Intelligent forestry safety management system based on big data

Abstract

The invention discloses an intelligent forestry safety management system based on big data, and the system comprises the steps: obtaining a pest image of a to-be-detected region, carrying out the data preprocessing of the pest image, obtaining a data set, inputting the data set into a preset biological recognition model, and carrying out the feature extraction to obtain a pest recognition result, inputting the pest recognition result into an improved YOLOv4-based forestry pest detection model for detection and analysis to obtain a pest classification result, constructing a forestry ecological safety index of the to-be-detected area according to the pest classification result, performing data management on the forestry ecological safety index, increasing the number of training samples by using data enhancement, and obtaining a forestry ecological safety index of the to-be-detected area. The method greatly improves the generalization ability of the model, improves the richness of feature information, combines the forest ecological safety index, the wetland safety index and the desert ecological safety index to construct the forestry ecological safety index, and improves the rationality and effectiveness of forestry safety management.

Description

Technical Field

[0001] The present invention belongs to the technical field of forestry ecological data processing, and in particular relates to a smart forestry safety management system based on big data. Background Art

[0002] Forest pest control is an important part of forestry prevention and control work, and it shoulders the important mission of protecting forest resources, maintaining ecological security, and promoting ecological civilization. However, with the increase in forest greening area, it is restricted by many factors, and pests and diseases have also become more common, increasing the difficulty of pest control. Forest biological disasters caused by forest pests are one of the major natural disasters in forestry. They not only destroy forest resources and threaten the safety of the three major ecosystems of forests, wetlands, and deserts, but also affect economic and trade security and forest product safety, weakening the role of forestry in ecological and people's livelihood construction. Since forest pest detection initially adopted manual statistical methods, the investment cost was huge, the detection results were highly dependent on expert professional knowledge, and the detection efficiency was low. Therefore, it is urgent to provide a smart forestry safety management system based on big data to solve the above-mentioned technical problems. Summary of the invention

[0003] In view of this, the present invention provides a smart forestry safety management system based on big data, which is specifically implemented by the following technical solutions.

[0004] The present invention provides a smart forestry safety management system based on big data, comprising:

[0005] A data acquisition unit, used to acquire images of harmful organisms in the area to be tested, and perform data preprocessing on the images of harmful organisms to obtain a data set;

[0006] A biological intelligent identification unit, used for inputting the data set into a preset biological identification model for feature extraction to obtain a harmful biological identification result, wherein the preset biological identification model includes a ResNet network, the ResNet network includes a plurality of residual modules, convolutional layers and short-circuit connection paths, the down-sampled residual module is composed of three convolutional layers and a short-circuit connection path with one convolutional layer, the three convolutional layers include two 1×1 convolutional layers and one 3×3 convolutional layer, and the short-circuit connection path includes a 1×1 convolution for downsampling and dimensionality control;

[0007] A forestry pest detection unit, used for inputting the harmful organism identification result into a forestry pest detection model based on improved YOLOv4 to perform detection and analysis to obtain a pest classification result;

[0008] The forestry ecological safety management unit is used to construct the forestry ecological safety index of the area to be tested according to the pest classification result, and to perform data management on the forestry ecological safety index.

[0009] As a preferred embodiment of the above scheme, the image data is expanded by data enhancement methods such as horizontal mirroring, brightness adjustment, Gaussian noise and Gaussian blur, specifically including:

[0010] The image color space is converted from RGB to HSV to perceive the brightness change of the image, and the brightness of the image is adjusted. The dark image is compensated for the brightness by increasing the brightness by 0.3, and the bright image is balanced by reducing the brightness by 0.3. The Gaussian distribution random number with a mean of 0, a standard deviation of 1, and a number of 3 is calculated and superimposed on the original image pixel value. The image is Gaussian processed, and each pixel of the image is weighted summed with its surrounding pixels. The length and width of the Gaussian matrix are set to 0, the standard deviation is set to 2, and the image is Gaussian blurred.

