Forest fire risk intelligent identification system and method based on visual large model

Abstract

This invention discloses an intelligent forest fire risk identification system and method based on a large visual model, belonging to the field of forest fire risk monitoring and early warning technology. The system includes an alarm data access module, a large visual model recognition module, a multi-scene feature library module, a tagging and risk rating module, and a result push and storage module. The alarm data access module connects to forest area video monitoring equipment, accesses alarm data, and standardizes it; the large visual model recognition module extracts features and identifies scenes from alarm images based on a pre-trained large visual model; the multi-scene feature library module provides identification criteria and supports iterative updates; the tagging and risk rating module completes category tagging and four-level risk level classification; and the result push and storage module implements hierarchical push and full storage. This invention simplifies the daily tens of thousands of alarm messages in forest areas to hundreds of high-risk messages, significantly reducing the workload of manual verification and improving the accuracy of fire risk identification and emergency response efficiency.

Description

Technical Field

[0001] This invention relates to the field of information technology for forest fire risk monitoring and early warning, specifically to a forest fire risk intelligent identification system and method based on a large visual model. Background Technology

[0002] Forest fire risk monitoring is a core component of forest fire prevention, and accurate and efficient fire risk identification directly determines the timeliness and effectiveness of forest fire emergency response. Currently, most forest areas in China deploy video surveillance cameras equipped with commonly used intelligent smoke and fire analysis algorithms to automatically capture and alert on fire risks. However, in practical applications, the following significant shortcomings and problems exist:

[0003] The algorithm has low recognition accuracy and a high false alarm rate: the general smoke analysis algorithm can only recognize basic visual features of smoke and cannot distinguish between forest fires and similar scenarios such as agricultural burning (straw burning, mountain clearing), industrial emissions, and natural fog, resulting in a large number of false alarms every day and failing to effectively identify real fire risks.

[0004] The workload of manual verification is too large, and the monitoring loses its practical significance: tens of thousands of smoke and fire alarm messages are uploaded daily by a large number of monitoring devices in the forest area. On-duty personnel cannot check and verify each one. The real forest fire risk information is easily buried by the massive amount of false alarm information, making it difficult to deal with it in a timely manner.

[0005] The lack of an alarm information classification mechanism leads to confusion in handling priorities: The existing system does not classify alarm information into categories and risk levels, leaving on-duty personnel without clear handling priorities, resulting in a waste of human resources and delaying emergency response to real fire hazards.

[0006] Lack of feature library iteration capability and poor adaptability: The feature recognition model of general algorithms is fixed and cannot be updated according to the regional characteristics of forest areas and seasonal changes (such as spring plowing and burning, winter fog). The false alarm rate cannot be reduced with optimization in actual application.

[0007] It is evident that existing forest fire risk monitoring technologies cannot meet the core needs of forest areas for accurate fire risk identification, alarm classification, and efficient response. There is an urgent need for an intelligent forest fire risk identification system based on high-precision visual recognition technology to resolve the contradiction between massive false alarms and accurate monitoring. Summary of the Invention

[0008] To address the issues of high false alarm rates and heavy verification workloads for on-duty personnel in existing intelligent smoke and fire analysis algorithms for forest fire risk video monitoring, this invention provides a forest fire risk intelligent identification system and method based on a visual large-scale model. The system, centered on visual large-scale models, image feature extraction, and risk level determination technologies, comprises five core modules: an alarm data access module, a visual large-scale model recognition module, a multi-scene feature library module, a tagging and risk grading module, and a result push and storage module. These modules are interconnected and interact with each other in real time, enabling intelligent forest fire risk identification and management throughout the entire process, from forest area alarm data access and secondary image recognition to category tagging, risk level classification, result push, and data traceability. This reduces the massive daily forest fire risk alarm information from tens of thousands to hundreds of high-risk alarms, significantly decreasing the verification and processing workload for on-duty personnel and improving the accuracy of forest fire risk identification and emergency response efficiency, thereby solving the monitoring failure problem caused by the high false alarm rate of traditional algorithms.

[0009] To achieve the above objectives, the specific solution of the present invention is as follows:

[0010] A forest fire risk intelligent identification system based on a large visual model includes:

[0011] The alarm data access module is used to access alarm data from video surveillance equipment in forest areas, standardize the alarm data, and output standardized alarm data.

[0012] The visual large model recognition module is connected to the alarm data access module. It is used to receive standardized alarm data, extract features from alarm images based on pre-trained visual large models, compare and recognize the extracted features with scene feature data in the multi-scene feature library module, and output the recognition results.

[0013] The multi-scene feature library module is connected to the visual large model recognition module. It is used to store scene feature data, provide recognition basis to the visual large model recognition module, and receive manual correction results for iterative updates of the feature library.

[0014] The labeling and risk rating module is connected to the visual large model recognition module. It is used to receive recognition results, label alarm images with category labels and classify risk levels, and output graded alarm information. The labeling and risk rating module is also connected to the multi-scene feature library module to provide feedback on manual correction results to the multi-scene feature library module.

