Real-time forest anomaly monitoring system based on artificial intelligence and satellite remote sensing

Abstract

The invention discloses a forest real-time anomaly monitoring system based on artificial intelligence and satellite remote sensing, which relates to the technical field of interdisciplinary, and comprises a data acquisition module, a data processing module, a data acquisition module, a feature decoupling module, a dynamic updating module and an early warning generation module, the data acquisition module is configured to synchronously acquire near-infrared polarization remote sensing data and laser radar snow depth data of a target area, and the near-infrared polarization remote sensing data comprises light intensity information of a plurality of polarization channels; the beneficial effects of the method are that near-infrared polarization remote sensing data and laser radar snow depth data are combined, a dual-channel adversarial neural network is introduced, accurate detection of forest fire points is realized, a snowfield reflection correction coefficient is calculated through the snow depth data for the problem of strong ice and snow reflection of a high-latitude snowfield forest, and the accuracy of the snow reflection correction coefficient is improved. And reflectivity correction is performed on the polarization remote sensing data, so that interference signals of ice and snow are eliminated, and the reliability of fire point detection is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of interdisciplinary technologies, and particularly to a forest real-time anomaly monitoring system based on artificial intelligence and satellite remote sensing. Background Art

[0002] Forest fire monitoring is one of the important tasks for global environmental protection and disaster prevention. Due to the suddenness and rapid spread of forest fires, traditional ground monitoring means (such as patrols, ground sensors, etc.) are difficult to provide timely and comprehensive fire information in large-scale and complex terrain environments. Therefore, forest fire monitoring technology based on satellite remote sensing has gradually become a research hotspot.

[0003] Current satellite remote sensing monitoring technologies mainly rely on visible light, infrared light or microwave remote sensing data for fire point identification. However, limited by atmospheric conditions, surface reflection characteristics and sensor accuracy, traditional remote sensing methods have certain false alarm and missed alarm problems in fire point detection. Especially in high-latitude snow-covered forest environments, the strong reflection characteristics of snow and ice will cover the optical signals of fires, making it difficult for conventional optical remote sensing to accurately identify fire points. Therefore, improving the accuracy of satellite remote sensing fire point identification, reducing false alarms and missed alarms, especially the monitoring ability in ice and snow-covered areas, has become an important research direction in this field.

[0004] In recent years, the rapid development of artificial intelligence technology has provided new methods for remote sensing data processing. Especially, deep learning models have shown powerful capabilities in pattern recognition and feature extraction. By combining artificial intelligence with multi-source remote sensing data, the intelligent level of forest fire monitoring systems can be effectively enhanced, and the accuracy and real-time performance of fire point identification can be improved.

[0005] Therefore, a forest real-time anomaly monitoring system based on artificial intelligence and satellite remote sensing is proposed. Summary of the Invention

[0006] The purpose of this part is to outline some aspects of the embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this part, as well as in the abstract and title of the present application, to avoid obscuring the purpose of this part, the abstract and the title. However, such simplifications or omissions shall not be used to limit the scope of the present invention.

[0007] In view of the problems of the above-mentioned forest real-time anomaly monitoring system based on artificial intelligence and satellite remote sensing, the present invention is proposed.

[0008] To solve the above technical problems, the present invention provides the following technical solution: A forest real-time anomaly monitoring system based on artificial intelligence and satellite remote sensing, comprising:

[0009] A data acquisition module, configured to synchronously acquire near-infrared polarization remote sensing data and lidar snow depth data of a target area, wherein the near-infrared polarization remote sensing data includes light intensity information of multiple polarization channels, and the lidar snow depth data includes spatially distributed snow depth values;

[0010] A data processing module, configured to generate a snow surface reflection correction coefficient according to the snow depth data, and perform reflectance correction on the near-infrared polarization remote sensing data to obtain corrected reflectance data;

[0011] A feature decoupling module, configured to input the corrected reflectance data into a two-channel adversarial neural network, where the two channels of the two-channel adversarial neural network respectively extract ice and snow reflection artifact features and fire point thermal radiation features, and generate fire point probability distribution data based on the feature differences between the two channels;

[0012] A dynamic update module, configured to dynamically adjust the snow surface reflection correction coefficient according to the newly added snow depth data and fire point verification results within a preset time period;

[0013] An early warning generation module, configured to fuse the fire point probability distribution data with geographical coordinate data to generate a real-time fire point heat map and trigger multi-level early warning instructions.