[0011] As a preferred embodiment of the above scheme, the data set is input into a preset biometric recognition model to extract features to obtain harmful organism recognition results, including:

[0012] The residual module of the ResNet network includes a semantic grouping layer and a group bilinear layer. The semantic grouping layer groups the semantic information distributed in different areas of the image by channel, and iteratively optimizes the channels of the same or different semantic groups to classify the semantic information of the image. The corresponding expression is:

[0013]

[0014] Among them, formula (1) is the calculation formula of the semantic mapping matrix, m represents the feature map of each channel, N represents the number of feature maps, A represents the semantic mapping matrix, which is the final calculation result of semantic grouping and needs to be iterated. Formulas (2) and (3) are the calculation formulas of the semantic grouping loss function, which are used to optimize the mapping matrix A, L g represents the semantic grouping loss function, i and j represent the horizontal and vertical coordinates respectively, and G represents the number of groups;

[0015] After semantic grouping, the group bilinear layer further performs bilinear transformation on the feature vector of each channel and each feature channel in each group. By calculating the outer product, the feature channels are made to interact in pairs, thus enhancing the representation of the corresponding semantics. The corresponding expression is:

[0016]

[0017] Among them, y represents the output feature, x represents the input feature, T B represents the group bilinear transformation, A is the mapping matrix calculated in formula (1), G represents the number of groups, j represents the number of groups, and I j Indicates the matrix of the group, vec means vectorization;

[0018] The bilinear change results obtained by pairwise interaction are summed up and vectorized to restore the original convolution feature dimension.

[0019] As a preferred embodiment of the above scheme, the evaluation criteria of the preset biometric recognition model use accuracy, recall, and F1 value as evaluation indicators, and the corresponding expressions are:

[0020]

[0021] Among them, TP represents the number of forest pests correctly identified, FP represents the number of forest pests mistakenly identified as a certain forest pest, FN represents the number of forest pests mistakenly identified as other types of forest pests, and the F1 value is the harmonic mean based on precision and recall. The average precision, average recall and average F1 value are the average precision, recall and F1 values ​​of all pest types, respectively.

[0022] As a preferred embodiment of the above scheme, the harmful organism identification result is input into a forestry pest detection model based on an improved TOLOv4 for detection and analysis to obtain a pest classification result, including:

[0023] The ECA-CBAM attention mechanism is integrated with the backbone network CSPDarknet53, and SPPF is used to replace SPP and build SPPF-PANet to fuse the feature information captured by multiple receptive fields. Among them, ECA-CBAM is a convolutional block attention module, including CAM structure and SAM structure. The fully connected layer used in the CAM structure is to capture nonlinear cross-channel interactions and involves dimensionality reduction to control the complexity of the model. After the channel-level global average pooling without reducing the dimensionality, the ECA module uses a one-dimensional convolution of size k and uses each channel machine k-1 neighbors to capture local cross-channel interaction information.

[0024] K represents the number of adjacent channels involved in the prediction of a channel attention weight. k is adaptively determined by mapping the channel dimension C. The corresponding expression is:

[0025]

[0026] Where |t| odd is the odd number closest to t, and γ and b are 2 and 1 respectively.

[0027] As a preferred embodiment of the above scheme, the ECA-CBAM attention mechanism performs channel attention weighting and spatial attention weighting on the feature graph inscription corresponding to the harmful organism identification result, and the corresponding expression is:

[0028]

[0029] Among them, F is the input feature map of ECA-CBAM, M c (·) is the ECA channel attention weight calculation formula, M s (·) is the SAM spatial attention weight calculation formula, is element-wise multiplication;

[0030] M c (·) and M s (·)The calculation method is:

[0031] M c (F) = σ(f 1×k (AvgPool(F))) (11)

[0032] M s (F) = σ(f 3×3 ([AvgPool(F)];MaxPool(F))) (11)

[0033] Among them, σ(·) represents the Sigmoid activation function, f a×b (·) represents a convolution operation with a kernel size of a×b, AvgPool(·) represents an average pooling operation, and MaxPool(·) represents a maximum pooling operation.

[0034] As a preferred embodiment of the above scheme, a deep separable convolution DSC is used to separate the mapping of cross-channel correlation and spatial correlation in the feature map of the convolutional neural network. The DSC method includes executing a deep convolution kernel to execute a point convolution. The deep convolution includes applying each filter to a single input channel at a time. The point convolution includes performing a convolution operation using a 1×1 filter to cover all channels.

[0035] If the network input D F ×D F ×M feature map, the convolution kernel size is D K ×D K ×M, when the output feature map size is D P ×D P ×N, the computational complexity of conventional convolution is (D P ×D P ×N)×D K ×D K ×M, the first step of DSC is to perform a depthwise convolution by using a convolutional layer of size D K ×D K ×1 convolution kernel convolves the input features of each layer. The network computation amount at this stage is (D K ×D K ×1)×M×D K ×D K, the feature map depth remains unchanged; the second step is the point convolution operation, the convolution kernel size is 1×1×M, and the output feature map size is D P ×D P ×N, the amount of calculation is D K ×D K ×N×(1×1×M); then the DSC calculation amount is equal to