[0015] The results push and storage module, connected to the tag and risk classification module, is used to receive classified alarm information, push classified alarm information, and store alarm data, identification process data and handling results.

[0016] Furthermore, the alarm data access module includes:

[0017] The multi-protocol adapter unit is used to support the SDK protocol and the GB / T28181 national standard protocol, and automatically access alarm images, device numbers, capture times, latitude and longitude, and alarm type data.

[0018] The data cleaning and standardization unit is used to perform format verification, missing value completion, and duplicate data filtering on the incoming alarm data, and to uniformly convert non-standard image formats into JPG or PNG formats.

[0019] The data caching and queue scheduling unit is used to start the data caching queue when the concurrent alarm data exceeds the preset threshold, and to perform identification and scheduling according to the capture time.

[0020] The equipment status monitoring unit is used to monitor the online status and capture status of video surveillance equipment in real time, and send equipment fault alerts when the equipment goes offline or captures abnormally.

[0021] Furthermore, the visual large model recognition module includes:

[0022] The lightweight model deployment unit is used to crop and fine-tune the pre-trained large visual model deployed in the large visual model recognition module, so that the single image recognition response time is less than or equal to 1 second.

[0023] The multi-dimensional feature extraction unit is used to extract color features, texture features, scene features and spatial features from alarm images.

[0024] The image enhancement processing unit is used to automatically enhance alarm images in backlight, heavy fog, and low-light conditions at night.

[0025] The batch recognition processing unit is used to perform batch parallel recognition of alarm images in the cache queue.

[0026] Furthermore, the multi-scene feature library module adopts a two-layer architecture of a basic feature library and a custom feature library, and the multi-scene feature library module includes:

[0027] The basic feature library unit contains visual feature data for four core scenarios: forest fires, agricultural burning, industrial emissions, and natural fog. Each scenario contains feature values ​​from ≥100,000 sample images.

[0028] A custom feature library unit is used to add feature samples of typical local interference scenarios in forest areas;

[0029] The feature library automatic iteration unit is used to add manually verified alarm images to the feature library after labeling them, and to optimize the pre-trained visual large model through incremental training.

[0030] The version management unit is used to record updates and modifications to the feature library and supports version rollback.

[0031] Furthermore, the labeling and risk rating module includes:

[0032] The multi-category labeling unit is used to label each alarm image with five categories of tags based on the recognition results: forest fire, agricultural burning, industrial emissions, natural fog, and other interference.

[0033] The four-level risk level identification unit is used to classify alarm information into four levels: high risk, low risk, no risk, and false alarm.

[0034] The manual correction unit is used to support on-duty personnel to manually correct labels and risk levels, and synchronously feed the correction results back to the multi-scenario feature library module.

[0035] The high-risk information automatic early warning unit is used to trigger a triple warning system of local pop-up window, voice reminder and mobile phone text message when high-risk alarm information is identified.

[0036] Furthermore, the high-risk category is forest fire, and the smoke and fire location is in the core forest area or densely vegetated area; the low-risk category is no open flame, the smoke coverage is less than 30%, or it is marked as agricultural burning or industrial emissions, and the distance between the alarm location and the forest boundary is ≤500 meters; the no-risk category is agricultural burning, industrial emissions or natural fog, and the distance between the alarm location and the forest boundary is >500 meters; the false alarm category is other interference, which is false alarm caused by camera lens contamination or light reflection.

[0037] Furthermore, the result push and storage module includes:

[0038] The multi-terminal result push unit is used to push the graded alarm information to the desktop terminal of the forest fire prevention command center and the mobile terminal of the duty personnel according to the principle of high risk priority. The push content includes alarm images, location information, capture time, category tags, risk level and AI fire analysis report, and supports map visualization display.

[0039] The full data storage unit is used to store all alarm data, including raw information, identification process data, tags and risk levels, and handling results. The storage period is ≥3 years, and it supports multi-dimensional queries by device number, time, risk level, and category tag.

[0040] The statistical report generation unit is used to perform statistical analysis on alarm data and generate forest fire risk alarm identification statistical reports including total alarms, percentage of each scenario label, percentage of each risk level, false alarm rate, and model recognition accuracy. The report can be downloaded and printed.

[0041] The handling process traceability unit is used to record the handling process of high-risk alarm information, including the verifier, handling time, handling result and record.

[0042] A forest fire risk intelligent identification method based on a large visual model, applied to the aforementioned system, includes the following steps:

[0043] S1, Real-time alarm data access: The alarm data captured by the forest area video surveillance camera is accessed to the alarm data access module through a dedicated network. The alarm data includes alarm images of smoke or fire and the device number, capture time, latitude and longitude, and alarm type data. The alarm data is cleaned and standardized, and standardized alarm data is output and stored in the data cache queue.

[0044] S2, Secondary recognition of the visual big model: The visual big model recognition module retrieves standardized alarm data from the data cache queue described in step S1, enhances the alarm images, extracts multi-dimensional features through the pre-trained visual big model, matches the extracted multi-dimensional features with the feature data in the multi-scene feature library module, completes scene classification, and outputs the recognition results.

