledge graph. 4.The forest fire danger assessment method based on interpretable deep learning according to claim 3, wherein: Convert the rules in the forest fire danger domain knowledge graph into a differentiable constraint loss function, and jointly optimize the training process of the concept bottleneck explainable deep learning model using the differentiable constraint loss function, including the following steps, Extract rules connected by logical relationships from the forest fire danger domain knowledge graph, replace the logical operators in the rules connected by logical relationships with fuzzy logic operators to generate a loss function template; Substitute the probability vector of semantic concepts into the rule template converted by fuzzy logic to obtain a rule loss value; According to the confidence score attached to each rule in the forest fire danger domain knowledge graph, weight the rule loss value to obtain a knowledge constraint loss; Add the binary cross-entropy loss of the knowledge constraint loss, the fire danger level prediction and the probability vector of the existence of semantic concepts to form a joint optimization objective function. 5.The forest fire danger assessment method based on interpretable deep learning according to claim 1, wherein: Deploy the concept bottleneck explainable deep learning model optimized by joint optimization as a risk assessment service, Including the following steps, Generate a lightweight high-performance inference model from the trained concept bottleneck explainable deep learning model, and define standardized input and output for the lightweight high-performance inference model as RESTful and gRPC service interfaces. The interpretive result of the high-frequency query sets a hash key cache mechanism, dynamically allocates inference tasks to edge devices according to the device computing power and network state of the request source, packages them into Docker images, deploys services in a Kubernetes cluster, and configures automatic scaling and load balancing strategies. 6.The forest fire danger assessment method based on interpretable deep learning according to claim 5, wherein: Based on the risk assessment service, the fire risk level result and the semantic concept layer activation result are received to generate an interpretable forest fire risk assessment report, The method comprises the following steps: Receiving the fire risk level probability distribution, the concept probability vector and the contribution matrix from the risk assessment service API, and analyzing them into structured data objects; Based on the fire risk level probability distribution, a fire risk probability map is generated by mapping through spatial grid points; Spatial clustering is performed on the fire risk probability map to identify high-risk areas and sort the dominant concept contribution, thereby obtaining the high-risk areas; The high-risk areas and the concept contribution are combined to generate an interpretable forest fire risk assessment report. 7.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is characterized in that: The processor executes the computer program to implement the steps of the forest fire risk assessment method based on interpretable deep learning according to any one of claims 1-6.

8. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by the processor to implement the steps of the forest fire risk assessment method based on interpretable deep learning according to any one of claims 1-6.

Images