[0036] As a preferred option of the above scheme, the ecological security status index and pressure index of the forestry ecosystem are calculated, and the forestry ecological security index is calculated based on the pressure index and status. The forestry ecological security index is obtained from the security indicators of forest, desert and wetland ecosystems, including:

[0037] S1: Standardize the collected data. The processing indicators are divided into positive correlation indicators and negative correlation indicators. The calculation of positive correlation indicators is as shown in formula (12), and the calculation of negative correlation indicators is as shown in formula (13), where y ij represents the standardized data of the jth indicator of the i-th sample, x ij represents the value of the jth index of the i-th sample, x max represents its maximum value, x min Indicates its minimum value:

[0038]

[0039] S2: Calculate the information utility value of the data. After obtaining the standardized data, the utility value of the data can be calculated. Y ij It represents the ratio of the standardized data of the jth indicator of the i-th sample to the standardized data of all indicators of the i-th sample. i is a positive integer from 1 to n. The information utility value expression of the corresponding indicator is:

[0040]

[0041] S3: Determine the weight w of forest ecological security status indicators j , the weight calculation expression of the state index of the sample is (14), where s represents the number of state indicators:

[0042]

[0043] S4: Modify the standardized data obtained in S1. The modification is as shown in formula (15). When the jth index of the sample is missing, it is replaced by 0. If the sample index data is not missing, the standardized data calculated by S1 is used:

[0044]

[0045] S5: The forestry ecological security index needs to use the data obtained in the first four steps, namely the weight of the state index and the standardized data and information utility value obtained after correction. The calculation is as shown in formula (16). s has the same meaning as in S3, and w j is the weight calculated by S3, y′ ij Corrected normalized data calculated for S4:

[0046]

[0047] As a preferred solution of the above scheme, after obtaining the state index and pressure index of forest ecological security, the forestry security index can be calculated. The forest ecological security index of the tth sample is expressed as IF t The pressure index of the tth sample is represented by Y Ft Indicates that the state index of the tth sample is Z Ft It is indicated that the safety factor of desert and wetland ecosystems can be calculated based on the calculation of forest ecological security index:

[0048]

[0049] As a preferred embodiment of the above scheme, the forestry ecological security index is related to the forest ecological security index, the desert ecological security index and the ecological security index, and its calculation is as shown in formula (18), I t represents the forestry ecological security index, I Ft represents the forest ecological security index of the tth sample, I wt represents the wetland ecological security index of the tth sample, I Dt represents the desert ecological security index of the tth sample, d1, d2 and d3 represent the proportion of forest area, wetland area and desert area to the total area of ​​these three categories respectively:

[0050] I t =d1I Ft +d2I wt +d3I Dt (18).

[0051] The present invention provides a smart forestry safety management system based on big data, which obtains harmful organism images in a test area, performs data preprocessing on the harmful organism images to obtain a data set, inputs the data set into a preset biometric recognition model for feature extraction to obtain a harmful organism recognition result, inputs the harmful organism recognition result into a forestry pest detection model based on an improved YOLOv4 for detection and analysis to obtain a pest classification result, constructs a forestry ecological safety index for the test area according to the pest classification result, performs data management on the forestry ecological safety index, uses data enhancement to expand the number of training samples, greatly improves the generalization ability of the model, and improves the richness of feature information, and constructs a forestry ecological safety index in combination with a forest ecological safety index, a wetland safety index and a desert ecological safety index, thereby improving the rationality and effectiveness of forestry safety management. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0053] Figure 1 This is a structural diagram of the smart forestry safety management system based on big data of the present invention;

[0054] Figure 2 This is a diagram of the calculation process of the forestry ecological security index of the present invention. DETAILED DESCRIPTION

[0055] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0056] See also Figure 1 The present invention provides a smart forestry safety management system based on big data, including:

[0057] A data acquisition unit, used to acquire images of harmful organisms in the area to be tested, and perform data preprocessing on the images of harmful organisms to obtain a data set;

[0058] A biological intelligent identification unit, used for inputting the data set into a preset biological identification model for feature extraction to obtain a harmful biological identification result, wherein the preset biological identification model includes a ResNet network, the ResNet network includes a plurality of residual modules, convolutional layers and short-circuit connection paths, the down-sampled residual module is composed of three convolutional layers and a short-circuit connection path with one convolutional layer, the three convolutional layers include two 1×1 convolutional layers and one 3×3 convolutional layer, and the short-circuit connection path includes a 1×1 convolution for downsampling and dimensionality control;

[0059] A forestry pest detection unit, used for inputting the harmful organism identification result into a forestry pest detection model based on improved YOLOv4 to perform detection and analysis to obtain a pest classification result;

[0060] The forestry ecological safety management unit is used to construct the forestry ecological safety index of the area to be tested according to the pest classification result, and to perform data management on the forestry ecological safety index.