[0045] S3, Labeling and Risk Rating: Obtain the recognition results output from step S2 through the labeling and risk rating module, label the category based on the scene classification results, complete the four-level risk level classification according to the quantification rules, generate graded alarm information, and trigger triple warning when the graded alarm information is high risk.

[0046] S4, Multi-terminal push of graded results: The graded alarm information from step S3 is obtained through the result push and storage module, and the graded alarm information is pushed to the command center desktop terminal and the mobile terminal of the on-duty personnel according to the principle of high risk priority.

[0047] S5, Manual Verification and Feedback Iteration: On-duty personnel verify and handle the graded alarm information generated in step S3 through the terminal, and manually correct the incorrectly identified information through the label and risk classification module. The correction results are synchronized to the multi-scene feature library module, which adds the manually verified alarm images to the feature library and optimizes the pre-trained visual large model through incremental training.

[0048] S6, Data Statistics and Traceability Management: The original alarm data, identification process data, hierarchical alarm information and handling results generated in steps S1 to S5 are fully stored through the result push and storage module. Statistical reports are automatically generated, and the verifier, handling time and result of high-risk alarms are recorded.

[0049] Furthermore, the triple warning mentioned in step S3 includes a command center desktop pop-up, voice reminders, and mobile phone text messages.

[0050] Advantages of the present invention

[0051] Compared with existing forest fire risk monitoring technologies, this invention has the following significant advantages, and core indicators have been greatly improved through actual scenario testing:

[0052] 1. This invention extracts multi-dimensional features from alarm images—including color, texture, scene, and spatial features—using a large visual model recognition module. Combined with a two-layer architecture of a basic feature library and a custom feature library within a multi-scene feature library module, it achieves accurate classification of scenarios such as forest fires, agricultural burning, industrial emissions, and natural fog. Real-world testing shows that the false alarm rate for forest fire risks is reduced by over 90%, enabling accurate identification of real fire hazards.

[0053] 2. This invention classifies alarm information into four risk levels—high risk, low risk, no risk, and false alarm—through a tagging and risk grading module. It also pushes graded alarm information according to the principle of prioritizing high risk through a result push and storage module, reducing the number of alarms in the forest area from tens of thousands to hundreds of high-risk alarms per day. On-duty personnel only need to process high-risk alarms, which greatly reduces labor costs.

[0054] 3. This invention triggers a triple warning system in the command center via a desktop pop-up, voice reminder, and SMS message through a high-risk warning unit of the tag and risk rating module. The pushed content includes alarm images, location information, category tags, risk level, and AI fire analysis report. This can improve the efficiency of fire emergency response by more than 80%, enabling on-duty personnel to quickly locate the fire risk and achieve early detection and early handling of forest fires.

[0055] 4. This invention supports on-duty personnel to manually correct incorrectly identified information through the manual correction unit of the tag and risk rating module. The correction results are synchronized to the multi-scene feature library module, and the feature library automatic iteration unit adds the manually verified alarm images to the feature library. Through incremental training to optimize the visual large model, the false alarm rate is gradually reduced with actual application, adapting to the monitoring needs of different forest areas and different seasons.

[0056] 5. This invention supports the SDK protocols of mainstream video surveillance manufacturers and the GB / T28181 national standard protocol through the multi-protocol adaptation unit of the alarm data access module. It can directly connect to the monitoring equipment already deployed in forest areas without the need for hardware modification. At the same time, the system is based on a distributed deployment architecture, supports dual-active disaster recovery and dual-machine hot standby, can adapt to the concurrent access of alarm data from more than 700 video surveillance cameras, supports unlimited expansion of the number of monitoring devices, and adapts to the monitoring needs of forest areas at the provincial, municipal, and county levels.

[0057] 6. This invention stores all original alarm data, identification process data, classification results, and handling results in the full storage unit of the result push and storage module, with a storage period of ≥3 years. The statistical report generation unit automatically generates statistical reports containing total alarms, tag percentage, and false alarm rate. The handling traceability unit records the verifier, handling time, and results of high-risk alarms, realizing full-process digital traceability of forest fire risk identification and handling.

[0058] 7. This invention relies on the image feature extraction and classification capabilities of a large visual model, and connects with alarm data from video surveillance cameras in forest areas. It performs secondary intelligent recognition and analysis on captured smoke and fire alarm images, labeling the alarm images with categories such as forest fire, agricultural burning, industrial emissions, and natural fog. It also completes risk level identification according to high risk, low risk, no risk, and false alarm, achieving accurate screening of fire risk alarm information. It simplifies the massive amount of daily fire risk alarm information in forest areas from tens of thousands to hundreds of high-risk information, significantly reducing the workload of on-duty personnel in verification and processing, improving the accuracy of forest fire risk identification and emergency response efficiency, and solving the monitoring failure problem caused by the high false alarm rate of traditional algorithms. Attached Figure Description

[0059] Figure 1 This is a flowchart illustrating the architecture of the intelligent forest fire risk identification system based on a large visual model, as described in this invention.

[0060] Figure 2 This is a flowchart of the intelligent forest fire risk identification method based on a large visual model according to the present invention. Detailed Implementation

[0061] The present invention will be further explained and described below with reference to the accompanying drawings and specific embodiments. It should be noted that the specific embodiments are not intended to limit the scope of the present invention.

