A forest surface fuel moisture content prediction method, device, equipment and medium

Abstract

This application discloses a method, apparatus, equipment, and medium for predicting the moisture content of combustibles on forest surfaces, relating to the field of combustible moisture content prediction. The method includes: acquiring multimodal meteorological factors of the target forest surface area at the current time; the multimodal meteorological factors include ground-based meteorological data, airborne satellite data, and space-based remote sensing data collected using four complementary methods; preprocessing the multimodal meteorological factors at the current time; performing decision-level information fusion processing on the preprocessed real-time data and a standard database to obtain a multi-source feature data set under the same time series; inputting the multi-source feature data set into a moisture content prediction model to obtain the current combustible moisture content; the moisture content prediction model is constructed based on a bidirectional long short-term memory deep learning algorithm, and the model parameters are optimized using a multiverse algorithm. This application improves the accuracy and practicality of predicting the moisture content of combustibles on forest surfaces.

Description

Technical Field

[0001] This application relates to the field of predicting the moisture content of combustibles, and in particular to a method, apparatus, equipment and medium for predicting the moisture content of combustibles on forest surfaces. Background Technology

[0002] Forest fires are characterized by their suddenness, destructiveness, and difficulty in control, seriously threatening the safety of forest ecosystems, the safety of people's lives and property, and the stability of the ecological environment. Small dead combustible materials on the forest surface (such as dead branches, fallen leaves, and small twigs) are the core carriers of forest fires and their rapid spread. Their moisture content directly determines the difficulty of ignition, the rate of spread, and the intensity of combustion, and is a key basic parameter for forest fire risk level forecasting and forest fire prevention and control decision-making.

[0003] Currently, technologies for predicting the moisture content of fine dead combustibles on the forest surface have gradually matured, resulting in various technical solutions, mainly including remote sensing estimation, equilibrium moisture content methods, meteorological element regression methods, and process model methods. Meanwhile, some regions have begun to experiment with using drones and fixed monitoring stations to assist in meteorological data collection, attempting to improve the convenience, coverage, and efficiency of prediction technologies. However, many technical problems still urgently need to be solved, severely limiting the accuracy of predictions and their practical application effectiveness.

[0004] First, the data collection methods are limited. Most schemes rely on a single fixed collection station or a single remote sensing method, resulting in limited collection range and insufficient data coverage. This makes it difficult to capture the differences in micro-meteorological factors in different areas of the forest, or the accuracy of small-scale predictions is low and the workload for large-scale applications is large, which cannot meet the prediction needs of complex forest areas.

[0005] Second, the data integration is insufficient. Most related technologies have not carried out effective decision-level fusion processing of historical standard meteorological data and real-time meteorological data, resulting in problems such as inconsistent time series, lack of completeness, and high redundancy of data, which directly affects the calculation accuracy of subsequent prediction models.

[0006] Third, the predictive models have poor adaptability. Most of the relevant models are general-purpose and have not been optimized in a targeted manner by fully combining the changes in the moisture content of small dead combustibles on the forest surface (such as the dynamic changes caused by the combined influence of multiple factors such as temperature, humidity, wind speed, and light). In addition, some models have the drawbacks of poor cross-regional generalization and large prediction errors. Among them, the average relative error of traditional models such as multiple regression and CART in predicting the moisture content of small dead combustibles is still at a high level, which cannot meet the high-precision requirements of forest fire risk level forecasting.

[0007] In addition, some solutions still rely on manual sampling and monitoring, which has limitations such as high time and labor costs and the inability to achieve real-time dynamic prediction, further restricting the promotion and application of prediction technology in forest fire prevention and control.

[0008] Therefore, improving the accuracy and practicality of predicting the moisture content of forest surface combustibles (such as fine dead combustibles) has become an urgent need in the field of forest fire prevention and control. Summary of the Invention

[0009] The purpose of this application is to provide a method, apparatus, equipment, and medium for predicting the moisture content of forest surface combustibles, which can improve the accuracy and practicality of predicting the moisture content of forest surface combustibles.

[0010] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a method for predicting the moisture content of forest surface combustibles, including: The target forest surface area is acquired at the current time using multimodal meteorological factors. These multimodal meteorological factors include ground-based meteorological data, airborne satellite data, and space-based remote sensing data collected by fixed multi-factor meteorological acquisition stations, mobile meteorological data calibration platforms, forest near-ground UAV monitoring units, and satellite remote sensing data acquisition units. The multimodal meteorological factors at the current moment are preprocessed to obtain preprocessed real-time data; The preprocessed real-time data is fused with multimodal meteorological factors from historical moments in a standard database at the decision-level information level to obtain a collection of multi-source feature data under the same time series. The multi-source feature data set is input into the moisture content prediction model to obtain the combustible moisture content of the target forest surface area at the current time. The moisture content prediction model is constructed based on the bidirectional long short-term memory deep learning algorithm and the multiverse algorithm is used to optimize the model parameters.

[0011] Secondly, this application provides a device for predicting the moisture content of forest surface combustibles, comprising: The system includes a fixed multi-factor meteorological data acquisition station, a mobile meteorological data calibration platform, a forest near-ground UAV monitoring unit, a satellite remote sensing data acquisition unit, a distributed forest data transmission network, and a moisture content prediction module; the distributed forest data transmission network includes a LoRa self-organizing communication network and a 5G long-distance communication network. The fixed multi-factor meteorological data acquisition station is connected to the 5G long-distance communication network via the LoRa self-organizing communication network; the mobile meteorological data calibration platform and the forest near-ground UAV monitoring unit are both connected to the distributed forest data transmission network; the satellite remote sensing data acquisition unit is connected to the 5G long-distance communication network. The fixed multi-factor meteorological data acquisition station is deployed within the target forest to collect baseline meteorological factors for the target forest surface area; the forest-based near-ground UAV monitoring unit is used to collect three-dimensional meteorological data for the target forest surface area; the mobile meteorological data calibration platform is used to perform on-site calibration and accuracy verification of the baseline meteorological factors and the three-dimensional meteorological data, using the verified baseline meteorological factors as ground-based meteorological data and the verified three-dimensional meteorological data as airborne satellite data; the satellite remote sensing data acquisition unit is used to collect space-based remote sensing data for the target forest surface area. The moisture content prediction module includes: The data acquisition unit is used to acquire multimodal meteorological factors of the target forest surface area at the current time; the multimodal meteorological factors include: ground-based meteorological data, air-based satellite data and space-based remote sensing data collected by fixed multi-factor meteorological acquisition stations, mobile meteorological data calibration platforms, forest near-ground UAV monitoring units and satellite remote sensing data acquisition units; The data preprocessing unit is used to preprocess the multimodal meteorological factors at the current moment to obtain preprocessed real-time data; The information fusion unit is used to perform decision-level information fusion processing on the preprocessed real-time data and the multimodal meteorological factors of historical time in the standard database to obtain a set of multi-source feature data under the same time series. The moisture content prediction unit is used to input the multi-source feature data set into the moisture content prediction model to obtain the combustible moisture content of the target forest surface area at the current time. The moisture content prediction model is constructed based on the bidirectional long short-term memory deep learning algorithm and the multiverse algorithm is used to optimize the model parameters.

[0012] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the computer program to implement the method for predicting the moisture content of forest surface combustibles as described above.

[0013] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for predicting the moisture content of forest surface combustibles as described above.

[0014] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a method, apparatus, equipment, and medium for predicting the moisture content of forest surface combustibles. It employs four complementary methods (fixed multi-factor meteorological data acquisition station, mobile meteorological data calibration platform, forest near-ground UAV monitoring unit, and satellite remote sensing data acquisition unit) to collect ground-based meteorological data, airborne satellite data, and space-based remote sensing data, meeting the prediction needs of complex forest areas. Preprocessed real-time data is fused with a standard database at the decision-level to obtain a multi-source feature data set under the same time series, improving the computational accuracy of subsequent prediction models. The moisture content prediction model is constructed based on a bidirectional long short-term memory deep learning algorithm, and the multiverse algorithm is used to optimize the model parameters, meeting the high-precision requirements of forest fire risk level forecasting. Furthermore, this application can automatically achieve dynamic prediction in real time, facilitating its widespread application in forest fire prevention and control. Therefore, this application improves the accuracy and practicality of predicting the moisture content of forest surface combustibles. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating a method for predicting the moisture content of forest surface combustibles, provided in an embodiment of this application; Figure 2 This is a schematic diagram of a convolutional neural network structure provided in an embodiment of this application; Figure 3 A schematic diagram of the convolution process provided in the embodiments of this application; Figure 4 A schematic diagram of the pooling process provided in an embodiment of this application; Figure 5 This is a schematic diagram of a long short-term memory neural network structure provided in an embodiment of this application; Figure 6 This is a schematic diagram of the Bi-LSTM neural network structure provided in the embodiments of this application; Figure 7 This is a schematic diagram of the MVO-Bi-LSTM prediction framework provided in the embodiments of this application; Figure 8 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] In one exemplary embodiment, such as Figure 1 As shown, a method for predicting the moisture content of forest surface combustibles is provided, including: Step 101: Obtain the multimodal meteorological factors of the target forest surface area at the current moment.