[0014] As a preferred solution of the forest real-time anomaly monitoring system based on artificial intelligence and satellite remote sensing of the present invention, wherein the data acquisition module includes:

[0015] A near-infrared polarization sensor, configured to collect polarization light intensity data of a target area in at least four polarization directions (0°, 45°, 90°, 135°) and store it as a time series in the form of a four-dimensional tensor;

[0016] A lidar scanning unit, configured to emit pulsed laser and receive reflected signals, and generate the snow depth data through an elevation difference algorithm, wherein the snow depth data is spatially aligned with the near-infrared polarization data.

[0017] As a preferred solution of the forest real-time anomaly monitoring system based on artificial intelligence and satellite remote sensing of the present invention, wherein the specific steps for the data processing module to perform reflectance correction include:

[0018] Calculate the snow depth attenuation factor for each pixel point according to the snow depth data;

[0019] Perform a multiplication operation on the snow depth attenuation factor, the solar zenith angle, and the snow crystal refraction coefficient to obtain a dynamic correction coefficient, wherein the initial value of the snow crystal refraction coefficient is determined by laboratory spectral measurement, and the solar zenith angle is synchronously recorded by the remote sensing satellite when collecting images;

[0020] Substitute the dynamic correction coefficient into the reflectivity correction formula to correct the original near-infrared polarization data pixel by pixel, where the logic of the correction formula is: original reflectivity / (1 + dynamic correction coefficient).

[0021] As a preferred solution of the forest real-time anomaly monitoring system based on artificial intelligence and satellite remote sensing of the present invention, wherein the steps for the dynamic update module to execute the adjustment of the snow reflectivity correction coefficient include:

[0022] Receive the true fire point coordinates and snow depth verification data fed back by the ground calibration station;

[0023] Optimize the snow crystal refraction coefficient and snow depth attenuation factor based on the gradient descent algorithm;

[0024] Synchronize the optimized snow crystal refraction coefficient and snow depth attenuation factor to the data processing module in real time.

[0025] As a preferred solution of the forest real-time anomaly monitoring system based on artificial intelligence and satellite remote sensing of the present invention, wherein the dual-channel adversarial neural network includes:

[0026] An artifact feature extraction channel, configured to learn the distribution law of ice and snow reflection artifacts in the polarization-spectral joint space and output artifact confidence data;

[0027] A fire point feature extraction channel, configured to learn the thermal radiation intensity and time change rate features of true fire points and output fire point confidence data;

[0028] An adversarial loss calculation unit, configured to separate the artifact and fire point features of the dual-channel adversarial neural network by minimizing the probability distribution difference between the outputs of the artifact feature extraction channel and the fire point feature extraction channel.

[0029] As a preferred solution of the forest real-time anomaly monitoring system based on artificial intelligence and satellite remote sensing of the present invention, wherein the input data of the fire point feature extraction channel includes:

[0030] Thermal radiation time series data, generated by fusing the corrected reflectivity data and the historical fire point temperature change rate, where the temperature change rate threshold is set to 0.5 °C per minute.

[0031] As a preferred solution of the forest real-time anomaly monitoring system based on artificial intelligence and satellite remote sensing of the present invention, wherein the conditions for the early warning generation module to trigger multi-level early warning instructions include:

[0032] When the fire point probability data is in the range of 0.3 - 0.6, send a coordinate review instruction to the UAV terminal;

[0033] When the fire point probability data is greater than 0.6, push the fire point coordinates to the fire fighting server.

[0034] As a preferred solution of the forest real-time anomaly monitoring system based on artificial intelligence and satellite remote sensing of the present invention, wherein, fusing the fire point probability distribution data with the geographic coordinate data includes:

[0035] Overlay the real-time fire point heat map on the satellite image layer of the geographic coordinate data map;

[0036] Dynamically render the color transparency according to the fire point probability data, where the transparency is positively correlated with the probability value.

[0037] As a preferred solution of the forest real-time anomaly monitoring system based on artificial intelligence and satellite remote sensing of the present invention, wherein, the system further includes an edge computing node, configured to:

[0038] Deploy a lightweight inference model at the satellite or ground station to perform real-time processing on the corrected reflectivity data;

[0039] When the fire point probability data exceeds a preset threshold, directly trigger a local warning instruction and synchronize it to the cloud server.