[0062] like Figure 1As shown in this specific embodiment, a forest fire risk intelligent identification system based on a large visual model is provided. Deployed on the e-government extranet and the dedicated forest fire prevention network, the system is centered on large visual models, image feature extraction, and risk level determination technologies. It includes an alarm data access module, a large visual model recognition module, a multi-scene feature library module, a tagging and risk grading module, and a result push and storage module. All modules are interconnected and interact in real time, enabling intelligent forest fire risk identification and management throughout the entire process, from forest area alarm data access and secondary image recognition to category tagging, risk level classification, result push, and data traceability. The system is based on a distributed deployment architecture, supporting dual-active disaster recovery and dual-machine hot standby. It can adapt to concurrent alarm data access from ≥700 video surveillance cameras and supports the daily processing of ≥10,000 alarm images, adapting to large-scale, high-concurrency fire risk monitoring applications in provincial forest areas. Simultaneously, it supports independent management by forest area, and each subsystem can customize feature library parameters and risk grading rules to meet the personalized monitoring needs of different forest areas.

[0063] The alarm data access module is used to access alarm data from video surveillance equipment in forest areas, and to standardize the alarm data and output standardized alarm data to achieve standardized and real-time access to alarm data from video surveillance cameras in forest areas.

[0064] The visual large model recognition module is connected to the alarm data access module. It is used to receive standardized alarm data, extract features from alarm images based on pre-trained visual large models, compare and recognize the extracted features with scene feature data in the multi-scene feature library module, and output the recognition results.

[0065] The multi-scene feature library module is connected to the visual large model recognition module. It is used to store scene feature data, provide recognition basis to the visual large model recognition module, and receive manual correction results for iterative updates of the feature library.

[0066] The labeling and risk rating module is connected to the visual large model recognition module to receive the recognition results, label the alarm images with category labels and classify them into four levels of risk, and output graded alarm information; the labeling and risk rating module is also connected to the multi-scene feature library module to provide feedback on the manual correction results to the multi-scene feature library module.

[0067] The result push and storage module is connected to the tag and risk classification module to receive classified alarm information, push classified alarm information, and store alarm data, identification process data and handling results, so as to realize multi-terminal push of identification results, full data storage and traceability of handling process.

[0068] The alarm data access module is the system's data source access module, enabling standardized and real-time access to alarm data from forest area video surveillance equipment. The alarm data access module includes a multi-protocol adaptation unit, a data cleaning and standardization unit, a data caching and queue scheduling unit, and an equipment status monitoring unit.

[0069] The multi-protocol adapter unit is used to support the SDK protocols of mainstream video surveillance manufacturers and the GB / T28181 national standard protocol, so as to realize the automatic access of alarm images, device numbers, capture time, latitude and longitude and alarm type data captured by fire risk video surveillance cameras already deployed in forest areas. The alarm types are divided into smoke and fire and smoke.

[0070] The data cleaning and standardization unit is used to perform format verification, missing value completion, and duplicate data filtering on the incoming alarm data, and to uniformly convert non-standard image formats into JPG or PNG formats to provide standardized data for subsequent identification.

[0071] The data caching and queue scheduling unit is used to start the data caching queue when the concurrent alarm data exceeds a preset threshold, and to perform identification and scheduling according to the capture time to avoid data loss and ensure the system's high concurrency processing capability.

[0072] The equipment status monitoring unit is used to monitor the online status and capture status of video surveillance equipment in real time, and send equipment fault reminders when the equipment goes offline or captures abnormally, so as to ensure the integrity of the monitoring network.

[0073] The visual large model recognition module is the core recognition module of the system. It relies on a pre-trained visual large model to achieve high-precision feature extraction and scene recognition of alarm images. The visual large model recognition module includes a lightweight model deployment unit, a multi-dimensional feature extraction unit, an image enhancement processing unit, and a batch recognition processing unit.

[0074] The lightweight model deployment unit is used to crop and fine-tune the pre-trained large visual model deployed in the large visual model recognition module, retaining the core capability of image feature extraction, reducing the model inference time, and achieving a single image recognition response time of ≤1 second;

[0075] The multi-dimensional feature extraction unit is used to extract color features, texture features, scene features, and spatial features from alarm images. Color features include red-orange flames and gray-white smoke. Texture features include flame pulsation and smoke diffusion. Scene features include forest vegetation, farmland, industrial areas, and open water areas. Spatial features include the distance between the fire and the forest area and the proportion of smoke coverage, thus distinguishing real forest fire risks from similar interference scenarios.

[0076] The image enhancement processing unit is used to automatically enhance alarm images in backlight, heavy fog, and low-light conditions at night. The enhancement processing includes brightness adjustment, defogging, and noise reduction to improve the model's recognition accuracy in complex environments.

[0077] The batch recognition processing unit is used to perform batch parallel recognition of alarm images in the cache queue, adapting to the processing needs of massive alarm data.