[0020] The multimodal meteorological factors include: ground-based meteorological data, airborne satellite data, and space-based remote sensing data collected by fixed multi-factor meteorological acquisition stations, mobile meteorological data calibration platforms, forest near-ground UAV monitoring units, and satellite remote sensing data acquisition units.

[0021] Step 102: Preprocess the multimodal meteorological factors at the current moment to obtain preprocessed real-time data.

[0022] Step 103: The preprocessed real-time data is fused with the multimodal meteorological factors of historical time in the standard database at the decision level to obtain a set of multi-source feature data under the same time series.

[0023] Step 104: Input the multi-source feature data set into the moisture content prediction model to obtain the combustible moisture content of the target forest surface area at the current moment; the moisture content prediction model is constructed based on the bidirectional long short-term memory deep learning algorithm and the multiverse algorithm is used to optimize the model parameters.

[0024] In another exemplary embodiment of this application, step 101 specifically includes: The baseline meteorological factors are acquired from fixed multi-factor meteorological data collection stations deployed in the target forest; the fixed multi-factor meteorological data collection stations transmit data sequentially through LoRa network and 5G network.

[0025] The system acquires three-dimensional meteorological data collected by a forest-based near-ground drone monitoring unit; the forest-based near-ground drone monitoring unit transmits data via a LoRa network or a 5G network.

[0026] A mobile meteorological data calibration platform is used to perform on-site calibration and accuracy verification of the reference meteorological factors and the stereo meteorological data. The verified reference meteorological factors are used as ground-based meteorological data, and the verified stereo meteorological data are used as airborne satellite data. The mobile meteorological data calibration platform transmits data through a LoRa network or a 5G network.

[0027] The system acquires space-based remote sensing data collected by a satellite remote sensing data acquisition unit; the satellite remote sensing data acquisition unit uses a 5G network for data transmission.

[0028] In another exemplary embodiment of this application, step 102 specifically includes: preprocessing the multimodal meteorological factors at the current moment using outlier removal algorithm, linear interpolation algorithm, neighborhood mean filling algorithm, Z-score data standardization algorithm and Gaussian filtering noise reduction algorithm to obtain preprocessed real-time data.

[0029] In another exemplary embodiment of this application, step 103 specifically includes: using a convolutional neural network to evaluate the credibility of each source data in the preprocessed real-time data, and based on the credibility evaluation results and evidence synthesis rules, performing information fusion processing on the preprocessed real-time data and the multimodal meteorological factors of historical moments in the standard database to obtain a set of multi-source feature data under the same time series; the convolutional neural network includes: an input layer, a convolutional layer, a pooling layer and a fully connected layer connected in sequence; the convolutional neural network adopts a dynamic weight adjustment strategy for fusion during the fusion process.

[0030] In another exemplary embodiment of this application, step 104, the method for determining the moisture content prediction model, includes: Obtain multimodal meteorological factors for each historical moment used for training.

[0031] The multimodal meteorological factors used for training at each historical moment are preprocessed to obtain preprocessed historical data.

[0032] A training dataset is constructed based on preprocessed historical data and corresponding measured moisture content of combustibles.

[0033] Construct a bidirectional long short-term memory deep learning network (Bi-LSTM neural network).

[0034] The training dataset is input into the bidirectional long short-term memory deep learning network. Based on the Adam optimizer and early stopping strategy, and using the multiverse algorithm, the model parameters in the bidirectional long short-term memory deep learning network are optimized and trained. The trained bidirectional long short-term memory deep learning network is determined as the water content prediction model.

[0035] This embodiment first constructs a distributed forest data transmission network within the forest using a combination of LoRa and 5G communication. Based on this, it employs four complementary methods: a self-developed fixed multi-factor meteorological data acquisition station, a mobile meteorological data calibration platform, a near-ground UAV monitoring unit within the forest, and a satellite remote sensing data acquisition unit. This acquires meteorological factors and surface fine dead combustible material moisture content in the forest's surface area, constructing a standard database of forest meteorological factors and surface fine dead combustible material moisture content. Second, a convolutional neural network (CNN) algorithm is used to perform decision-level information fusion processing on the standard database, obtaining a multi-source feature data set for the same time series. Then, a moisture content prediction model based on a bidirectional long short-term memory (Bi-LSTM) deep learning algorithm is constructed, and the model parameters are optimized using a multiverse multiverse (MVO) algorithm. Finally, the multi-source data set is input into the prediction model, outputting the moisture content prediction result. This embodiment significantly improves the prediction accuracy and generalization ability of forest surface combustible moisture content through multi-source complementary data acquisition, decision-level information fusion, and targeted optimization of deep learning models, providing high-precision basic data support for forest fire risk level forecasting.

[0036] The following section uses small dead combustibles on the surface of a forest as an example to further explain the above-mentioned method for predicting the moisture content of combustibles on the forest surface.

[0037] This embodiment strictly follows the core logic of "acquisition-processing-fusion-computation-output", linking various modules to work together, and specifically includes the following steps.

[0038] Step 1: Complementary acquisition of multi-source data and construction of a standard database.

[0039] This step aims to address the issues of limited data sources and insufficient coverage in existing technologies. By employing five complementary data collection methods, meteorological factors and land cover data for areas containing small dead combustibles on the forest surface are obtained, and a standard database is constructed.

[0040] This embodiment is a multimodal meteorological factor acquisition system based on a distributed forest transmission network. It employs a fusion architecture of LoRa self-organizing communication network and 5G long-range communication network, using a star topology layout to adapt to the complex terrain of forests and the need for large-scale, low-power transmission. The LoRa network handles local, short-range, low-power data transmission within the forest, providing 3-15km short-range coverage, multi-node access, and data caching and retransmission capabilities. The 5G network serves as the backbone transmission layer, enabling long-distance, high-speed data transmission, multi-source data integration, seamless network switching, and secure encryption. The system integrates four acquisition units: a fixed multi-factor meteorological acquisition station, a mobile meteorological data calibration platform, a near-ground UAV within the forest, and satellite remote sensing. This constructs an integrated air-ground acquisition system combining "ground-based precise acquisition + airborne three-dimensional supplementary acquisition + space-based macroscopic monitoring." Each unit relies on the dual networks to collaboratively complete data acquisition, transmission, and calibration. By implementing a hierarchical transmission and unified scheduling mechanism, multi-source data closed-loop calibration, complementary fusion of air, space, and ground data, and a stable transmission mechanism throughout the entire process, the system effectively solves the problems of difficult collection, transmission, and low accuracy of meteorological data in forest areas. Ultimately, it achieves intelligent and comprehensive collection of multimodal meteorological factors in forests, providing data support for forest fire risk early warning, ecological environment protection, carbon sink monitoring, and refined forestry management.

[0041] The distributed forest data transmission network constructed in this embodiment adopts a deep integration mode of LoRa self-organizing communication network and 5G long-distance communication network. The entire network adopts a unified star topology layout, which is the core carrier for realizing efficient collection and stable transmission of multimodal meteorological factors in air, space and ground. The two communication modules have clear division of labor and work together to perfectly adapt to the application scenarios of complex forest terrain, large-area coverage and low power consumption transmission.

[0042] Specifically, this step involves the synchronous and collaborative collection of data through the following five units.

[0043] (I) Core architecture and functions of distributed forest data transmission network.

[0044] The distributed forest data transmission network constructed in this embodiment adopts a deep integration mode of LoRa self-organizing communication network and 5G long-distance communication network. The entire network adopts a unified star topology layout, which is the core carrier for realizing efficient collection and stable transmission of multimodal meteorological factors in air, space and ground. The two communication modules have clear division of labor and work together to perfectly adapt to the application scenarios of complex forest terrain, large-area coverage and low power consumption transmission.

[0045] (1) Functions and layout of LoRa self-organizing communication network.

[0046] LoRa self-organizing communication networks are specifically designed for local data communication within forests. Addressing the challenges of dense vegetation, complex terrain, severe signal obstruction, and weak coverage of traditional communication networks within forests, they enable short-range, low-power, and wide-area wireless transmission of meteorological data. In a star topology, the LoRa main gateway, deployed in a relatively open area within the core of the forest, serves as the network's central node. Various meteorological data acquisition terminals within the forest act as terminal sub-nodes, directly establishing wireless communication connections with the LoRa main gateway. This eliminates the need for multiple relays, reducing data transmission loss and latency, and ensuring stable communication within the forest.

[0047] The network has four core functions: ①Low power consumption and long distance transmission: Using LoRa spread spectrum communication technology, a single gateway can achieve signal coverage of 3-15km in forest areas, support long-term battery power supply for terminal devices, and meet the continuous communication needs of outdoor devices without external power supply.

[0048] ② Multi-node concurrent access can simultaneously handle data upload requests from hundreds of forest data collection terminals, adapting to the multi-point, high-density meteorological data collection layout in forest areas and avoiding data transmission conflicts.

[0049] ③ Local data caching: When the network signal in the forest area is interrupted, the meteorological data collected by the terminal is automatically cached and automatically retransmitted after the communication is restored, thus preventing data loss.