[0040] As a preferred solution of the forest real-time anomaly monitoring system based on artificial intelligence and satellite remote sensing of the present invention, wherein, the lightweight inference model is generated through the following steps:

[0041] Perform knowledge distillation on the dual-channel adversarial neural network, and retain the weights of the fire point feature extraction channel;

[0042] Compress the model parameters and quantize and store them in the eight-bit integer format.

[0043] Advantages of the present invention:

[0044] 1. By combining near-infrared polarization remote sensing data and lidar snow depth data, and introducing a dual-channel adversarial neural network, accurate detection of forest fire points is achieved. Aiming at the strong reflection problem of ice and snow in high-latitude snow-covered forests, a snow ground reflection correction coefficient is calculated through the snow depth data, and the reflectivity of the polarization remote sensing data is corrected, thereby eliminating the interference signal of ice and snow and improving the reliability of fire point detection.

[0045] 2. Adopt a dynamic update mechanism to improve the monitoring accuracy under different meteorological conditions by adjusting the snow ground reflection correction coefficient in real time. In addition, the introduction of a multi-level warning mechanism enables the rapid transmission of fire information and improves the forest fire warning and response capabilities. Description of the Drawings

[0046] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for description in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings. Among them:

[0047] Figure 1 This is the system framework diagram of the forest real-time anomaly monitoring system based on artificial intelligence and satellite remote sensing of the present invention.

[0048] Figure 2 This is the reflectance correction step diagram of the forest real-time anomaly monitoring system based on artificial intelligence and satellite remote sensing of the present invention.

[0049] Figure 3 This is the dynamic adjustment of snow reflectance correction coefficient step diagram of the forest real-time anomaly monitoring system based on artificial intelligence and satellite remote sensing of the present invention. Detailed implementation manners

[0050] To make the above objects, features and advantages of the present invention more obvious and understandable, the following will make a detailed description of the specific implementation manners of the present invention in conjunction with the drawings of the specification.

[0051] In the following description, many specific details are set forth to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0052] Secondly, the so-called "one embodiment" or "embodiment" herein refers to a specific feature, structure or characteristic that can be included in at least one implementation manner of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that excludes other embodiments.

[0053] Furthermore, the present invention is described in detail in conjunction with schematic diagrams. When detailing the embodiments of the present invention, for the convenience of explanation, the cross-sectional views showing the device structure will be enlarged locally not in accordance with the general scale, and the schematic diagrams are only examples and should not limit the scope of protection of the present invention herein. In addition, in actual production, three-dimensional spatial dimensions including length, width and depth should be included.

[0054] Embodiment

[0055] The forest real-time anomaly monitoring system based on artificial intelligence and satellite remote sensing includes: a data acquisition module, a data processing module, a feature decoupling module, a dynamic update module, and an early warning generation module.

[0056] For the data acquisition module, it is configured to synchronously acquire near-infrared polarization remote sensing data and lidar snow depth data of the target area. Among them, the near-infrared polarization remote sensing data contains the light intensity information of multiple polarization channels, and the lidar snow depth data contains the snow depth values with spatial distribution.

[0057] As above, the data acquisition module solves the limitations of a single data source through the data complementarity of multiple sensors. Specifically, the near-infrared polarization data can penetrate the strong reflection interference on the ice and snow surface and capture the masked fire thermal radiation signal, while the lidar accurately measures the snow thickness, providing a spatial distribution basis for subsequent reflectivity correction.

[0058] For example, when the snow thickness exceeds 0.5 meters, traditional optical sensors may be completely unable to detect the fire point below, but the near-infrared polarization data can still retain valid information.

[0059] The data acquisition module specifically includes: a near-infrared polarization sensor and a lidar scanning unit;

[0060] Among them, the near-infrared polarization sensor is configured to collect the polarization light intensity data of the target area in at least four polarization directions (0°, 45°, 90°, 135°) and store it as a time series in the form of a four-dimensional tensor;

[0061] The reason for choosing four polarization directions is to completely calculate the Stokes parameters, which is the basis for analyzing the polarization state. Storing the structured data as a four-dimensional tensor (longitude, latitude, polarization direction, time) is not only convenient for subsequent processing but also supports dynamic analysis. For example, in three consecutive observations, the four-way polarization data of a certain coordinate point may show a sequence change of [120, 95, 110, 98], [115, 90, 105, 93],

[0062] [125, 100, 115, 102]. This time dimension information helps to identify instantaneous interference.