[0061] In this embodiment, the image data is expanded by data enhancement methods such as horizontal mirroring, brightness adjustment, Gaussian noise and Gaussian blur, specifically including: converting the image color space from RGB to HSV to perceive the brightness change of the image, adjusting the brightness of the image, compensating the brightness of the overly dark image by increasing the brightness by 0.3, balancing the brightness of the overly bright image by reducing the brightness by 0.3, calculating a Gaussian distributed random number with a mean of 0, a standard deviation of 1 and a number of 3, superimposing it on the original image pixel value, performing Gaussian processing on the image, performing a weighted sum on each pixel of the image and its surrounding pixels, setting the length and width of the Gaussian matrix to 0, the standard deviation to 2, and performing Gaussian blur processing on the image. Due to the wide distribution of forest pests and complex backgrounds, the same type of forest pests are affected differently by seasonal climate. In order to reduce the impact of the environment on image features, redundant background information is removed by cropping. The improvement of deep learning model recognition effect requires the support of a large number of data sets. However, the current collection of samples of some forest pests is difficult, resulting in a small sample size. At the same time, different shooting angles and light also affect the recognition accuracy of the model. In order to increase the sample size of some forest pests and simulate more actual usage scenarios, data enhancement is used to expand the data set.

[0062] It should be noted that the dataset includes training set, validation set and test set. The training set generally has the largest number of images and is used for learning model parameters. The validation set is used to determine the model structure and fine-tune the experimental parameters. The test set is used to evaluate the detection effect of the model. The data structure of each set is generally image data and its corresponding label file. Image translation generates new images by randomly moving image pixels in the horizontal or vertical direction of the image. Image translation can improve the diversity of the dataset and make the deep learning model have better robustness and generalization ability. In the process of implementing the translation operation, first, read the original image to obtain the width and height information of the image, then randomly determine the offset and translation direction according to the predefined offset range and the translation direction, and then translate the image pixels along the translation direction. The translation operation will cause the loss of image edge pixel information. It is necessary to fill the image edge after translation to avoid changing the image size. Finally, the image obtained by the operation is saved as a new image for subsequent use. Gaussian noise generates new images by adding random noise that obeys Gaussian distribution to the image to increase the diversity of the dataset and provide the generalization ability of the deep learning model. In the process of adding Gaussian noise to an image, first, read the original image and obtain the width and height information. Then, generate a random noise image that obeys the Gaussian distribution. The pixel values ​​of the image are limited to between 0 and 255. Finally, superimpose the random noise image and the original image according to a certain weight to generate a new image with Gaussian noise.

[0063] It should be understood that the construction is carried out by calculating the pressure index and the state index. The state index indicates the quality of forestry ecological security. The larger the index, the better the ecological security; the pressure index indicates the pressure on the ecosystem. The larger the index, the greater the pressure on the system. The forestry safety index is constructed on the basis of the forest ecological safety index, the wetland ecological safety index and the desert ecological safety index. The larger the forestry ecological safety index, the healthier the ecological environment. The harmful biological images of the test area are obtained, and the harmful biological images are preprocessed to obtain a data set. The data set is input into the preset biometric recognition model for feature extraction to obtain the harmful biological recognition result. The harmful biological recognition result is input into the forestry pest detection model based on the improved YOLOv4 for detection and analysis to obtain the pest classification result. The forestry ecological safety index of the test area is constructed according to the pest classification result, and the forestry ecological safety index is managed. The number of training samples is expanded by using data enhancement, which greatly improves the generalization ability of the model and the richness of feature information. The forestry ecological safety index is constructed by combining the forest ecological safety index, the wetland safety index and the desert ecological safety index to improve the rationality and effectiveness of forestry safety management.