[0078] Batch recognition processing: Supports batch parallel recognition of alarm images in the cache queue, meeting the business needs of processing ≥10,000 alarm messages per day and ensuring recognition efficiency.

[0079] The multi-scene feature library module provides recognition basis for the large visual model and is the foundation for achieving accurate scene classification. The multi-scene feature library module adopts a two-layer architecture of basic feature library and custom feature library, and includes a basic feature library unit, a custom feature library unit, an automatic feature library iteration unit and a version management unit.

[0080] The basic feature library unit contains visual feature data for four core scenarios: forest fires, agricultural burning, industrial emissions, and natural fog. Each scenario includes feature values ​​from ≥100,000 sample images, covering scenario features from different regions, seasons, and weather conditions. Agricultural burning includes straw burning and scorched earth; industrial emissions include exhaust gas and smoke; and natural fog includes morning fog and haze.

[0081] The custom feature library unit allows users to add feature samples of typical local interference scenarios, such as burning paper offerings during forest rituals and cooking smoke from farmhouses, based on the actual conditions of the forest area, thus achieving localized adaptation of the feature library.

[0082] Automatic feature library iteration: The system automatically adds manually verified alarm images (correctly identified or incorrectly identified) to the feature library, and optimizes the pre-trained visual model through incremental training to gradually reduce the false alarm rate;

[0083] The version management unit is used to record updates and modifications to the feature library and supports version rollback to ensure the stability of model recognition.

[0084] The labeling and risk rating module is the system's decision output module. Based on the recognition results of the large visual model, it completes the category labeling and risk level classification of alarm images. The labeling and risk rating module includes a category labeling unit, a four-level risk level unit, a manual correction unit, and a high-risk warning unit.

[0085] The multi-category labeling unit is used to label each alarm image with five categories of tags based on the recognition results: forest fire, agricultural burning, industrial emissions, natural fog, and other interference, so as to achieve accurate classification of alarm scenarios.

[0086] The four-level risk level labeling unit is used to formulate quantitative risk classification rules, dividing alarm information into four levels: high risk, low risk, no risk, and false alarm. High risk is indicated by forest fires, with smoke and fire located in core forest areas or densely vegetated areas, or smoke covering more than 30% of the screen. Low risk is indicated by no open flames, smoke covering less than 30% indicating agricultural burning or industrial emissions, and the distance between the alarm location and the forest boundary is ≤500 meters, indicating a risk of spread. No risk is indicated by agricultural burning, industrial emissions, or natural fog, and the distance between the alarm location and the forest boundary is >500 meters, indicating no risk of spread. False alarms are indicated by other interference, such as camera lens contamination or light reflection causing false alarms.

[0087] The manual correction unit is used to support on-duty personnel to manually correct labels and risk levels, and synchronously feed the correction results back to the multi-scenario feature library module for model iteration.

[0088] The high-risk information automatic early warning unit is used to trigger a triple warning system of local pop-up window, voice reminder and mobile phone text message when high-risk alarm information is identified, so as to ensure that the on-duty personnel can perceive it at the first time.

[0089] The result push and storage module is the system's result output and data management module, realizing multi-terminal push of identification results and full-process data storage traceability. The result push and storage module includes a multi-terminal push unit, a full storage unit, a statistical report generation unit, and a handling traceability unit.

[0090] The multi-terminal result push unit is used to push the graded alarm information to the desktop terminal of the forest fire prevention command center and the mobile terminal of the duty personnel according to the principle of high risk priority. The push content includes alarm images, location information, capture time, category tags, risk level and AI fire analysis report, and supports map visualization display.

[0091] The full data storage unit is used to store all alarm data, including raw information, identification process data, tags and risk levels, and handling results. The storage period is ≥3 years, and it supports multi-dimensional queries by device number, time, risk level, and category tag.

[0092] The statistical report generation unit is used to perform statistical analysis on alarm data and generate forest fire risk alarm identification statistical reports including total alarms, percentage of each scenario label, percentage of each risk level, false alarm rate, and model recognition accuracy. The report can be downloaded and printed.

[0093] The handling process traceability unit is used to record the handling process of high-risk alarm information, including the verifier, handling time, handling results and records, so as to realize the closed-loop traceability of the entire process of forest fire risk from identification to handling.

[0094] like Figure 2 As shown, a forest fire risk intelligent identification method based on a large visual model, applied to the above-mentioned system, includes the following steps:

[0095] S1, Real-time alarm data access: Alarm data captured by more than 700 forest fire risk video surveillance cameras deployed in the forest area is accessed to the alarm data access module through a dedicated network. The alarm data includes alarm images of smoke or fire, as well as equipment number, capture time, latitude and longitude, and alarm type data. The alarm data is cleaned and standardized, and standardized alarm data is output and stored in the data cache queue.

[0096] S2, Secondary recognition of visual large model: The visual large model recognition module retrieves standardized alarm data from the data cache queue described in step 1, enhances the alarm image, extracts multi-dimensional features through the pre-trained visual large model, matches the extracted multi-dimensional features with the feature data in the multi-scene feature library module, completes scene classification, and outputs the recognition result.