[0063] The lidar scanning unit is configured to emit pulsed laser and receive the reflected signal, and generate snow depth data through the elevation difference algorithm, where the snow depth data is spatially aligned with the near-infrared polarization data;

[0064] The reason for spatially aligning the snow depth data with the near-infrared polarization data is that there is a field of view difference between the lidar and the optical sensor on the satellite platform, and it is necessary to interpolate the snow depth data into the same grid coordinate system as the remote sensing data through coordinate transformation.

[0065] For the data processing module, it is configured to generate a snow surface reflection correction coefficient based on the snow depth data, and perform reflectivity correction on the near-infrared polarization remote sensing data according to the reflection correction data to obtain the corrected reflectivity data.

[0066] AsFigure 2 As shown, specifically, the data processing module performs reflectivity correction in three steps:

[0067] In step S11, the snow depth attenuation factor for each pixel is calculated based on the snow depth data;

[0068] Specifically, the snow depth attenuation factor represents the contribution intensity of the reflectivity attenuation for each additional meter of snow depth. Through the snow depth attenuation factor, the reflectivity attenuation ratio e caused by the snow layer thickness for this pixel can be obtained -βd , where d is the snow depth and β is the snow depth attenuation factor;

[0069] Here, the calculation of the snow depth attenuation factor can adopt a segmented strategy: the thinner the snow cover, the weaker the attenuation effect;

[0070] For example, when the snow depth d = 0.2 m, the snow depth attenuation factor β is taken as 0.2. At this time, e^{-0.2×0.2}=0.96, and the correction effect is small; while when the snow depth d = 1.5 m, β is taken as 0.12, e^{-0.12×1.5}=0.83, significantly suppressing the reflection interference of deep snow cover. This dynamic adjustment enables the system to adapt to different snow cover conditions.

[0071] In step S12, the snow depth attenuation factor is multiplied by the solar zenith angle and the snow crystal refractive index to obtain a dynamic correction coefficient. The initial value of the snow crystal refractive index is determined by laboratory spectral measurement, and the solar zenith angle is synchronously recorded by the remote sensing satellite when collecting the image;

[0072] This step aims to quantify the influence of the polarization characteristics of ice and snow reflection. Among them, the typical value of the snow crystal refractive index is 0.3. The solar zenith angle can also be calculated using astronomical algorithms based on the longitude and latitude of the target point. When actually performing the operation, the solar zenith angle needs to be converted to radians. For example, in the early morning, the solar zenith angle θ = 15°, and the converted radian is 15°×(π / 180)≈0.2618 radians. Assuming the snow depth d = 0.8 m and the snow depth attenuation factor β is taken as 0.15, then the correction coefficient W = 0.3×0.262×e^{-0.15×0.8}≈0.055. At this time, the correction amplitude is small, while at noon when the solar zenith angle θ = 60°, the converted radian is 60°×(π / 180)≈1.0472 radians. At this time, W = 0.3×1.047×e^{-0.15×0.8}≈0.278, and the correction effect is significantly enhanced. This dynamic characteristic enables the system to adapt to all-weather light changes.

[0073] In step S13, the dynamic correction coefficient is substituted into the reflectivity correction formula to perform per-pixel correction on the original near-infrared polarization data. The logic of the correction formula is: original reflectivity / (1 + dynamic correction coefficient);

[0074] In this step, pixel-by-pixel processing ensures spatial accuracy. For example, in the fire edge area, significant differences may occur in the reflectivity of adjacent pixels after correction, such as from 0.72 to 0.53 (fire area) and from 0.68 to 0.61 (ice and snow area). This enhanced contrast makes subsequent feature extraction more accurate.

[0075] Through the above three steps, the mathematical formula for correcting near-infrared polarization remote sensing data in the data processing module can be obtained as follows:

[0076]

[0077] Where R is the near-infrared polarization remote sensing data and α is the snow crystal refraction coefficient. This correction will reduce the intensity of ice and snow reflection noise while retaining the true fire signal.

[0078] For the dynamic update module, it is configured to dynamically adjust the snow surface reflection correction coefficient according to the newly added snow depth data and fire point verification results within a preset time period;

[0079] The update mechanism design of this module realizes the self-adaptability of the system. For example, when continuous snowfall for three days causes the average snow depth to increase from 0.8 meters to 1.2 meters, the system automatically adjusts the snow depth attenuation factor from 0.15 to 0.13 in the previous text by analyzing the newly collected snow depth samples, making the snow surface reflection correction coefficient more in line with the actual snow layer attenuation characteristics. Compared with the traditional fixed parameter model, this dynamic adjustment reduces the false alarm rate caused by environmental changes in temperature and season.