[0050] ④ Flexible network expansion: As the monitoring range of the forest area expands, new LoRa sub-gateways can be added and connected to the main gateway to quickly expand the network coverage and adapt to the monitoring needs of forests of different sizes.

[0051] (2) Functions and layout of 5G long-distance communication network.

[0052] A 5G long-range communication network is deployed in the edge of forest areas, establishing 5G core gateways and base stations as the backbone transmission layer of the distributed transmission network. It receives forest meteorological data aggregated by the LoRa self-organizing communication network, enabling long-distance, high-speed, and high-capacity wireless data transmission and establishing a transmission channel between forest data and the backend data management platform. In a star topology, the 5G core gateway directly interfaces with the LoRa main gateway, acting as the upper-level aggregation node of the entire distributed network. It uniformly receives and integrates all ground meteorological data transmitted by the LoRa network, while also undertaking some direct communication tasks for mobile data acquisition devices. The core functions of this network are as follows:

[0053] ① Ultra-long-distance high-speed transmission: Relying on the technical advantages of 5G's large bandwidth, low latency, and wide coverage, massive amounts of multimodal meteorological data collected in forest areas can be transmitted without loss to monitoring centers tens or even hundreds of kilometers away, breaking through the transmission limitations caused by the remote geographical location of forests.

[0054] ② Multi-source data integration and uploading: it can synchronously receive aggregated data from LoRa network, real-time data transmitted by unmanned aerial vehicles, and docking data from satellite remote sensing, realizing unified packaging and encrypted transmission of multi-type data.

[0055] ③ Seamless switching of communication signals: it supports automatic switching between LoRa and 5G networks for mobile collection equipment in forests. When the equipment is close to the core forest area, it will automatically access the LoRa network, and when it is close to forest edges or open areas, it will automatically switch to 5G network, so as to ensure the continuity of data transmission.

[0056] ④ Data security encryption: it encrypts the transmitted meteorological data to prevent data from being tampered with or stolen, and ensures the security and confidentiality of forest meteorological monitoring data.

[0057] (II) Fixed multi-factor meteorological collection station.

[0058] The fixed multi-factor meteorological collection station is the core terminal for ground meteorological data collection in forest areas. It is deployed at key monitoring points with different altitudes and different stand types in forests, including ridges, valleys, forest edges, high fire risk areas, ecologically sensitive areas, etc., so as to realize all-weather, continuous and stable collection of basic meteorological factors in forest areas. The collection station has built-in high-precision meteorological sensors, which can synchronously collect multi-dimensional meteorological factors including temperature, humidity, wind speed, wind direction, air pressure, precipitation, light intensity, soil temperature and humidity, and moisture content of surface combustibles, and is equipped with a Beidou positioning module to facilitate linkage monitoring with the mobile meteorological data calibration platform. On this basis, the equipment is equipped with a LoRa communication module, which serves as a basic terminal sub-node of the star topology and directly establishes a communication connection with the LoRa main gateway. Its function realization completely relies on the LoRa ad hoc communication network: the collected reference meteorological data is uploaded to the LoRa main gateway in real time through the LoRa network, and then uniformly transmitted by the main gateway to the 5G core gateway, and finally remotely pushed to the data monitoring center through the 5G network; parameters such as equipment operation status, sensor power, and collection frequency can be下发 reversely through the distributed network to realize remote monitoring and parameter adjustment; in response to the long-term monitoring demand at fixed points, relying on the low power consumption feature of LoRa, the equipment does not need frequent battery replacement, and can achieve continuous and uninterrupted collection for 6 to 12 months, providing high-precision and high-stability ground reference data for the entire meteorological monitoring system, and making up for the shortcoming that space-based and air-based monitoring cannot achieve long-term accurate observation at a single ground point.

[0059] (III) Mobile meteorological data calibration platform.

[0060] The mobile meteorological data calibration platform is a core device for ensuring the data accuracy of the entire monitoring system. Utilizing portable monitoring kits and field monitoring vehicles, it boasts flexible deployment advantages and is accessible across all regions. It is primarily used to correct meteorological data acquisition errors caused by sensor drift, terrain obstruction, and vegetation interference in complex forest environments, enabling on-site calibration and accuracy verification of data collected by fixed multi-factor meteorological data collection stations and UAV monitoring units. The platform is equipped with high-precision calibration sensors and dual communication modules (LoRa+5G), allowing for flexible switching to a distributed transmission network: when operating inside the forest, it automatically connects to the LoRa self-organizing communication network, uploading the on-site collected calibration reference data and error correction parameters to the LoRa main gateway in real time, and simultaneously distributing them to surrounding fixed meteorological data collection stations and host computer terminals for real-time data correction; when operating at the forest edge, during long-distance mobile relocation, or when high-speed transmission of calibration data is required, it automatically switches to the 5G communication network to directly communicate with the backend data platform, uploading complete calibration reports and correction coefficients, while simultaneously receiving calibration task instructions from the platform. Its core functions are reflected in the following five aspects.

[0061] (1) On-site dynamic calibration can be carried out in forest areas without fixed base stations and in remote areas, and various data collection terminals can be calibrated on-site to solve the problem that field equipment cannot be regularly retrieved and calibrated.

[0062] (2) Cross-terminal data correction: The calibration data is distributed across the entire domain through a distributed network, and the data collected by multiple fixed collection stations and UAVs are corrected simultaneously, forming a closed loop of "collection-calibration-correction".

[0063] (3) Emergency supplementary data transmission: When the local LoRa network fails or the data of the fixed collection station is interrupted, it can be used as a temporary collection terminal to directly transmit emergency meteorological data through the 5G network to ensure that the monitoring data is not interrupted.

[0064] (4) Equipment status inspection: rely on the communication network to obtain the operating status of all data collection terminals in the forest area, quickly locate faulty equipment, and improve the system operation and maintenance efficiency.

[0065] (5) Meteorological changes that fill gaps between various fixed equipment.

[0066] (iv) Forest near-ground drone monitoring unit.

[0067] Forest meteorological monitoring is greatly affected by its internal geographical structure and forest stand composition. The forest near-ground UAV monitoring unit is the key to realizing low-altitude, three-dimensional and areal meteorological monitoring in forest areas. While identifying geographical structure and forest stand type, it makes up for the shortcomings of fixed collection stations in single-point monitoring, limited coverage and inability to monitor canopy micro-meteorology. It is responsible for collecting multimodal meteorological factors in forest canopy, vertical stratification in forest, and remote and complex terrain areas.

[0068] The drone is equipped with lightweight meteorological sensors, high-definition imaging equipment, and a LoRa+5G dual communication module. Following preset routes or ad-hoc commands, it conducts near-ground low-altitude patrols, collecting multi-dimensional data such as canopy temperature, canopy humidity, vertical temperature and humidity profiles within the forest, vegetation growth, hotspot hazards, and local microclimates. Its communication transmission and functionality rely on precise coordination through a distributed network: when flying at low altitudes within the forest, it prioritizes access to the LoRa self-organizing communication network, transmitting low-altitude meteorological data in real time for stable short-range transmission and reduced flight energy consumption; when flying to the forest edge, high-altitude open areas, or when high-definition, large-capacity data transmission is required, it automatically switches to the 5G network for high-speed transmission of complete monitoring data and video footage.

[0069] In a star topology, the drone acts as a mobile terminal sub-node, with its communication timing uniformly scheduled by the LoRa master gateway and the 5G core gateway to avoid data conflicts with other terminals. The core functions include the following four points.

[0070] (1) Three-dimensional supplementary measurement covers all areas, realizing vertical meteorological monitoring from the forest ground to the canopy and horizontal surface monitoring from a single point to a region, covering areas such as mountains and dense forests that cannot be reached by manpower and fixed equipment.

[0071] (2) Real-time data linkage: The collected airborne meteorological data is synchronously transmitted to the data platform through a distributed network and integrated with the ground fixed station data and calibration data in real time.

[0072] (3) Flexible scheduling of operations: the backend platform issues flight routes and data collection instructions through the 5G network, and realizes real-time data interaction during flight through the LoRa network, supporting temporary cruise data collection in emergency situations.

[0073] (4) Collaborative calibration operation: receive correction parameters from the mobile calibration platform, optimize its own acquisition accuracy in real time, and ensure consistency between airborne and ground-based data.

[0074] (v) Satellite remote sensing data acquisition unit.

[0075] The satellite remote sensing data acquisition unit is the core space-based system for large-scale, comprehensive meteorological monitoring of forest areas. Relying on high-resolution meteorological satellites and ecological environment monitoring satellites, it constructs a space-based monitoring network responsible for collecting macroscopic meteorological factors and ecological environment parameters across the entire forest area. This complements ground-based and low-altitude monitoring, achieving integrated space-air-ground coverage. This unit does not directly connect to the LoRa self-organizing communication network. Instead, it transmits the collected remote sensing data in real-time to the satellite data ground receiving station via a dedicated satellite data transmission channel. The data is then connected to the distributed forest data transmission network via the internet and 5G communication networks, pushing the data to the backend data management platform, thus completing cross-network integration with ground-based and low-altitude data. Its core functions emphasize macroscopic monitoring and data complementarity.

