A system and method for satellite terrain intelligent identification and bird risk assessment

Abstract

The application provides a system and method for satellite terrain intelligent identification and bird risk assessment, and relates to the cross field of aviation safety and geographic information technology. In view of the defects of limited monitoring range, low identification accuracy and single evaluation dimension of traditional bird strike prevention technology, the system comprises a satellite data acquisition module, a terrain intelligent identification module, a bird activity risk assessment module and a warning and prevention decision module. Through multi-source satellite remote sensing data preprocessing, improved U-Net++ model pixel-level terrain classification, risk assessment of fuzzy logic and Bayesian network fusion, combined with spatio-temporal context and human activity data, a four-level risk heat map is generated; according to the risk level, differential warning and prevention measures are pushed, forming an evaluation, warning, prevention and feedback closed loop.

Description

Technical Field

[0001] This invention relates to the intersection of aviation safety technology and geographic information technology, and in particular to a system and method based on intelligent satellite terrain recognition and bird risk assessment, for bird strike risk early warning and prevention decision support in airport perimeters and low-altitude flight areas. Background Technology

[0002] Bird strikes are a significant threat to aviation safety. Traditional bird monitoring relies primarily on manual patrols, ground-based radar, or fixed cameras, which suffer from limited monitoring range, insufficient real-time performance, and delayed warnings. For example, some existing technologies combine radar and optical equipment for bird monitoring, but there is still considerable room for improvement in recognition accuracy in complex scenarios and the efficiency of multi-source data fusion. In recent years, the development of satellite remote sensing technology has provided new solutions for large-scale, periodic environmental monitoring. For instance, some projects have used satellite tracking technology to analyze bird activity patterns; however, this technology primarily focuses on bird migration studies and has not yet fully integrated high-resolution terrain recognition and dynamic risk assessment to achieve accurate early warnings for aviation safety.

[0003] On the other hand, existing bird strike risk assessment systems are mostly based on local sensor data, such as identifying bird targets and assessing risks through radar images. While these methods can provide real-time bird information, they are limited by the sensor coverage area, making it difficult to conduct comprehensive and forward-looking risk assessments of the ecological environment and bird habitats in large areas around airports.

[0004] Existing technologies employ a combination of radar and optical equipment, with a monitoring coverage area of ​​≤5km; satellite terrain recognition methods based on random forests achieve only 85% accuracy in identifying bird habitats. Neither addresses the dual requirements of large-scale monitoring and high-precision identification, nor do they integrate multi-dimensional data for dynamic risk assessment. Therefore, there is an urgent need for a system and method that integrates satellite remote sensing and large-scale terrain intelligent recognition technologies to achieve accurate and efficient assessment and early warning of bird activity risks, thus overcoming the shortcomings of existing technologies in spatial coverage, forward-looking early warning, and multi-source information fusion. Summary of the Invention

[0005] The purpose of this invention is to provide a system and method based on intelligent satellite terrain recognition and bird risk assessment to solve the problems of limited monitoring range, lack of forward-looking risk assessment and insufficient utilization of multi-source data in existing bird strike prevention technologies, and to achieve high-precision and intelligent assessment and early warning of bird activity risks over a wide range of airspace.

[0006] To achieve the above objectives, the present invention provides the following technical solution: Satellite Data Acquisition Module: This module acquires multi-source satellite remote sensing data and supporting auxiliary data, and generates standardized remote sensing datasets through a data preprocessing unit. Specifically, it integrates optical satellite and synthetic aperture radar satellite data to acquire high-resolution optical images with a resolution ≤10m and multispectral data covering the visible to near-infrared bands. Supporting auxiliary data includes satellite orbital parameters, atmospheric correction parameters, and a digital elevation model (DEM) with a resolution of 30m. Radiometric correction uses the 6S model to convert the original DN values ​​into surface reflectance, using the following formula: This formula is a classic atmospheric correction formula used to convert satellite sensor DN values ​​into surface reflectance, and is a standard method for remote sensing data preprocessing (the 6S model is the industry-standard model). For sensor radiance, For path radiation, Solar irradiance at the top of the atmosphere. The zenith angle of the sun. Atmospheric transmittance was used. Coordinate registration was achieved using a quadratic polynomial model based on ground control points (GCPs) and satellite orbital parameters, with the planar position error controlled within 3 meters. A multi-scale convolutional network (MSBDN algorithm) was employed to repair cloud-covered areas, with a cloud cover threshold set to ≤10%. Wavelet transform algorithms were used to fuse multi-source data (such as optical and SAR data) to improve spatial detail to 5m resolution. A standardized dataset was generated by unifying timestamps using linear interpolation based on satellite transit time.