[0097] S3, Labeling and Risk Rating: The identification results output from step 2 are obtained through the labeling and risk rating module. Category labels are labeled according to the scene classification results. Based on the quantification rules, four levels of risk are classified: high risk, low risk, no risk, and false alarm. Graded alarm information is generated. When the graded alarm information is high risk, a triple warning is triggered. The triple warning includes a pop-up window on the command center desktop, voice reminder, and mobile phone text message.

[0098] S4, Multi-terminal push of graded results: The graded alarm information from step 3 is obtained through the result push and storage module, and the graded alarm information is pushed to the command center desktop terminal and the mobile terminal of the duty personnel according to the high-risk priority principle. The duty personnel only need to focus on the high-risk information.

[0099] S5, Manual Verification and Feedback Iteration: On-duty personnel verify and handle high-risk information in the graded alarm information generated in step 3 through the terminal, and manually correct the incorrectly identified information through the label and risk classification module. The correction results are synchronized to the multi-scene feature library module, which adds the manually verified alarm images to the feature library and optimizes the pre-trained visual large model through incremental training.

[0100] S6, Data Statistics and Traceability Management: The result push and storage module fully stores the original alarm data, identification process data, hierarchical alarm information and handling results generated in steps 1 to 5, automatically generates statistical reports, records the verifier, handling time and results of high-risk alarms, and realizes full-process digital traceability of forest fire risk identification and handling.

[0101] The aforementioned intelligent forest fire risk identification system based on a large visual model was deployed on the Guangxi e-government extranet. In accordance with the forest fire prevention needs of Guangxi Zhuang Autonomous Region, it was connected to more than 700 forest fire risk video surveillance cameras in key forest areas throughout the region. The system supports processing ≥10,000 alarm images per day, with a single image recognition response time of ≤1 second, realizing intelligent identification and hierarchical handling of forest fire risks in forest areas at the autonomous region, city, and county levels.

[0102] Example 1: Identification of High-Risk Forest Fire Information and Emergency Response

[0103] Alarm data access: On March 9, 2026, at 11:55:46, a video surveillance camera in the Wenwei Muhu Forest Area of ​​Mengshan County, Wuzhou City, captured a picture of dense smoke. The device automatically triggered a smoke alarm and connected the alarm image, device latitude and longitude, capture time, and alarm type data to the forest fire risk intelligent identification system based on the visual big model in the above embodiment in real time through the dedicated forest fire prevention network. After the system completes data cleaning, the data enters the identification queue.

[0104] Secondary model recognition: After dehazing and enhancing the alarm images, the system extracts image features using a lightweight visual model: The image shows a forest area shrouded in dense fog, with thick smoke rising from the trees. The smoke covers a large area, is whitish in color, and is located deep within the forest. The smoke concentration is high, but there are no flames or firelight features. There are personnel on site, possibly forest rangers or forest managers. Through multi-dimensional feature matching, the system determines the scene to be a "forest fire."

[0105] Tagging and Risk Classification: The system automatically tags the alarm image with the category "forest fire". Based on the rules, it is determined to be high risk, which immediately triggers a pop-up window and voice reminder on the command center desktop terminal, and at the same time sends a text message warning to the mobile phone of the on-duty personnel.

[0106] AI Analysis Report Generation: The system automatically generates an AI fire analysis report, which includes:

[0107] Time Analysis: March 9, 2026, 11:55:46

[0108] Location Analysis: Wooden Protection at Wenwei, Mengshan, Wuzhou, Guangxi - Natural Disaster

[0109] Scene Analysis: The image shows a forest area shrouded in dense fog, with thick smoke rising from the trees. The smoke covers a large area, is whitish in color, and is located deep within the forest. The smoke concentration is high, but there are no signs of flame or fire. Personnel are present at the scene, possibly forest rangers or forest supervisors. The smoke may have been caused by a forest fire, but there are no visible flames or signs of fire, and the smoke area does not exceed 30% of the image. Furthermore, the dense fog may impair visibility, making it difficult to clearly identify the specific location of the smoke and the direction of fire spread.

[0110] Disaster analysis: Dense smoke is a characteristic of forest fires, but there is a lack of open flames and traces of fire. The smoke coverage is less than 30%, and dense fog may affect the assessment of the fire situation, so the disaster risk is relatively low.

[0111] Fire hazard type: Forest fire

[0112] Risk level: High risk

[0113] Recommendations: It is suggested that you immediately contact the local emergency management bureau and forestry bureau, and in case of emergency, dispatch personnel directly to the local township.

[0114] Results push and handling: The system pushes high-risk alarm information and AI analysis reports to the Wuzhou Forest Fire Prevention Command Center, including images of dense smoke, precise latitude and longitude coordinates, and a map of the forest area location. On-duty personnel immediately dispatch nearby forest rangers to the scene to verify the situation. Once confirmed as a forest fire, the emergency response process is initiated.

[0115] Data storage and traceability: The system stores the original data of this alarm, the identification results, the verifier, the handling time, and the handling process of the fire. It can be queried and replayed online at any time to achieve full-process traceability.