[0080] As Figure 3 shown, specifically, the steps for the dynamic update module to execute the adjustment of the snow surface reflection correction coefficient include three steps:

[0081] In step S21, receive the true fire point coordinates and snow depth verification data fed back by the ground calibration station;

[0082] In this step, the main function of the ground calibration station is to verify the accuracy of satellite observations. If the difference between the reflectivity after satellite correction and the measured value exceeds the threshold, it triggers the dynamic update module to execute the adjustment of the snow surface reflection correction coefficient. For example, if a calibration station detects a fire point at 62.3°N and 129.7°E, the system will extract the satellite data at this location: the reflectivity before correction R = 0.58, and after correction R' = 0.33, while the actual fire point thermal radiation should make R'≥0.45. This deviation will trigger the dynamic update module.

[0083] In step S22, optimize the snow crystal refraction coefficient and snow depth attenuation factor based on the gradient descent algorithm;

[0084] Specifically, taking the mean square error between the satellite-corrected reflectivity data and the ground-measured values as the loss function, by calculating the parameter gradients and iteratively updating with a predetermined learning rate, the snow crystal refraction coefficient and the snow depth attenuation factor are optimized. For example, initially setting the snow crystal refraction coefficient to 0.3 and the snow depth attenuation factor to 0.15, after multiple iterations, the loss function converges, and the optimized snow crystal refraction coefficient is 0.287 and the snow depth attenuation factor is 0.142. At this time, substituting into the mathematical formula for correcting the near-infrared polarization remote sensing data, the calculated corrected reflectivity is increased from 0.33 to 0.47, fitting the radiation characteristics of the real fire point.

[0085] In step S23, the optimized snow crystal refraction coefficient and snow depth attenuation factor are synchronized to the data processing module in real time.

[0086] For the feature decoupling module, it is configured to input the corrected reflectivity data into a two-channel adversarial neural network. The two channels of the two-channel adversarial neural network respectively extract the ice and snow reflection artifact features and the fire point thermal radiation features, and generate the fire point probability distribution data based on the feature differences between the two channels;

[0087] The structural core of this two-channel adversarial neural network is reflected in the two-channel interaction mechanism: the first channel focuses on identifying abnormal polarization characteristics, and the second channel detects the pattern of increasing thermal radiation over time. The two channels compete with each other during the training process, and finally the two-channel adversarial neural network can strip the ice and snow reflection interference and determine the true fire point probability of each pixel.

[0088] As above, in the feature decoupling module, the two-channel adversarial neural network should include an artifact feature extraction channel, a fire point feature extraction channel, and an adversarial loss calculation unit;

[0089] The artifact feature extraction channel is configured to learn the distribution law of ice and snow reflection artifacts in the polarization-spectral joint space and output the artifact confidence data;

[0090] The training data of this artifact feature extraction channel should include typical ice and snow reflection patterns. For example, the reflectivity of fresh snow in the 90° polarization direction is usually 1.2 - 1.5 times that in the 0° direction. The two-channel adversarial neural network extracts these spatial-polarization correlation features through a 3D convolutional layer and outputs a high artifact confidence when detecting such patterns, thereby suppressing the misjudgment of fire points in this area during subsequent fusion.

[0091] The fire point feature extraction channel is configured to learn the thermal radiation intensity and the time change rate characteristics of real fire points and output the fire point confidence data. The input data of the fire point feature extraction channel includes: the thermal radiation time series data, which is generated by fusing the corrected reflectivity data and the historical fire point temperature change rate, where the temperature change rate threshold is set to 0.5 °C per minute;

[0092] For the fire point feature extraction channel, the key is to capture the continuous growth characteristic of thermal radiation. For example, the radiation intensity of a real fire point may change from 0.45 → 0.62 → 0.81 in three consecutive observation cycles, showing continuous growth. In contrast, the snow and ice reflection artifacts usually show a fluctuating pattern of 0.58 → 0.61 → 0.59, that is, the radiation value fluctuates slightly or changes randomly. The dual-channel adversarial neural network can memorize the time series features through LSTM units. When the temperature change rate in two consecutive cycles is detected to be greater than 0.5 °C / min, the fire point confidence is increased, that is, the probability of being determined as a fire point is increased;