[0076] (1) Large-scale full-area monitoring breaks through the limitations of ground and low-altitude monitoring range, and realizes the collection of macro meteorological factors such as surface temperature, vegetation coverage, soil moisture, cloud cover, precipitation distribution, and large-scale meteorological background field of the entire forest or even the regional forest area.

[0077] (2) Spatiotemporal data matching: The collected space-based remote sensing data is spatiotemporally registered and fused with the precise point data and UAV low-altitude surface data transmitted by LoRa on the ground via 5G network.

[0078] (3) Long-term dynamic monitoring provides periodic and comprehensive remote sensing data on forest meteorology and ecology, providing macro data support for forest ecological changes, long-term meteorological pattern analysis, and large-scale fire risk warning.

[0079] (4) Complete data coverage: For areas with weak communication and insufficient ground acquisition equipment in local forest areas, satellite remote sensing data is used to supplement and improve the data, so as to achieve meteorological monitoring without blind spots.

[0080] (vi) Integrated air-space-ground multimodal meteorological factor collection and collaborative operation mechanism.

[0081] Relying on the star topology scheduling capability of the distributed forest data transmission network, the four major acquisition units—fixed multi-factor meteorological acquisition stations, mobile meteorological data calibration platform, forest near-ground UAV monitoring unit, and satellite remote sensing data acquisition unit—form an efficient and collaborative acquisition mechanism to achieve full-area, all-time, and high-precision acquisition of multimodal meteorological factors.

[0082] (1) Layered transmission and unified scheduling: The LoRa network carries short-range data transmission of fixed, mobile and UAV terminals in the forest, while the 5G network undertakes long-range backbone transmission and satellite data docking. The star topology structure uniformly allocates communication channels and schedules data upload timing to avoid data transmission conflicts between multiple terminals and ensure smooth communication across the entire network.

[0083] (2) Multi-source data closed-loop calibration: Fixed acquisition stations provide reference data, UAVs complete three-dimensional supplementary acquisition, mobile calibration platforms correct errors on-site, and calibration parameters are synchronized to all acquisition terminals through a distributed network, forming a data closed loop of "acquisition-transmission-calibration-correction-reacquisition" to comprehensively improve data accuracy.

[0084] (3) Complementary integration of air, space and ground data: Ground-based terminals achieve high-precision point monitoring, air-based UAVs achieve three-dimensional surface monitoring, and space-based satellites achieve macro-domain monitoring. The three types of data are converged and integrated through LoRa+5G network to generate a multi-dimensional, high spatiotemporal resolution forest multimodal meteorological dataset.

[0085] (4) Stable and reliable transmission throughout the entire process: LoRa network ensures low-power and stable communication in complex forest environments, 5G network enables long-distance and high-speed transmission, seamless switching between the two networks, combined with data caching, encryption and retransmission mechanisms, completely solves the pain points of difficult collection, difficult transmission and low accuracy of meteorological data in forest areas.

[0086] The entire system uses a distributed forest data transmission network as its communication core, fully leverages the technological advantages of LoRa and 5G networks, accurately adapts to the functional requirements of each acquisition unit, and ultimately achieves intelligent acquisition of forest meteorological factors in an integrated, multimodal, and comprehensive manner, providing all-round and high-precision data support for forest fire early warning, ecological environment protection, carbon sink monitoring, and refined forestry management.

[0087] Step 2: Data preprocessing and decision-level information fusion.

[0088] This step is performed using the standard database and real-time acquired data from Step 1. It aims to address the issues of insufficient data fusion, consistency, and completeness in existing technologies. Through preprocessing and decision-level fusion, it generates a high-quality multi-source data collection.

[0089] (a) Outlier removal algorithm.

[0090] First, data preprocessing is performed to clean meteorological, satellite, and remote sensing data: an outlier removal algorithm (3σ criterion) is used to automatically identify and remove outliers and duplicates from multi-source data. The Laida criterion, also known as the 3σ criterion or PauTa criterion, is the most commonly used outlier removal algorithm in engineering and data analysis. It relies on the probability characteristics of the normal distribution to identify outliers caused by gross errors. It is simple to calculate, highly practical, and widely used for data cleaning in scenarios such as sensor data, experimental measurements, and environmental monitoring. The principle of the outlier removal algorithm (Laida criterion) is shown below.

[0091] (1) Algorithm principle.

[0092] ① The underlying theory of this algorithm: If a set of random measurement data strictly follows a normal distribution X ~ N ( µ , σ 2 ); data in [ µ - σ , µ + σ The probability of the interval is 68.27%, and the data is in [...]. µ -2 σ , µ +2 σ The probability of the interval is 95.45%, and the data is in [...]. µ -3 σ , µ+3 σ The probability of the interval is 99.73%.

[0093] ② Exception detection logic refers to exceeding... µ ±3 σ The data in this range has a natural probability of only 0.27%, which is an extremely low probability event. This type of data is generally caused by gross errors such as human mistakes, equipment failures, and environmental interference, and is defined as an outlier that needs to be removed.

[0094] ③ In real-world scenarios, the overall true mean µ Overall standard deviation σ Unable to obtain, use the sample mean instead. Unbiased standard deviation of the sample s Replace the completion of the judgment.

[0095] (2) Complete formula.

[0096] Given the original dataset: x 1, x 2, x 3, ..., x n ,in n This represents the number of samples.

[0097] ① Sample mean formula.

[0098] Reflecting the overall centrality of the data: .

[0099] ② Sample unbiased standard deviation (Bessel's formula).

[0100] To measure the dispersion of data, the Laida algorithm must use... n -1 unbiased estimate: .

[0101] ③ Single-point residual formula.

[0102] Deviation of a single data point from the overall mean: .

[0103] ④ Core discrimination formula (threshold rule): .

[0104] Equivalence check: Satisfies This point is an outlier and should be removed; it meets the following requirements. This data point is normal and should be retained.

[0105] (3) Standard algorithm execution steps.

[0106] ① Input the original dataset and calculate the overall sample mean. Sample standard deviation ② Calculate the residuals of all data point by point. ③According to ④ Filter outliers, removing only the outlier with the largest deviation each time; ⑤ After removing outliers, update the mean and standard deviation, and iterate twice; ⑥ Iterate repeatedly until all data meet the criteria. The algorithm terminates.

[0107] In this algorithm, multiple outliers cannot be removed in batches at once, and statistical parameters must be updated iteratively; otherwise, false positives and false negatives will occur. Relaxed optimization: 2 is used in some scenarios. σ As a threshold, it is suitable for mild noise reduction and lenient anomaly detection; compared with similar algorithms: for small samples, the Grubbs criterion and Dixon criterion are preferred, while for large samples, the Laida criterion is preferred.

[0108] (ii) Linear interpolation algorithm.

[0109] Missing data is filled using linear interpolation. The principle and complete formula of the linear interpolation algorithm are shown below.

[0110] (1) Algorithm principle.

[0111] Given two discrete sampling points ( x 0, y 0), ( x 1, y 1) Assuming the function change between two points follows a linear relationship, the original continuous function is approximated by the straight line connecting the two points, thus calculating the interval [ x 0, x Any unknown x-coordinate within 1] x corresponding ordinate y The algorithm essentially combines the fitting of straight lines between two points with the weighted ratio calculation, making it suitable for scenarios with gradual data changes. Its extremely simple calculation and fast processing speed are applied in this embodiment to complete unexpectedly missing data.

[0112] (2) Complete formula.

[0113] ① Slope derivation formula: .

[0114] ② Weighted proportional formula: The two points are weighted inversely proportional to their distance: .

[0115] ③ Symmetrical proportion: .

[0116] In this algorithm ( x 0, y 0), ( x 1, y 1) Given two known data points, andx 0≠ x 1; x The x-coordinate of the target to be solved (generally required) x 0< x < x 1. For interpolation, use interpolation; for interpolation outside the interval, use extrapolation, which increases the error. y The unknown value is obtained through linear interpolation.

[0117] (iii) Neighborhood mean filling algorithm.

[0118] For missing areas in satellite imagery and near-ground UAV data, a neighborhood mean imputation method is used to fill in the gaps. The principle and complete formula of the neighborhood mean imputation algorithm are shown below.

[0119] (1) Algorithm principle.

[0120] Data restoration for missing or anomalous meteorological data from satellite remote sensing imagery and UAV near-ground gridded meteorological data is a fundamental spatial interpolation algorithm for integrated air-space-ground meteorology, forest combustible moisture content, and surface environmental parameter detection. The core theoretical support of the neighborhood mean filling algorithm is based on generating supplementary information from the spatiotemporal proximity of meteorological elements within geographic pixels and monitoring grids, achieving data filling and bias correction. Centered on the missing pixel or measuring point, a fixed-range neighborhood window is extracted, all valid observation data within the window are filtered, the missing value is estimated using the arithmetic mean, and the filling is completed globally point-by-point by sliding the window.

[0121] (2) Symbol definition.