[0064] Optionally, the data set is input into a preset biometric recognition model to extract features to obtain harmful organism recognition results, including:

[0065] The residual module of the ResNet network includes a semantic grouping layer and a group bilinear layer. The semantic grouping layer groups the semantic information distributed in different areas of the image by channel, and iteratively optimizes the channels of the same or different semantic groups to classify the semantic information of the image. The corresponding expression is:

[0066]

[0067] Among them, formula (1) is the calculation formula of the semantic mapping matrix, m represents the feature map of each channel, N represents the number of feature maps, A represents the semantic mapping matrix, which is the final calculation result of semantic grouping and needs to be iterated. Formulas (2) and (3) are the calculation formulas of the semantic grouping loss function, which are used to optimize the mapping matrix A, L g represents the semantic grouping loss function, i and j represent the horizontal and vertical coordinates respectively, and G represents the number of groups;

[0068] After semantic grouping, the group bilinear layer further performs bilinear transformation on the feature vector of each channel and each feature channel in each group. By calculating the outer product, the feature channels are made to interact in pairs, thus enhancing the representation of the corresponding semantics. The corresponding expression is:

[0069]

[0070] Among them, y represents the output feature, x represents the input feature, T B represents the group bilinear transformation, A is the mapping matrix calculated in formula (1), G represents the number of groups, j represents the number of groups, and I j Indicates the matrix of the group, vec means vectorization;

[0071] The bilinear change results obtained by pairwise interaction are summed up and vectorized to restore the original convolution feature dimension.

[0072] In this embodiment, the evaluation criteria of the preset biometric recognition model use accuracy, recall, and F1 value as evaluation indicators, and the corresponding expressions are:

[0073]

[0074] Among them, TP represents the number of forest pests correctly identified, FP represents the number of forest pests mistakenly identified as a certain forest pest, FN represents the number of forest pests mistakenly identified as other types of forest pests, and the F1 value is the harmonic mean based on precision and recall. The average precision, average recall and average F1 value are the average precision, recall and F1 values ​​of all pest types, respectively.

[0075] It should be noted that the pest identification result is input into the forestry pest detection model based on the improved TOLOv4 for detection and analysis to obtain the pest classification result, including: fusing the ECA-CBAM attention mechanism and the backbone network CSPDarknet53, replacing SPP with SPPF and constructing SPPF-PANet to fuse the feature information captured by multiple receptive fields, wherein ECA-CBAM is a convolutional block attention module, including a CAM structure and a SAM structure. The fully connected layer used in the CAM structure is to capture nonlinear cross-channel interactions and involves dimensionality reduction to control the complexity of the model; after the channel-level global average pooling without reducing the dimensionality, the ECA module uses a one-dimensional convolution of size k, and uses each channel machine k-1 neighbors to capture local cross-channel interaction information; K represents the number of adjacent channels involved in the prediction of a channel attention weight, and k is adaptively determined by mapping the channel dimension C, and the corresponding expression is:

[0076]

[0077] Where |t| odd is the odd number closest to t, γ and b are 2 and 1 respectively. The ResNet network alleviates the problem of deep learning network degradation as the depth increases to a certain extent, but due to the high complexity of the deep learning network, the network still encounters the problem of large amount of calculation in the application process. The deep bilinear transformation module is embedded in the residual module of the RestNet network, which improves the fine-grained feature extraction capability of the network without increasing the amount of calculation.

[0078] Optionally, the ECA-CBAM attention mechanism performs channel attention weighting and spatial attention weighting on the feature graph inscription corresponding to the harmful organism identification result, and the corresponding expression is:

[0079]

[0080] Among them, F is the input feature map of ECA-CBAM, M c (·) is the ECA channel attention weight calculation formula, M s (·) is the SAM spatial attention weight calculation formula, is element-wise multiplication;

[0081] M c (·) and M s (·)The calculation method is:

[0082] M c (F) = σ(f 1×k (AvgPool(F))) (11)

[0083] Ms (F) = σ(f 3×3 ([AvgPool(F)];MaxPool(F))) (11)

[0084] Where σ(·) represents the Sigmoid activation function, f a×b (·) represents a convolution operation with a kernel size of a×b, AvgPool(·) represents an average pooling operation, and MaxPool(·) represents a maximum pooling operation.

[0085] It should be noted that the mapping of cross-channel correlation and spatial correlation in the feature map of the convolutional neural network is separated by using the deep separable convolution DSC. The DSC process includes performing point convolution with a deep convolution kernel. The deep convolution process includes applying each filter to a single input channel at a time. The point convolution process includes performing convolution operations with a 1×1 filter covering all channels.