[0093] An adversarial loss calculation unit, configured to separate the artifacts and fire point features of the dual-channel adversarial neural network by minimizing the difference in the probability distributions output by the artifact feature extraction channel and the fire point feature extraction channel;

[0094] Specifically, by minimizing the difference in the probability distributions output by the artifact feature extraction channel and the fire point feature extraction channel, the dual-channel adversarial neural network separates the artifacts and fire point features. That is, the difference in the probability distributions output by the two channels is calculated as the loss function, and the network is automatically optimized by comparing the result differences: when the artifact feature extraction channel determines a certain area as an artifact, while the fire point feature extraction channel determines it as a real fire point, if the confidence of the determination result of the artifact feature extraction channel is greater than that of the fire point feature extraction channel, the system reduces the update amplitude of the artifact channel parameters and increases the learning rate of the fire point channel parameters through backpropagation. During training, every certain number of batches, the neural network parameters of one channel are fixed alternately, and only the parameters of the other channel are updated;

[0095] For example: Assume that during the initial training, the fire point feature extraction channel misjudges a strongly reflective snowfield (reflectivity 0.68) as a fire point (confidence 0.7), and the artifact feature extraction channel correctly identifies it as an artifact (confidence 0.85). Through alternating updates: freeze the fire point feature extraction channel and update the artifact feature extraction channel → the artifact feature extraction channel learns such reflection features and the output confidence increases to 0.92; freeze the artifact feature extraction channel and update the fire point feature extraction channel → the fire point feature extraction channel is forced to find other features (such as the time change rate) and reduces the confidence of this area to 0.3; after multiple rounds of alternating training, the false alarm rate of the system for such scenarios can be reduced, achieving the separation of artifacts and fire point features by the dual-channel adversarial neural network.

[0096] For the early warning generation module, it is configured to fuse the fire point probability distribution data with the geographical coordinate data to generate a real-time fire point heat map and trigger multi-level early warning instructions;

[0097] Specifically, the fusion of fire point probability distribution data and geographic coordinate data includes: overlaying the real-time fire point heat map on the satellite image layer of the GIS map; dynamically rendering the color transparency according to the fire point probability data, where the transparency is positively correlated with the probability value;

[0098] In the above warning generation module, the conditions for triggering multi-level warning instructions include: when the fire point probability data is in the range of 0.3 - 0.6, sending a coordinate review instruction to the UAV terminal; when the fire point probability data is greater than 0.6, pushing the fire point coordinates to the fire server.

[0099] The system also includes an edge computing node, configured to: deploy a lightweight inference model at the satellite or ground station to perform real-time processing on the corrected reflectivity data; when the fire point probability data exceeds a preset threshold, directly trigger a local warning instruction and synchronize it to the cloud server;

[0100] The above lightweight inference model is generated through the following steps: performing knowledge distillation on a dual-channel adversarial neural network and retaining the weights of the fire point feature extraction channel; compressing the model parameter quantity and quantifying and storing it in an eight-bit integer format;

[0101] By deploying a lightweight model at the satellite or ground station, the edge computing node in this system can achieve proximal real-time processing of the corrected reflectivity data, immediately trigger a local warning when detecting that the fire point probability exceeds the threshold, and synchronize the streamlined data to the cloud; its functions include real-time response, offline emergency, resource optimization, and low-power operation.

[0102] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limitations. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A real-time forest anomaly monitoring system based on artificial intelligence and satellite remote sensing, characterized in that, Including: A data acquisition module configured to synchronously acquire near-infrared polarization remote sensing data and lidar snow depth data of a target area, wherein the near-infrared polarization remote sensing data contains light intensity information of multiple polarization channels, and the lidar snow depth data contains spatially distributed snow depth values; A data processing module configured to generate a snow surface reflection correction coefficient based on the snow depth data and perform reflectance correction on the near-infrared polarization remote sensing data to obtain corrected reflectance data; A feature decoupling module configured to input the corrected reflectance data into a dual-channel adversarial neural network, where the two channels of the dual-channel adversarial neural network respectively extract ice and snow reflection artifact features and fire point thermal radiation features, and generate fire point probability distribution data based on the feature differences between the two channels; A dynamic update module configured to dynamically adjust the snow surface reflection correction coefficient according to newly added snow depth data and fire point verification results within a preset time period; An early warning generation module configured to fuse the fire point probability distribution data with geographic coordinate data to generate a real-time fire point heat map and trigger multi-level early warning instructions.