[0122] In this embodiment, satellite imagery and discrete UAV measurement points are uniformly mapped into a two-dimensional spatial raster matrix, and standard mathematical notation is established: ① For global data matrix; M Vertical spatial sampling number / number of image rows; N The number of horizontal spatial samples / number of image columns; X i,j The original observation value at coordinates (i, j).

[0123] ② Missing mask matrix Mask Used to distinguish between valid and missing data: .

[0124] A value of 1 indicates valid normal meteorological or satellite remote sensing data; A value of 0 indicates missing or abnormal data.

[0125] ③ Targets to be repaired: All targets that meet the requirements A spatial unit of 0 ( i , j).

[0126] ④ Neighborhood window: based on missing points ( i , j Centered on , the window size is (2k+1)×(2k+1). Where k is the neighborhood radius, k=1 corresponds to a 3×3 window, and k=2 corresponds to a 5×5 window (generally applicable to meteorological remote sensing).

[0127] ⑤ Basic neighborhood coordinate set: .

[0128] (3) Complete formula.

[0129] ①Statistical analysis of valid data.

[0130] Total number of valid data within the neighborhood: .

[0131] ② Standard filling formula (non-boundary region).

[0132] Missing location ( i , j Repair estimate : .

[0133] It is a subset of valid data coordinates within the neighborhood.

[0134] ③ Boundary correction formula (image or measurement point edge adaptation).

[0135] Window overflows may occur at the edges of satellite and drone monitoring areas, requiring the trimming of legal neighboring areas. .

[0136] Final filling formula for the boundary region: .

[0137] ④ Extreme scenario constraints.

[0138] If there is no valid data in the neighborhood n i,j =0, automatically expand the neighborhood radius. k The calculation is iterated until a valid sample is obtained.

[0139] (4) Standard algorithm execution steps.

[0140] ① Rasterize satellite and UAV data to construct a two-dimensional parameter matrix; ② Detect anomalies and generate a missing mask matrix; ③ Traverse all missing cells, delineate the central neighborhood window and perform boundary clipping; ④ Filter valid measurement points and image data within the window; ⑤ Substitute the mean formula to calculate the repair value and replace the original missing value; ⑥ Output a complete and continuous meteorological remote sensing dataset.

[0141] (iv) Z-score data standardization algorithm.

[0142] The Z-score standardization algorithm is used to transform multi-source data of different units and magnitudes into a unified standard (mean of 0 and standard deviation of 1), eliminating the influence of dimensions. Z-score standardization, also known as standard deviation standardization or zero-mean standardization, is the most commonly used linear standardization method in machine learning, remote sensing data analysis, and multi-source meteorological data fusion (satellite imagery + UAV near-ground monitoring).

[0143] (1) Algorithm principle.

[0144] The Z-score data standardization algorithm can solve the problem of inconsistent units and large numerical ranges of multi-source heterogeneous parameters such as combustible material moisture content, air temperature, humidity, remote sensing reflectance, and surface temperature in this embodiment, and provides data preprocessing for neighborhood filling, deep learning, time series prediction, and multi-factor coupling analysis. Its principle is as follows.

[0145] ① Centralization: Subtract the mean of the data in this dimension from the original data to eliminate the overall data bias and make the mean of the dataset zero; ② Scale normalization: Divide the centered data by the standard deviation of the data in this dimension to eliminate the differences in the dimensions, orders of magnitude, and dispersion of different indicators and make the standard deviation of the dataset 1; ③ The transformed data follows a standard normal distribution N(0,1), retaining the distribution characteristics, relative relationships, and degree of abnormal deviation of the original data, without changing the internal correlation of the data.

[0146] (2) Symbol definition.

[0147] Let the original dataset of the features be... X ={ x 1, x 2, x 3, ..., x n}; x i For the first i One original sample value; n This represents the total number of samples. μ This is the population mean; The mean of the sample; σ The population standard deviation; s The standard deviation is the sample standard deviation. z i This is the output value after Z-score normalization.

[0148] (3) Complete formula.

[0149] ① The overall Z-score is suitable for scenarios where the data covers the entire research population and there is no sampling error: .

[0150] In the formula: , .

[0151] ② The sample Z-score is suitable for measured data. In practical applications, only a portion of the sampled data can be obtained, so the unbiased sample standard deviation can be used. .

[0152] In the formula: , .

[0153] (4) Standard algorithm execution steps.

[0154] ① For a single indicator (such as meteorological data, moisture content of combustibles, or satellite band reflectivity), statistically analyze all samples; ② Calculate the sample mean of that indicator. ③ Calculate the sample standard deviation of this indicator. s ④ Substitute each sample into the formula, subtract the mean from the original value, and divide by the standard deviation to obtain the standardized value. z i ⑤ Standardize independently by dimension (calculate the mean and standard deviation separately for different meteorological parameters and remote sensing bands), and prohibit cross-dimensional mixed calculations.

[0155] (v) Gaussian filtering noise reduction algorithm.

[0156] (1) Algorithm overview and core positioning.

[0157] Satellite imagery undergoes additional noise reduction using a Gaussian filtering algorithm to improve data purity. Gaussian filtering is a linear low-pass spatial filtering algorithm based on a two-dimensional Gaussian normal distribution, and it is a core algorithm for noise reduction and purification of remote sensing satellite imagery, UAV near-Earth meteorological raster data, and environmental monitoring time-series data. The difference between this algorithm and traditional simple mean filtering is that the latter uses equal-weighted forced smoothing, while Gaussian filtering employs a distance-weighted smoothing mechanism, reasonably completing the noise reduction task without affecting the data's accuracy. Its core functions are shown below.

[0158] ① Directional suppression of Gaussian white noise, sensor random jitter noise, isolated bad spots in remote sensing images, and local high-frequency anomalies; ② Preservation of the low-frequency true spatial distribution characteristics and gradient changes of surface and meteorological elements; ③ Smoothing out ineffective random errors and weakening discrete outliers, significantly improving the purity, spatial continuity, and reliability of the original observation data; ④ Adaptable to multimodal data fusion preprocessing, often used in conjunction with missing value imputation, Z-score normalization, and Laida anomaly detection.

[0159] (2) Core noise reduction principle.

[0160] ① Spatial correlation weighting: In geospatial data and remote sensing rasters, the closer the central pixel / monitoring point is to its surrounding neighboring points, the stronger the data correlation and the higher the proportion of effective information; the farther away they are, the weaker the correlation and the higher the proportion of interference noise. The Gaussian function naturally satisfies the condition that the closer to the center, the greater the weight; and the farther away they are, the more the weight decays exponentially.

[0161] ② Low-pass noise reduction mechanism: Noise is generally a high-frequency abrupt signal (single-point numerical abrupt change, no spatial continuity), while real meteorological / remote sensing data is a low-frequency continuous signal; Gaussian filtering filters high-frequency noise components through neighborhood weighted convolution, retains low-frequency effective physical information, and mathematically removes random errors to achieve data purification.

[0162] ③ Controllable smoothing intensity: through standard deviation σ Adjust noise reduction level flexibly: σ The larger the value, the wider the weight diffusion range and the stronger the smoothing and noise reduction. σ The smaller the size, the better the detail is preserved, with weak noise reduction.

[0163] (3) Complete formula.

[0164] ① One-dimensional Gaussian function (theoretical basis).

[0165] .

[0166] In the formula x This is the offset distance from the center pixel; σ The standard deviation is Gaussian, and it is the core control parameter for noise reduction intensity.

[0167] ② Two-dimensional Gaussian function (core of remote sensing / raster data).

[0168] Satellite imagery and UAV-based two-dimensional gridded meteorological data both employ a two-dimensional Gaussian distribution, which serves as the theoretical prototype for Gaussian filtering. .

[0169] In the formula, ( x , y () represents the horizontal and vertical offset coordinates of any pixel within the neighborhood relative to the central target pixel; It is a natural exponential function; 2 πσ 2 The normalization coefficients ensure that the sum of the weights across the entire domain is 1.

[0170] Based on this, the forest meteorological factors and land cover standard database constructed in step one is invoked, and the preprocessed real-time data is matched with the historical data in the database.

[0171] (vi) Decision-level data fusion.

[0172] Finally, a decision-level information fusion algorithm is employed to fuse real-time and historical data. This embodiment utilizes a Convolutional Neural Network (CNN) fusion algorithm. First, the credibility of each source data is assessed (based on data acquisition accuracy and transmission stability). Then, through evidence synthesis rules, the feature information of the multi-source data is fused, redundant information is removed, and missing data is supplemented, generating a multi-source data set with the same time series and a unified format. During the fusion process, a dynamic weight adjustment strategy is employed, adjusting the fusion weights of each source data in real time based on the operational status feedback from subsequent model computation layers, thereby improving the quality of the fused data.