[0086] If the network input D F ×D F ×M feature map, the convolution kernel size is D K ×D K ×M, when the output feature map size is D P ×D P ×N, the computational complexity of conventional convolution is (D P ×D P ×N)×D K ×D K ×M, the first step of DSC is to perform a depthwise convolution by using a convolutional layer of size D K ×D K ×1 convolution kernel convolves the input features of each layer. The network computation amount at this stage is (D K ×D K ×1)×M×D K ×D K , the feature map depth remains unchanged; the second step is the point convolution operation, the convolution kernel size is 1×1×M, and the output feature map size is D P ×D P ×N, the amount of calculation is D K ×D K ×N×(1×1×M); then the DSC calculation amount is equal to

[0087] Among them, the deep separable convolution DSC replaces the ordinary convolution through deep convolution and point convolution, which greatly reduces the number of parameters and theoretical calculations. At the same time, it improves the feature extraction ability of the network by increasing the channel dimension of the output feature map, and achieves a good balance between accuracy and computational cost. The SSP module is a commonly used deep learning module that can be used for tasks such as image classification and target detection. The main function of the SPP module is to pool the input feature map at different scales, splice multiple pooling results together, and then perform subsequent processing. The advantage of this is that local features and global features can be obtained at the same time, thereby improving the receptive field and recognition ability of the model.

[0088] See also Figure 2 , calculate the ecological security status indicators and pressure index of the forestry ecosystem, and calculate the forestry ecological security index based on the pressure index and status. The forestry ecological security index is obtained from the security indicators of forest, desert and wetland ecosystems, including:

[0089] S1: Standardize the collected data. The processing indicators are divided into positive correlation indicators and negative correlation indicators. The calculation of positive correlation indicators is as shown in formula (12), and the calculation of negative correlation indicators is as shown in formula (13), where y ij represents the standardized data of the jth indicator of the i-th sample, x ij represents the value of the jth index of the i-th sample, x max represents its maximum value, x min Indicates its minimum value:

[0090]

[0091] S2: Calculate the information utility value of the data. After obtaining the standardized data, the utility value of the data can be calculated. Y ij It represents the ratio of the standardized data of the jth indicator of the i-th sample to the standardized data of all indicators of the i-th sample. i is a positive integer from 1 to n. The information utility value expression of the corresponding indicator is:

[0092]

[0093] S3: Determine the weight w of forest ecological security status indicators j , the weight calculation expression of the state index of the sample is (14), where s represents the number of state indicators:

[0094]

[0095] S4: Modify the standardized data obtained in S1. The modification is as shown in formula (15). When the jth index of the sample is missing, it is replaced by 0. If the sample index data is not missing, the standardized data calculated by S1 is used:

[0096]

[0097] S5: The forestry ecological security index needs to use the data obtained in the first four steps, namely the weight of the state index and the standardized data and information utility value obtained after correction. The calculation is as shown in formula (16). s has the same meaning as in S3, and w j is the weight calculated by S3, y′ ij Corrected normalized data calculated for S4:

[0098]

[0099] In this embodiment, after obtaining the state index and pressure index of the forest ecological security, the forestry security index can be calculated. The forest ecological security index of the tth sample is expressed as I Ft Indicates that the pressure index of the tth sample is expressed as Y Ft Indicates that the state index of the tth sample is Z Ft It is indicated that the safety factor of desert and wetland ecosystems can be calculated based on the calculation of forest ecological security index:

[0100]

[0101] It should be noted that the forestry ecological security index is related to the forest ecological security index, desert ecological security index and ecological security index. Its calculation is as shown in formula (18): I t represents the forestry ecological security index, I Ft represents the forest ecological security index of the tth sample, I wt represents the wetland ecological security index of the tth sample, I Dt I represents the desert ecological security index of the tth sample, d1, d2 and d3 represent the proportion of forest area, wetland area and desert area to the total area of ​​these three categories respectively: t =d1I Ft +d2I wt +d3I Dt (18).

[0102] In all examples shown and described herein, any specific values ​​should be interpreted as merely exemplary and not as limiting, and thus other examples of the exemplary embodiments may have different values.

[0103] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0104] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. It should be pointed out that for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A smart forestry safety management system based on big data, characterized in that: include: A data acquisition unit, used to acquire images of harmful organisms in the area to be tested, and perform data preprocessing on the images of harmful organisms to obtain a data set; A biological intelligent identification unit, used for inputting the data set into a preset biological identification model for feature extraction to obtain a harmful biological identification result, wherein the preset biological identification model includes a ResNet network, the ResNet network includes a plurality of residual modules, convolutional layers and short-circuit connection paths, the down-sampled residual module is composed of three convolutional layers and a short-circuit connection path with one convolutional layer, the three convolutional layers include two 1×1 convolutional layers and one 3×3 convolutional layer, and the short-circuit connection path includes a 1×1 convolution for downsampling and dimensionality control; A forestry pest detection unit, used for inputting the harmful organism identification result into a forestry pest detection model based on improved YOLOv4 to perform detection and analysis to obtain a pest classification result; The forestry ecological safety management unit is used to construct the forestry ecological safety index of the tested area according to the pest classification result, and perform data management on the forestry ecological safety index.