[0116] Implementation results: From the moment the camera captured the alarm to the moment the on-duty personnel received the high-risk warning, the entire process took only 15 seconds. Compared to the original system where personnel had to manually sift through tens of thousands of messages, the response efficiency was improved by 95%, enabling early detection and early response to the fire and effectively controlling its spread.

[0117] Example 2: Identification and Management of Low-Risk Information on Agricultural Burning

[0118] Alarm data access: A video surveillance camera at the boundary of a forest area in Guigang City, Guangxi Province, captured a smoke image, triggering a smoke alarm. The alarm image, the device's latitude and longitude, and the capture time (16:20:15 in the afternoon) were then accessed into the system.

[0119] Secondary recognition of visual model: The large visual model extracts image features: the smoke is grayish-white, the diffusion speed is slow, the background is farmland, the action of farmers burning straw can be seen in the picture, and the location of the smoke is 300 meters away from the forest boundary, which matches the feature of "agricultural burning".

[0120] Labeling and Risk Assessment: The system labels agricultural burning as "agricultural burning". According to the rules, if it is close to the forest boundary and there is a risk of spread, it is judged as low risk and the information is pushed to the local county-level forest fire prevention duty terminal.

[0121] Results Handling and Feedback: After receiving a low-risk information, the on-duty personnel dispatched local township staff to the site for control, requiring the burning personnel to clear flammable materials to prevent the fire from spreading to the forest area. At the same time, the handling results were entered into the system, and the system marked the verified result and added it to the multi-scenario feature database.

[0122] Implementation results: The agricultural burning alarms were accurately classified and classified as low risk, which not only achieved effective control of risks at the forest boundary, but also did not take up the time of the on-duty personnel to deal with high-risk information. Compared with the indiscriminate alarms of the original system, the control was more targeted.

[0123] Example 3: Identification and Filtering of False Alarm Information for Natural Fog

[0124] Alarm data access: A video surveillance camera in a forest area of ​​Hezhou City, Guangxi Province, captured a white fog image due to heavy fog in the early morning. The device erroneously triggered a smoke alarm and accessed the alarm image and capture time (time: 06:10:00 AM) into the system.

[0125] Secondary recognition of visual model: The large visual model extracts image features: The image is uniform gray-white fog with no fixed smoke source and no flame features. The background is forest vegetation, which highly matches the features of "natural fog".

[0126] Labeling and Risk Assessment: The system labels the fog as "natural fog" and determines it as risk-free according to the rules. This information is then added to the backend database and not pushed to on-duty personnel.

[0127] Data statistics and model optimization: The system includes the identification result of this risk-free information in the daily statistical report, and the data on the proportion of false alarms related to natural fog is updated in real time, providing data support for forest fire prevention work.

[0128] Implementation Results: The false alarms caused by natural fog were automatically filtered by the system and not pushed to the on-duty personnel, effectively reducing meaningless alarm information. On that day, the system filtered more than 9,800 such risk-free / false alarm messages and only pushed 120 high-risk messages, significantly reducing the workload of the on-duty personnel.

[0129] The above embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Any modifications, equivalent substitutions, improvements, etc., made based on the technical solutions of the present invention should be included within the protection scope of the present invention.

Claims

1. A forest fire risk intelligent identification system based on a large visual model, characterized in that, include: The alarm data access module is used to access alarm data from video surveillance equipment in forest areas, standardize the alarm data, and output standardized alarm data. The visual large model recognition module is connected to the alarm data access module. It is used to receive standardized alarm data, extract features from alarm images based on pre-trained visual large models, compare and recognize the extracted features with scene feature data in the multi-scene feature library module, and output the recognition results. The multi-scene feature library module is connected to the visual large model recognition module. It is used to store scene feature data, provide recognition basis to the visual large model recognition module, and receive manual correction results for iterative updates of the feature library. The labeling and risk rating module is connected to the visual large model recognition module. It is used to receive recognition results, label alarm images with category labels and classify risk levels, and output graded alarm information. The labeling and risk rating module is also connected to the multi-scene feature library module to provide feedback on manual correction results to the multi-scene feature library module. The results push and storage module, connected to the tag and risk classification module, is used to receive classified alarm information, push classified alarm information, and store alarm data, identification process data and handling results.

2. The system according to claim 1, characterized in that, The alarm data access module includes: The multi-protocol adapter unit is used to support the SDK protocol and the GB / T28181 national standard protocol, and automatically access alarm images, device numbers, capture times, latitude and longitude, and alarm type data. The data cleaning and standardization unit is used to perform format verification, missing value completion, and duplicate data filtering on the incoming alarm data, and to uniformly convert non-standard image formats into JPG or PNG formats. The data caching and queue scheduling unit is used to start the data caching queue when the concurrent alarm data exceeds the preset threshold, and to perform identification and scheduling according to the capture time. The equipment status monitoring unit is used to monitor the online status and capture status of video surveillance equipment in real time, and send equipment fault alerts when the equipment goes offline or captures abnormally.