[0173] This embodiment aims to enhance the correlation between meteorological data and forest combustible moisture content. By establishing a mapping relationship between the two data, key meteorological features at each moment are identified, thereby improving the data quality of forest surface meteorological factors and forest combustible moisture content while ensuring the continuity of the time series. Figure 2 The convolutional neural network structure shown in the diagram uses convolutional layers and pooling layers as key steps in feature generation. Figure 2 The first layer is the input data layer. Convolutional operations are used to obtain features for the second layer, and pooling is then performed on the second layer's feature map to obtain features for the third layer. This process is repeated to obtain connection vectors, which are then fed into the fully connected layer. The fully connected layer can be viewed as a negative feedback neural network used for further processing and feature generation. In the entire network structure, each feature data point can be considered as neurons arranged in a matrix. These neurons perform feature recognition and fusion at different levels of the input information, thus achieving effective processing and learning of complex data. Convolutional neural networks effectively reduce the number of parameters and computational complexity through local feature recognition and weight sharing mechanisms, while maintaining good model performance.

[0174] In this embodiment, the convolutional layer uses convolutional kernels to identify the input data features of meteorological factors. The convolution operation can be viewed as sliding a convolutional kernel window across the input data sequence and calculating the weights of the portion covered by the window.

[0175] A data matrix consisting of input meteorological factor data I and a convolution kernel K The convolution operation can be represented as: .

[0176] In the formula F ( i , j ) is the output feature at position ( i , j The value of ); I ( i + m , j +n ) indicates the location of the input data ( i + m , j + n The window value; K ( m , n ) is the location of the convolution ( m , n The weight of ) m and n This represents the size of the convolution kernel. The convolution process is as follows: Figure 3 As shown.

[0177] The pooling layer is located after the convolutional layer. It reduces the spatial size of the feature matrix by downsampling, thereby reducing the amount of computation and preventing overfitting. The most commonly used pooling methods in this layer are max pooling and average pooling.

[0178] Max pooling involves selecting the maximum value within the pooling window as the output. Let the size of the pooling window be... p × p Max pooling can be expressed as: .

[0179] F ( i , j ) is the location of the feature after pooling ( i , j The value of ); Average pooling: calculates the average of all values ​​in the pooling window. For average pooling, the output can be represented as: .

[0180] In the formula F ( i , j ) is the feature map after pooling at position ( i , j The value of ). Average pooling calculates the average of all values ​​in the pooling window. The pooling process is as follows: Figure 4 As shown.

[0181] Among the two pooling methods mentioned above, max pooling, which selects the maximum value within the coverage area as the output, is one of the most commonly used pooling methods. It tends to preserve detailed data and edge information, maintains the data magnitude of the features unchanged, and helps reduce noise by preserving the time series, thus reducing the computational requirements of fully connected layers. Average pooling, on the other hand, calculates the average of all values ​​within the coverage area as the output. This pooling method tends to smooth features. Compared to max pooling, average pooling retains more background information to some extent, and by averaging the input features, it can avoid overfitting the model to the training data to some extent. In summary, max pooling is more suitable for capturing significant features in the data and is beneficial for data fusion to generate feature data. Therefore, this embodiment uses max pooling for decision-level data fusion.

[0182] Step 3: Prediction model construction, optimization and computation.

[0183] This step receives the multi-source data set generated in step two and aims to address the problems of poor adaptability and low accuracy of existing prediction models. By constructing and optimizing a deep learning model based on Bi-LSTM, high-precision prediction can be achieved.

[0184] Model Construction and Training: A Bi-LSTM deep learning algorithm was used to construct a moisture content prediction model, taking into account the variation pattern of moisture content in small dead combustibles on the forest surface (the characteristic of being affected by multiple factors such as temperature, humidity, wind speed, and light). The model input layer incorporates multi-dimensional data including meteorological factors, satellite-extracted features, and remote sensing-extracted features. A 6-layer encoder is set in the hidden layer, and a self-attention mechanism is used to capture the synergistic influence relationship between multiple factors. The output layer is the predicted moisture content value. The model training uses the gradient descent algorithm (Adam optimizer), with the measured moisture content by the drying method as the label. Iterative training is performed using historical multi-source data (1000 iterations, learning rate 0.001). An early stopping strategy (training stops if the validation set error does not decrease for 10 consecutive iterations) is used to avoid overfitting and ensure the model training accuracy.

[0185] Parameter optimization: The Multiverse of Elements (MVO) algorithm is used for dynamic optimization of model parameters, including the number of encoder layers, the number of attention heads, the learning rate, and the regularization coefficient. To address the environmental differences in different forest regions, factors such as forest stand type and terrain slope are introduced as adjustment parameters to establish an adaptive parameter adjustment model, achieving generalization and adaptation of the model across various forest region scenarios. The model parameters are periodically (once a month) recalibrated using newly added measured data to further reduce prediction errors.

[0186] Prediction calculation: A dual-mode control method of batch calculation + real-time calculation is adopted. Batch calculation is used for large-scale forest areas, using GPU parallel computing to accelerate the calculation and ensure efficiency; real-time calculation is used for key prevention and control areas, using CPU fast calculation mode with a response time of ≤30 seconds. During the calculation process, an error verification algorithm is built in to calculate the deviation between the prediction result and the historical measured data. If the deviation exceeds a preset threshold (5%), the model parameters are automatically re-optimized to ensure prediction accuracy.

[0187] (a) Construction of Bi-LSTM time series prediction model.

[0188] Long Short-Term Memory (LSTM) neural networks are recurrent neural networks suitable for time series modeling and prediction. Compared to traditional measurement methods such as Back Propagation Neural Networks and Random Forest algorithms, this network introduces the ability to characterize time series, establishing a connection between current input factors and historical information. This enhances the correlation of the input sequence while improving prediction accuracy. It is suitable for measuring the moisture content of fine dead combustibles on the forest surface, which exhibits dynamic changes in environmental factors at different times. The structure of an LSTM neural network is as follows: Figure 5 As shown.

[0189] Figure 5 middle , , for , , Meteorological factor input data at any given time; , , for , , Predicted output at time step; , for , Output memory unit at any time (moisture content of fine dead combustibles on the forest surface); , These are the forget gate and the output gate, respectively. , To update the door; This represents the activation function, whose output is between 0 and 1; It is the hyperbolic tangent function, and its output is... arrive between, and Representing neurons is the weight matrix. With bias term The way various gate systems work together in a network. The operation of a Long Short-Term Memory (LSTM) neural network begins with the forgetting gate. Initially, the purpose was to Spectrum of time - meteorological data and The moisture content at each moment is combined to make the calculation results... to The calculation principle between them is shown in the following formula, where and Here are the weight matrix and bias terms for the forget gate.

[0190] .

[0191] In obtaining The data was then processed Unit and Multiplying digit by digit, when One of them is At that time, the opposite position This is invalid data, which is precisely the purpose of the forget gate. Meanwhile, the update gate has two parts. and right The input at any given time is effectively preserved and normalized, and its working principle is shown in the following formula: .

[0192] .

[0193] As can be seen from the current step, by updating the door pair Meteorological data at any time and After processing the moisture content at a given time... Unit and Adding them together creates a new memory unit. The calculation principle is shown in the following formula: .

[0194] In the final output gate section go through Processing yields a result in to The value between, and and Unit output Multiplying these values ​​yields the desired moisture content output. The calculation principle is illustrated in the following formula, which shows the output moisture content. This information will be used as a condition for the network's operation in the next moment during training; this is the temporal characteristic of Long Short-Term Memory (LSTM) neural networks.

[0195] .

[0196] .

[0197] Although LSTM can effectively learn historical information and long-term dependencies in time series forecasting, its information transmission direction is usually "past → present," meaning the model at time step 1 is not directly related to the previous information. The hidden representations primarily rely on previous observations within the historical window. In sequences with obvious periodicity, stage-like characteristics, or local structural features, unidirectional information alone may be insufficient to fully characterize the sequence morphology. Therefore, this embodiment further introduces a Bi-LSTM structure, simultaneously performing forward and backward modeling within the input window and fusing the hidden states from both directions to enhance the representation and prediction capabilities of sequence contextual information.

[0198] Bi-LSTM neural networks are an extension of LSTM neural networks. By performing two rounds of deep learning on the same input sequence, they can effectively improve the ability to model complex time series patterns, thereby increasing the accuracy of model predictions. Bi-LSTM neural networks, for example... Figure 6 As shown.

[0199] Bi-LSTM builds upon LSTM by incorporating two separate hidden layers: a forward hidden layer for processing sequential data in chronological order and a backward hidden layer for processing sequential data in reverse chronological order. The hidden states from both directions are then summed and fed into the same output layer to obtain the final prediction. The computation process for both the forward and backward directions in Bi-LSTM is identical to that of LSTM; therefore, simply fusing and summing the predictions from both directions yields the final Bi-LSTM neural network prediction output. .

[0200] (ii) Multiverse Optimization Algorithm.

[0201] In the training process of the Bi-LSTM neural network, this embodiment introduces the MVO (Multi-Verse Optimizer) algorithm. This algorithm utilizes a "white hole-black hole-wormhole" mechanism to achieve information exchange and search updates among different populations, thereby finding the globally optimal solution. The multiverse optimization algorithm is inspired by the multiverse theory in physics, which posits the existence of other universes besides the known ones. It solves the optimization problem through the "white hole-black hole-wormhole" mechanism. The algorithm structure is as follows: .

[0202] Where d represents the quantity in the variables, and n represents the number of universes.

[0203] .