2. The smart forestry safety management system based on big data according to claim 1 is characterized in that: The image data is expanded through data enhancement methods such as horizontal mirroring, brightness adjustment, Gaussian noise and Gaussian blur, including: The image color space is converted from RGB to HSV to perceive the brightness change of the image, and the brightness of the image is adjusted. The dark image is compensated for the brightness by increasing the brightness by 0.3, and the bright image is balanced by reducing the brightness by 0.

3. The Gaussian distribution random number with a mean of 0, a standard deviation of 1, and a number of 3 is calculated and superimposed on the original image pixel value. The image is Gaussian processed, and each pixel of the image is weighted summed with its surrounding pixels. The length and width of the Gaussian matrix are set to 0, the standard deviation is set to 2, and the image is Gaussian blurred.

3. The smart forestry safety management system based on big data according to claim 1 is characterized in that: The data set is input into a preset biometric recognition model to extract features to obtain harmful organism recognition results, including: The residual module of the ResNet network includes a semantic grouping layer and a group bilinear layer. The semantic grouping layer groups the semantic information distributed in different areas of the image by channel, and iteratively optimizes the channels of the same or different semantic groups to classify the semantic information of the image. The corresponding expression is: Among them, formula (1) is the calculation formula of the semantic mapping matrix, m represents the feature map of each channel, N represents the number of feature maps, A represents the semantic mapping matrix, which is the final calculation result of semantic grouping and needs to be iterated. Formulas (2) and (3) are the calculation formulas of the semantic grouping loss function, which are used to optimize the mapping matrix A, L g represents the semantic grouping loss function, i and j represent the horizontal and vertical coordinates respectively, and G represents the number of groups; After semantic grouping, the group bilinear layer further performs bilinear transformation on the feature vector of each channel and each feature channel in each group. By calculating the outer product, the feature channels are made to interact in pairs, thus enhancing the representation of the corresponding semantics. The corresponding expression is: Among them, y represents the output feature, x represents the input feature, T B represents the group bilinear transformation, A is the mapping matrix calculated in formula (1), G represents the number of groups, j represents the number of groups, and I j Indicates the matrix of the group, vec means vectorization; The bilinear change results obtained by pairwise interaction are summed up and vectorized to restore the original convolution feature dimension.

4. The smart forestry safety management system based on big data according to claim 3 is characterized in that: The evaluation criteria of the preset biometric recognition model use accuracy, recall, and F1 value as evaluation indicators, and the corresponding expressions are: Among them, TP represents the number of forest pests correctly identified, FP represents the number of forest pests mistakenly identified as a certain forest pest, FN represents the number of forest pests mistakenly identified as other types of forest pests, and the F1 value is the harmonic mean based on precision and recall. The average precision, average recall and average F1 value are the average precision, recall and F1 values ​​of all pest types, respectively.

5. The smart forestry safety management system based on big data according to claim 1 is characterized in that: The harmful organism identification results are input into the forestry pest detection model based on the improved TOLOv4 for detection and analysis to obtain pest classification results, including: The ECA-CBAM attention mechanism is integrated with the backbone network CSPDarknet53, and SPPF is used to replace SPP and build SPPF-PANet to fuse the feature information captured by multiple receptive fields. Among them, ECA-CBAM is a convolutional block attention module, including CAM structure and SAM structure. The fully connected layer used in the CAM structure is to capture nonlinear cross-channel interactions and involves dimensionality reduction to control the complexity of the model. After the channel-level global average pooling without reducing the dimensionality, the ECA module uses a one-dimensional convolution of size k and uses each channel machine k-1 neighbors to capture local cross-channel interaction information. K represents the number of adjacent channels involved in the prediction of a channel attention weight. k is adaptively determined by mapping the channel dimension C. The corresponding expression is: Where |t| odd is the odd number closest to t, and γ and b are 2 and 1 respectively.

6. The smart forestry safety management system based on big data according to claim 5 is characterized in that: The ECA-CBAM attention mechanism performs channel attention weighting and spatial attention weighting on the feature graph inscription corresponding to the pest identification result, and the corresponding expression is: Among them, F is the input feature map of ECA-CBAM, M c (·) is the ECA channel attention weight calculation formula, M s (·) is the SAM spatial attention weight calculation formula, is element-wise multiplication; M c (·) and M s (·)The calculation method is: M c (F)=σ(f 1×k (AvgPool(F))) (11) M s (F)=σ(f 3×3 ([AvgPool(F)];MaxPool(F))) (11) Among them, σ(·) represents the Sigmoid activation function, f a×b (·) represents a convolution operation with a kernel size of a×b, AvgPool(·) represents an average pooling operation, and MaxPool(·) represents a maximum pooling operation.