2. The real-time forest anomaly monitoring system based on artificial intelligence and satellite remote sensing according to claim 1, characterized in that, The data acquisition module includes: A near-infrared polarization sensor configured to collect polarization light intensity data of the target area in at least four polarization directions (0°, 45°, 90°, 135°) and store it as a time series in the form of a four-dimensional tensor; A lidar scanning unit configured to emit pulsed laser and receive reflected signals, and generate the snow depth data through an elevation difference algorithm, where the snow depth data is spatially aligned with the near-infrared polarization data.

3. The forest real-time anomaly monitoring system based on artificial intelligence and satellite remote sensing according to claim 1, characterized in that, The specific steps for the data processing module to perform reflectance correction include: Calculating the snow depth attenuation factor for each pixel point according to the snow depth data; Performing a multiplication operation on the snow depth attenuation factor, the solar zenith angle, and the snow crystal refractive index to obtain a dynamic correction coefficient, where the initial value of the snow crystal refractive index is determined by laboratory spectral measurement, and the solar zenith angle is synchronously recorded by the remote sensing satellite when collecting images; Substituting the dynamic correction coefficient into the reflectance correction formula to perform per-pixel correction on the original near-infrared polarization data, where the logic of the correction formula is: original reflectance / (1 + dynamic correction coefficient).

4. The forest real-time anomaly monitoring system based on artificial intelligence and satellite remote sensing according to claim 3, characterized in that, The steps for the dynamic update module to perform adjusting the snow surface reflection correction coefficient include: Receiving the true fire point coordinates and snow depth verification data feedback from the ground calibration station; Optimizing the snow crystal refractive index and the snow depth attenuation factor based on the gradient descent algorithm; Real-time synchronizing the optimized snow crystal refractive index and snow depth attenuation factor to the data processing module.

5. The forest real-time anomaly monitoring system based on artificial intelligence and satellite remote sensing according to claim 1, wherein The dual-channel adversarial neural network includes: An artifact feature extraction channel configured to learn the distribution law of ice and snow reflection artifacts in the polarization-spectral joint space and output artifact confidence data; A fire point feature extraction channel configured to learn the thermal radiation intensity and time change rate features of real fire points and output fire point confidence data; An adversarial loss calculation unit configured to separate the artifact and fire point features of the dual-channel adversarial neural network by minimizing the probability distribution difference between the outputs of the artifact feature extraction channel and the fire point feature extraction channel.

6. The real-time forest anomaly monitoring system based on artificial intelligence and satellite remote sensing according to claim 5, wherein, The input data of the fire point feature extraction channel includes: The thermal radiation time series data is generated by fusing the corrected reflectivity data with the historical fire point temperature change rate, where the temperature change rate threshold is set at 0.5 °C per minute.

7. The forest real-time anomaly monitoring system based on artificial intelligence and satellite remote sensing according to claim 1, characterized in that The conditions for the warning generation module to trigger multi-level warning instructions include: When the fire point probability data is in the range of 0.3 - 0.6, send a coordinate review instruction to the UAV terminal; When the fire point probability data is greater than 0.6, push the fire point coordinates to the fire protection server.

8. The forest real-time anomaly monitoring system based on artificial intelligence and satellite remote sensing according to claim 1, characterized in that, Fusing the fire point probability distribution data with the geographical coordinate data includes: Overlaying the real-time fire point heat map on the satellite image layer of the geographical coordinate data map; Dynamically rendering the color transparency according to the fire point probability data, where the transparency is positively correlated with the probability value.

9. The real-time forest anomaly monitoring system based on artificial intelligence and satellite remote sensing according to claim 1, characterized in that The system further includes an edge computing node, configured to: Deploy a lightweight inference model at the satellite or ground station to perform real-time processing on the corrected reflectivity data; When the fire point probability data exceeds the preset threshold, directly trigger a local warning instruction and synchronize it to the cloud server.

10. The forest real-time anomaly monitoring system based on artificial intelligence and satellite remote sensing according to claim 9, characterized in that, The lightweight inference model is generated through the following steps: Perform knowledge distillation on the dual-channel adversarial neural network, retaining the weights of the fire point feature extraction channel; Compress the model parameter quantity and quantize and store it in the eight-bit integer format.

Images