[0204] in For the first The first universe The location of the black hole For the first The location of a universe It is the first The standard expansion rate of the universe. This represents the Kth universe generated through a roulette wheel betting mechanism. The location of a black hole. Therefore, the location of a black hole in the optimal universe can be obtained: .

[0205] There are two key control curves: The probability of wormhole existence (gradually increasing with iterations: use wormholes sparingly in the early stages to avoid premature convergence; use them more frequently in the later stages to accelerate fine-grained search) and The travel distance rate (gradually decreases with iteration: large perturbation step size in the early stage, small step size in the later stage). The two are complementary. In the early stage, large step size and few wormholes are needed to continuously expand and explore the universe. In the later stage, small step size and more wormholes are used to refine the optimal solution.

[0206] It is a dynamic parameter. The threshold value is taken here. Experience points. and This indicates the upper and lower bounds of the dimension of this universe. , for The current optimal black hole position is obtained by using random numbers within the range. As the algorithm continues to evolve and iterate to find the optimal position, if another optimal black hole position is found... Afterwards, and The matching and adaptation abilities of the black holes are compared, and the one with better fitness is selected as the next generation of new black hole individuals.

[0207] Wormhole Existence Rate and travel distance rate The adaptive formula is as follows: .

[0208] in It is the minimum value, which is 0.2. It is the maximum value, which is 1. Indicates the current iteration number. This indicates the maximum number of iterations.

[0209] .

[0210] p Defined as the utilization accuracy during iterations, the higher the value, the faster and more accurate the search.

[0211] The characteristics of MVO (Mean-of-Volume Optimization) are clear structure, low dependence on parameters, and the ability to maintain strong global search capabilities in complex, nonlinear spaces without being limited by local optimization. Currently, MVO is widely used in machine learning model hyperparameter tuning, engineering optimization, and parameter optimization. This embodiment uses Bi-LSTM as the time-series prediction network in water content prediction modeling and utilizes MVO to optimize the learning rate of the Bi-LSTM, iteratively updating it to obtain a better network training scheme. Figure 6 It can be seen that MVO and Bi-LSTM work together through "outer layer global optimization - inner layer temporal modeling": MVO continuously adjusts candidate solutions according to the objective function, while Bi-LSTM updates network parameters and feeds back prediction errors under a given configuration, forming a dynamic parameter optimization process, which helps to improve the prediction accuracy and generalization ability of the model under different regions and climate conditions.

[0212] The MVO-Bi-LSTM prediction framework, such as Figure 7 As shown.

[0213] Step 4: Output and display of prediction results.

[0214] This step receives the prediction results output from step three and synchronously shares the prediction information with relevant systems through a visualization interface and data interface.

[0215] Results Display: A visual display interface, combined with a chart library, presents the moisture content prediction results intuitively in the form of dashboards, line graphs, heat maps, etc., supporting filtering and viewing by region and time period. A real-time refresh control strategy is adopted, allowing for on-demand refresh setting to ensure that the displayed data is synchronized with the model calculation results. The interface has a built-in data annotation function, automatically annotating abnormal prediction values ​​and corresponding collection points, facilitating staff to quickly locate problems.

[0216] Data Interface: The interface adopts a standardized RESTful API design, supporting HTTP / HTTPS protocols to achieve seamless integration with forest fire risk forecasting systems and forest fire prevention and control decision-making platforms. The interface uses encrypted transmission (e.g., national cryptographic standards, AES encryption algorithms) to ensure data transmission security, supports batch data retrieval and real-time querying, and has a response time of ≤5 seconds, meeting the data sharing needs of multiple platforms.

[0217] Data Export and Retention: Employing a multi-format export control method, it supports exporting data in various formats such as Excel, CSV, and PDF. During the export process, the data is automatically formatted and labeled with key information such as data collection time, forecast time, and data source. Combined with the encrypted storage function of the storage module, the original exported data files are encrypted and retained. Data retrieval and export are supported by time range and region range, facilitating subsequent review analysis and model optimization.

[0218] Step 5: Assistance and Comprehensive Support Throughout the various steps of this method, auxiliary modules provide full support to ensure its stable and reliable operation: the power supply module provides stable power to each core module, adapting to the field environment and supporting solar power (fixed data collection stations) and lithium battery power (drones, portable data collection devices); the communication module enables data transmission and remote control between modules, supporting both wireless (4G / 5G, Bluetooth, WiFi, LoRa, 485, etc.) and wired dual communication modes, adapting to the communication environment of complex forest areas; the storage module stores raw collected data, standard databases, real-time data, prediction results, and model parameters, supporting large-capacity storage and enabling long-term data retention and rapid retrieval.

[0219] This embodiment has the following advantages.

[0220] More comprehensive data sources and wider coverage: Five-dimensional multi-source collaborative data collection is adopted, including drones, portable mobile devices, fixed collection stations, satellites, and airborne / ground remote sensing. It takes into account large-scale macro-monitoring of forest areas, coverage of complex terrain, and precise collection of micro-meteorological data in key areas, solving the problems of incomplete coverage and missing local data in traditional single data sources.

[0221] Enhanced data acquisition stability and reliability: Positioning utilizes GPS + BeiDou dual-mode, and communication employs redundant transmission via WiFi / 4G / 5G / LoRa, coupled with data caching and breakpoint resume capabilities to prevent data loss due to signal interruptions in the field; fixed stations are equipped with dual power supply of solar energy and lithium batteries, supporting 24-hour uninterrupted data acquisition, and the sensors feature built-in anomaly self-detection and outlier removal, resulting in higher data accuracy.

[0222] Data processing is more standardized and fusion is more scientific: the system completes outlier removal, missing value completion, dimensional standardization, and remote sensing noise reduction, and achieves decision-level information fusion through DS evidence theory and dynamic weight adjustment, effectively eliminating redundancy and completing information. Compared with traditional simple weighted fusion, data consistency and usability are greatly improved.

[0223] The prediction model boasts significantly superior accuracy and generalization ability: It employs Bi-LSTM deep learning with a self-attention mechanism to capture complex coupling relationships among multiple factors and fit the variation pattern of combustible moisture content; combined with MVO optimization, early stopping to prevent overfitting, grid search + Bayesian optimization, the model parameters adapt to different forest areas, forest stands, and terrain environments; it supports monthly recalibration + real-time error verification, automatic optimization when prediction deviation exceeds the limit, and stronger long-term stability.

[0224] High computational efficiency, adapted to practical emergency needs: It adopts a dual mode of GPU batch parallel computing + CPU real-time fast computing, with a response time of ≤30 seconds in key areas. It takes into account both large-scale forest area measurement and rapid prediction of fire risk emergency, breaking through the bottleneck of traditional models being slow to compute and difficult to apply in real time.

[0225] The results are presented in a more user-friendly way, making business integration easier: the results are visualized through dashboards, heat maps, and line graphs, and automatic marking of abnormal points is supported; a RESTful API encrypted interface is provided, which can be seamlessly connected to forest fire risk forecasting and forest fire prevention and control platforms, with fast interface response and secure data sharing.

[0226] Full-process traceability and easy review and optimization: Supports export of data in multiple formats, encrypted storage and retrieval. Historical data collection, fused data, model parameters and prediction results are all traceable, providing support for continuous model iteration and fire situation review.

[0227] More adaptable to field environments: The entire method is adapted to complex scenarios such as forest areas without network, remoteness, and extreme weather. The entire chain from power supply, communication, data collection to computing is designed for practical field applications, making it far more practical than laboratory-type prediction methods.

[0228] Based on the same inventive concept, this application also provides a forest surface combustible moisture content prediction device for implementing the above-mentioned method for predicting forest surface combustible moisture content. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more forest surface combustible moisture content prediction device embodiments provided below can be found in the limitations of the forest surface combustible moisture content prediction method described above, and will not be repeated here.

[0229] In one exemplary embodiment, a forest surface combustible moisture content prediction device is provided, comprising: a fixed multi-factor meteorological acquisition station, a mobile meteorological data calibration platform, a forest near-ground UAV monitoring unit, a satellite remote sensing data acquisition unit, a distributed forest data transmission network, and a moisture content prediction module; the distributed forest data transmission network comprises: a LoRa self-organizing communication network and a 5G long-distance communication network.

[0230] The fixed multi-factor meteorological data acquisition station is connected to the 5G long-distance communication network via the LoRa self-organizing communication network; the mobile meteorological data calibration platform and the forest near-ground UAV monitoring unit are both connected to the distributed forest data transmission network; the satellite remote sensing data acquisition unit is connected to the 5G long-distance communication network.

[0231] The fixed multi-factor meteorological acquisition station is deployed within the target forest to collect baseline meteorological factors for the target forest surface area; the forest-based near-ground UAV monitoring unit is used to collect three-dimensional meteorological data for the target forest surface area; the mobile meteorological data calibration platform is used to perform on-site calibration and accuracy verification of the baseline meteorological factors and the three-dimensional meteorological data, using the verified baseline meteorological factors as ground-based meteorological data and the verified three-dimensional meteorological data as airborne satellite data; the satellite remote sensing data acquisition unit is used to collect space-based remote sensing data for the target forest surface area.