7. The smart forestry safety management system based on big data according to claim 6 is characterized in that: A deep separable convolution (DSC) is used to separate the mapping of cross-channel correlation and spatial correlation in the feature map of the convolutional neural network. The DSC method includes performing point convolution with a deep convolution kernel. The deep convolution includes applying each filter to a single input channel at a time. The point convolution includes performing a convolution operation with a 1×1 filter covering all channels. If the network input D F ×D F ×M feature map, the convolution kernel size is D K ×D K ×M, when the output feature map size is D P ×D P ×N, the computational complexity of conventional convolution is (D P ×D P ×N)×D K ×D K ×M, the first step of DSC is to perform a depthwise convolution by using a convolutional layer of size D K ×D K ×1 convolution kernel convolves the input features of each layer. The network computation amount at this stage is (D K ×D K ×1)×M×D K ×D K , the feature map depth remains unchanged; The second step is the point convolution operation, the convolution kernel size is 1×1×M, and the output feature map size is D P ×D P ×N, the amount of calculation is D K ×D K ×N×(1×1×M); then the DSC calculation amount is equal to 8. The smart forestry safety management system based on big data according to claim 1 is characterized in that: Calculate the ecological security status indicators and pressure index of the forestry ecosystem, and calculate the forestry ecological security index based on the pressure index and status. The forestry ecological security index is derived from the security indicators of forest, desert and wetland ecosystems, including: S1: Standardize the collected data. The processing indicators are divided into positive correlation indicators and negative correlation indicators. The calculation of positive correlation indicators is as shown in formula (12), and the calculation of negative correlation indicators is as shown in formula (13), where y ij represents the standardized data of the jth indicator of the i-th sample, x ij represents the value of the jth index of the i-th sample, x max represents its maximum value, x min Indicates its minimum value: S2: Calculate the information utility value of the data. After obtaining the standardized data, the utility value of the data can be calculated. Y ij It represents the ratio of the standardized data of the jth indicator of the i-th sample to the standardized data of all indicators of the i-th sample. i is a positive integer from 1 to n. The information utility value expression of the corresponding indicator is: S3: Determine the weight w of forest ecological security status indicators j , the weight calculation expression of the state index of the sample is (14), where s represents the number of state indicators: S4: Modify the standardized data obtained in S1. The modification is as shown in formula (15). When the jth index of the sample is missing, it is replaced by 0. If the sample index data is not missing, the standardized data calculated by S1 is used: S5: The forestry ecological security index needs to use the data obtained in the first four steps, namely the weight of the state index and the standardized data and information utility value obtained after correction. The calculation is as shown in formula (16). s has the same meaning as in S3, and w j is the weight calculated by S3, y′ ij Corrected normalized data calculated for S4:

9. The smart forestry safety management system based on big data according to claim 8 is characterized in that: Also includes: After obtaining the state index and pressure index of forest ecological security, the forestry security index can be calculated. The forest ecological security index of the tth sample is expressed as I Ft Indicates that the pressure index of the tth sample is expressed as Y Ft Indicates that the state index of the tth sample is Z Ft It is indicated that the safety factor of desert and wetland ecosystems can be calculated based on the calculation of forest ecological security index:

10. The smart forestry safety management system based on big data according to claim 9 is characterized in that: The forestry ecological security index is related to the forest ecological security index, desert ecological security index and ecological security index. Its calculation is as shown in formula (18): I t represents the forestry ecological security index, I Ft represents the forest ecological security index of the tth sample, I wt represents the wetland ecological security index of the tth sample, I Dt represents the desert ecological security index of the tth sample, d1, d2 and d3 represent the proportion of forest area, wetland area and desert area to the total area of ​​these three categories respectively: <h2 style=";text-align:left;direction:ltr">I<h2 style=";text-align:left;direction:ltr"> t <h2 style=";text-align:left;direction:ltr"> =d1I<h2 style=";text-align:left;direction:ltr"> Ft <h2 style=";text-align:left;direction:ltr"> +d2I<h2 style=";text-align:left;direction:ltr"> wt <h2 style=";text-align:left;direction:ltr"> +d3I<h2 style=";text-align:left;direction:ltr"> Dt <h2 style=";text-align:left;direction:ltr"> (18)

Images