3. The system according to claim 1, characterized in that, The visual large model recognition module includes: The lightweight model deployment unit is used to crop and fine-tune the pre-trained large visual model deployed in the large visual model recognition module, so that the single image recognition response time is less than or equal to 1 second. The multi-dimensional feature extraction unit is used to extract color features, texture features, scene features and spatial features from alarm images. The image enhancement processing unit is used to automatically enhance alarm images in backlight, heavy fog, and low-light conditions at night. The batch recognition processing unit is used to perform batch parallel recognition of alarm images in the cache queue.

4. The system according to claim 1, characterized in that, The multi-scene feature library module adopts a two-layer architecture of basic feature library and custom feature library, and the multi-scene feature library module includes: The basic feature library unit contains visual feature data for four core scenarios: forest fires, agricultural burning, industrial emissions, and natural fog. Each scenario contains feature values ​​from ≥100,000 sample images. A custom feature library unit is used to add feature samples of typical local interference scenarios in forest areas; The feature library automatic iteration unit is used to add manually verified alarm images to the feature library after labeling them, and to optimize the pre-trained visual large model through incremental training. The version management unit is used to record updates and modifications to the feature library and supports version rollback.

5. The system according to claim 1, characterized in that, The labeling and risk rating module includes: The multi-category labeling unit is used to label each alarm image with five categories of tags based on the recognition results: forest fire, agricultural burning, industrial emissions, natural fog, and other interference. The four-level risk level identification unit is used to classify alarm information into four levels: high risk, low risk, no risk, and false alarm. The manual correction unit is used to support on-duty personnel to manually correct labels and risk levels, and synchronously feed the correction results back to the multi-scenario feature library module. The high-risk information automatic early warning unit is used to trigger a triple warning system of local pop-up window, voice reminder and mobile phone text message when high-risk alarm information is identified.

6. The system according to claim 5, characterized in that, High-risk areas are marked as forest fires, with smoke and fire located in core forest areas or densely vegetated areas; low-risk areas are marked as no open flames, with smoke coverage of less than 30%, or as agricultural burning or industrial emissions, with the distance between the alarm location and the forest boundary ≤ 500 meters; no-risk areas are marked as agricultural burning, industrial emissions, or natural fog, with the distance between the alarm location and the forest boundary > 500 meters; false alarms are marked as other interference, which includes false alarms caused by camera lens contamination or light reflection.

7. The system according to claim 1, characterized in that, The result push and storage module includes: The multi-terminal result push unit is used to push the graded alarm information to the desktop terminal of the forest fire prevention command center and the mobile terminal of the duty personnel according to the principle of high risk priority. The push content includes alarm images, location information, capture time, category tags, risk level and AI fire analysis report, and supports map visualization display. The full data storage unit is used to store all alarm data, including raw information, identification process data, tags and risk levels, and handling results. The storage period is ≥3 years, and it supports multi-dimensional queries by device number, time, risk level, and category tag. The statistical report generation unit is used to perform statistical analysis on alarm data and generate forest fire risk alarm identification statistical reports, including total alarms, percentage of each scenario label, percentage of each risk level, false alarm rate, and model recognition accuracy. The unit also supports report downloading and printing. The handling process traceability unit is used to record the handling process of high-risk alarm information, including the verifier, handling time, handling result, and record.

8. A method for intelligent identification of forest fire risk based on a large visual model, applied to the system described in any one of claims 1 to 7, characterized in that, Includes the following steps: S1, Real-time alarm data access: The alarm data captured by the forest area video surveillance camera is accessed to the alarm data access module through a dedicated network. The alarm data includes alarm images of smoke or fire and the device number, capture time, latitude and longitude, and alarm type data. The alarm data is cleaned and standardized, and standardized alarm data is output and stored in the data cache queue. S2, Secondary recognition of the visual big model: The visual big model recognition module retrieves standardized alarm data from the data cache queue described in step S1, enhances the alarm images, extracts multi-dimensional features through the pre-trained visual big model, matches the extracted multi-dimensional features with the feature data in the multi-scene feature library module, completes scene classification, and outputs the recognition results. S3, Labeling and Risk Rating: Obtain the recognition results output from step S2 through the labeling and risk rating module, label the category based on the scene classification results, complete the four-level risk level classification according to the quantification rules, generate graded alarm information, and trigger triple warning when the graded alarm information is high risk. S4, Multi-terminal push of graded results: The graded alarm information from step S3 is obtained through the result push and storage module, and the graded alarm information is pushed to the command center desktop terminal and the mobile terminal of the on-duty personnel according to the principle of high risk priority. S5, Manual Verification and Feedback Iteration: On-duty personnel verify and handle the graded alarm information generated in step S3 through the terminal, and manually correct the information with identification errors through the label and risk classification module. The correction results are synchronized to the multi-scene feature library module, which adds the manually verified alarm images to the feature library and optimizes the pre-trained visual large model through incremental training. S6, Data Statistics and Traceability Management: The result push and storage module fully stores the original alarm data, identification process data, hierarchical alarm information and handling results generated in steps S1 to S5, automatically generates statistical reports, and records the verifier, handling time and result of high-risk alarms.

9. The method according to claim 8, characterized in that, The triple warning mentioned in step S3 includes a pop-up window on the command center desktop, voice reminders, and SMS messages.

Images