[0232] The moisture content prediction module includes: The data acquisition unit is used to acquire multimodal meteorological factors of the target forest surface area at the current time; the multimodal meteorological factors include: ground-based meteorological data, air-based satellite data and space-based remote sensing data collected by fixed multi-factor meteorological acquisition stations, mobile meteorological data calibration platforms, forest near-ground UAV monitoring units and satellite remote sensing data acquisition units.

[0233] The data preprocessing unit is used to preprocess the multimodal meteorological factors at the current moment to obtain preprocessed real-time data.

[0234] The information fusion unit is used to perform decision-level information fusion processing on the preprocessed real-time data and the multimodal meteorological factors of historical time in the standard database to obtain a set of multi-source feature data under the same time series.

[0235] The moisture content prediction unit is used to input the multi-source feature data set into the moisture content prediction model to obtain the combustible moisture content of the target forest surface area at the current time. The moisture content prediction model is constructed based on the bidirectional long short-term memory deep learning algorithm and the multiverse algorithm is used to optimize the model parameters.

[0236] As an optional implementation, the LoRa self-organizing communication network includes: a LoRa main gateway; the 5G long-distance communication network includes: a 5G core gateway.

[0237] As an optional implementation, the fixed multi-factor meteorological data acquisition station includes: a meteorological sensor, a BeiDou positioning module, and a LoRa communication module; the meteorological sensor is used to collect baseline meteorological factors for the target forest surface area; the meteorological sensor is connected to the mobile meteorological data calibration platform through the BeiDou positioning module; and the meteorological sensor transmits data through the LoRa communication module.

[0238] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 8 As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores the current moisture content of combustibles in the target forest surface area. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements a method for predicting the moisture content of combustibles on the forest surface.

[0239] Those skilled in the art will understand that Figure 8 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer equipment to which the present application is applied. Specific computer equipment may include, for example, [the following is a list of possible additional structures]. Figure 8 The embodiments show more or fewer components, combinations of certain components, or different component arrangements. In one exemplary embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, which the processor executes to implement the steps in the above-described method embodiments.

[0240] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0241] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0242] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0243] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0244] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, etc., and are not limited to these.

[0245] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0246] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method of predicting forest floor fuel moisture content, characterized by, include: Obtain multimodal meteorological factors for the target forest surface area at the current moment; The multimodal meteorological factors include: ground-based meteorological data, airborne satellite data, and space-based remote sensing data collected using fixed multi-factor meteorological acquisition stations, mobile meteorological data calibration platforms, forest near-ground UAV monitoring units, and satellite remote sensing data acquisition units; The multimodal meteorological factors at the current moment are preprocessed to obtain preprocessed real-time data; The preprocessed real-time data is fused with multimodal meteorological factors from historical moments in a standard database at the decision-level information level to obtain a collection of multi-source feature data under the same time series. The multi-source feature data set is input into the moisture content prediction model to obtain the combustible moisture content of the target forest surface area at the current time. The moisture content prediction model is constructed based on the bidirectional long short-term memory deep learning algorithm and the multiverse algorithm is used to optimize the model parameters.

2. The forest ground fuel moisture content prediction method according to claim 1, characterized by, Obtain multimodal meteorological factors for the target forest surface area at the current moment, specifically including: The baseline meteorological factors are acquired from fixed multi-factor meteorological data collection stations deployed in the target forest; the fixed multi-factor meteorological data collection stations transmit data sequentially through LoRa network and 5G network. Acquire three-dimensional meteorological data collected by a forest-based near-ground drone monitoring unit; the forest-based near-ground drone monitoring unit transmits data via a LoRa network or a 5G network; A mobile meteorological data calibration platform is used to perform on-site calibration and accuracy verification of the reference meteorological factors and the stereo meteorological data. The verified reference meteorological factors are used as ground-based meteorological data, and the verified stereo meteorological data are used as airborne satellite data. The mobile meteorological data calibration platform transmits data through a LoRa network or a 5G network. The system acquires space-based remote sensing data collected by a satellite remote sensing data acquisition unit; the satellite remote sensing data acquisition unit uses a 5G network for data transmission.

3. The method for predicting the moisture content of forest surface combustibles according to claim 1, characterized in that, The multimodal meteorological factors at the current moment are preprocessed to obtain preprocessed real-time data, specifically including: The multimodal meteorological factors at the current moment are preprocessed using outlier removal algorithm, linear interpolation algorithm, neighborhood mean filling algorithm, Z-score data standardization algorithm and Gaussian filtering noise reduction algorithm to obtain preprocessed real-time data.

4. The method for predicting the moisture content of forest surface combustibles according to claim 1, characterized in that, The preprocessed real-time data is fused with multimodal meteorological factors from historical moments in a standard database at the decision-level to obtain a multi-source feature data set for the same time series, specifically including: A convolutional neural network is used to evaluate the credibility of each source data in the preprocessed real-time data. Based on the credibility evaluation results and evidence synthesis rules, the preprocessed real-time data is fused with multimodal meteorological factors from historical moments in a standard database to obtain a set of multi-source feature data under the same time series. The convolutional neural network includes an input layer, a convolutional layer, a pooling layer, and a fully connected layer connected in sequence. The convolutional neural network uses a dynamic weight adjustment strategy for fusion during the fusion process.

5. The method for predicting the moisture content of forest surface combustibles according to claim 1, characterized in that, The method for determining the moisture content prediction model includes: Obtain multimodal meteorological factors for each historical moment used for training; The multimodal meteorological factors used for training at each historical moment are preprocessed to obtain preprocessed historical data; A training dataset is constructed based on the preprocessed historical data and the corresponding measured moisture content of combustibles. Construct a bidirectional long short-term memory deep learning network; The training dataset is input into the bidirectional long short-term memory deep learning network. Based on the Adam optimizer and early stopping strategy, and using the multiverse algorithm, the model parameters in the bidirectional long short-term memory deep learning network are optimized and trained. The trained bidirectional long short-term memory deep learning network is determined as the water content prediction model.

6. A device for predicting the moisture content of forest surface combustibles, characterized in that, include: Fixed multi-factor meteorological data acquisition station, mobile meteorological data calibration platform, forest near-ground UAV monitoring unit, satellite remote sensing data acquisition unit, distributed forest data transmission network, and moisture content prediction module; The distributed forest data transmission network includes: a LoRa self-organizing communication network and a 5G long-distance communication network; The fixed multi-factor meteorological data acquisition station is connected to the 5G long-distance communication network via the LoRa self-organizing communication network; the mobile meteorological data calibration platform and the forest near-ground UAV monitoring unit are both connected to the distributed forest data transmission network; the satellite remote sensing data acquisition unit is connected to the 5G long-distance communication network. The fixed multi-factor meteorological data acquisition station is deployed within the target forest to collect baseline meteorological factors for the target forest surface area; the forest-based near-ground UAV monitoring unit is used to collect three-dimensional meteorological data for the target forest surface area; the mobile meteorological data calibration platform is used to perform on-site calibration and accuracy verification of the baseline meteorological factors and the three-dimensional meteorological data, using the verified baseline meteorological factors as ground-based meteorological data and the verified three-dimensional meteorological data as airborne satellite data; the satellite remote sensing data acquisition unit is used to collect space-based remote sensing data for the target forest surface area. The moisture content prediction module includes: The data acquisition unit is used to acquire multimodal meteorological factors of the target forest surface area at the current time; the multimodal meteorological factors include: ground-based meteorological data, air-based satellite data and space-based remote sensing data collected by fixed multi-factor meteorological acquisition stations, mobile meteorological data calibration platforms, forest near-ground UAV monitoring units and satellite remote sensing data acquisition units; The data preprocessing unit is used to preprocess the multimodal meteorological factors at the current moment to obtain preprocessed real-time data; The information fusion unit is used to perform decision-level information fusion processing on the preprocessed real-time data and the multimodal meteorological factors of historical time in the standard database to obtain a set of multi-source feature data under the same time series. The moisture content prediction unit is used to input the multi-source feature data set into the moisture content prediction model to obtain the combustible moisture content of the target forest surface area at the current time. The moisture content prediction model is constructed based on the bidirectional long short-term memory deep learning algorithm and the multiverse algorithm is used to optimize the model parameters.

7. The forest surface combustible moisture content prediction device according to claim 6, characterized in that, The LoRa self-organizing communication network includes: a LoRa main gateway; the 5G long-distance communication network includes: a 5G core gateway.

8. The forest surface combustible moisture content prediction device according to claim 6, characterized in that, The fixed multi-factor meteorological data acquisition station includes: a meteorological sensor, a BeiDou positioning module, and a LoRa communication module; The meteorological sensor is used to collect baseline meteorological factors for the target forest surface area; the meteorological sensor is connected to the mobile meteorological data calibration platform through the Beidou positioning module; the meteorological sensor transmits data through the LoRa communication module.

9. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that the processor executes the computer program to implement the method for predicting the moisture content of forest surface combustibles according to any one of claims 1-5.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the method for predicting the moisture content of forest surface combustibles as described in any one of claims 1-5.

Images