[0007] Terrain intelligent recognition module: Employs an improved U-Net++ network with a ResNet50 backbone. Pre-trained weights are derived from ImageNet, integrating channel attention (CBAM) and spatial attention mechanisms to enhance the extraction of fine terrain features. Its loss function is Dice Loss. ,in To predict probabilities, For real labels, The environmental weights are a classic loss function commonly used in semantic segmentation to address class imbalance and are standard in the U-Net series models. In this invention, they are innovatively applied directly to the training of a terrain classification model by increasing the weights. The model enhances its focus on bird habitats and micro-topography. Terrain categories are further subdivided, including water bodies (rivers, lakes, marshes, artificial ponds), woodlands (coniferous forests, broadleaf forests, mixed forests, shrublands), farmland, built-up areas, bare land, and transition zones. For bird habitats, quantitative standards are set: water depth ≤ 2m, woodland canopy closure ≥ 0.6, and farmland crops must be Gramineae (such as wheat and rice). Transfer learning is performed using remote sensing datasets such as LoveDA, combined with data augmentation (rotation, scaling) to expand the sample. The model achieves an overall classification accuracy of 94.2% (IoU metric), with 100 training epochs and a batch size of 16.

[0008] Bird Activity Risk Assessment Module: The bird activity risk assessment model adopts a deep spatiotemporal fusion network as its backbone architecture, integrating a spatial and temporal dual-branch processing mechanism: the spatial branch extracts terrain features based on a ResNet-50 backbone of a convolutional neural network (CNN), while the temporal branch analyzes historical bird activity data (such as bird activity frequency) and spatiotemporal context data (such as weather sequences) through a long short-term memory network, and dynamically weights and fuses features using an attention mechanism to output a spatial feature map. Output time series feature vector The formula is Weight This formula, dynamically calculated using an attention mechanism, is an improvement on the classic linear fusion formula. The weights are dynamically calculated by the attention mechanism (rather than fixed values), enabling adaptive fusion of spatial features (topography) and temporal features (bird activity, weather), thus fitting the "spatiotemporal dual-branch" evaluation architecture—an innovative design of this invention. Input data is rigorously quantified, historical bird activity data is normalized hourly, and spatiotemporal context data such as wind force are processed using triangular membership functions (e.g., weak wind membership). To handle uncertainty, this formula is the most basic membership function form in fuzzy logic, used to transform qualitative indicators into quantitative membership degrees. The formula structure is the textbook standard form; risk value is calculated using a weighted formula. To achieve this, the formula is an innovative adaptation of weighted summation: 1. It clarifies the three core dimensions of risk assessment; 2. It dynamically adjusts the weights for the "mountainous scenario" (reducing the weight of migration risk and increasing the weight of human activity interference), solving the problem of "fixed weights and inability to adapt to terrain differences" in traditional assessments. This is a scenario design specific to this invention, where risks are clustered. Migration risk based on the fusion of Bayesian network conditional probability and terrain features Interference risk depends on path overlap and seasonal factors. By normalizing the intensity of human activities, default weights are used. (Mountainous areas will be dynamically adjusted to) The final output is a four-level risk heatmap (low: ,middle: (High: R≥0.8), forming a closed-loop assessment system from data quantification to risk visualization.

[0009] Early warning and prevention decision-making module: Based on different risk levels, the system generates corresponding early warning information. For each risk level, the system also provides a series of targeted prevention and control measures to help relevant units effectively deal with potential risks and ensure safety and stability.

[0010] Furthermore, the satellite data acquisition module supports access to multiple satellite data sources, including optical satellites and synthetic aperture radar satellites, and can perform registration and fusion operations on data of different resolutions and different time phases; among them, Gaofen-3 C-band SAR eliminates ground object interference through polarization calibration, Sentinel-1 achieves time series data alignment through orbit parameter correction, and Sentinel-2 adapts to multispectral analysis through band calibration.

[0011] Furthermore, the terrain intelligent recognition module uses convolutional neural networks and attention mechanisms to achieve pixel-level terrain classification. For typical bird habitats, the terrain intelligent recognition module constructs a habitat suitability feature set to achieve accurate identification of key habitats such as wetlands, grasslands, and orchards. The habitat suitability feature set includes water depth, forest canopy closure, and crop type. The quantification standards for the habitat suitability feature set are: water depth ≤ 2m, forest canopy closure ≥ 0.6, and crop type is Gramineae. Gramineae includes wheat and rice.

[0012] Furthermore, the bird activity risk assessment module adopts a risk assessment model that combines fuzzy logic and Bayesian networks. The input features include terrain type, season, meteorological data, proximity to water sources, and vegetation coverage.

[0013] Furthermore, the early warning and prevention decision-making module can trigger different levels of early warning signals based on the risk level, and recommend ecological governance, sound and light bird deterrence, or flight scheduling prevention and control measures for identified high-risk areas.

[0014] A method for intelligent satellite terrain recognition and bird risk assessment includes the following steps: Step 1: Acquire multi-source satellite remote sensing data and supporting auxiliary data for the target area through the satellite data acquisition module. Perform radiometric correction, geometric fine correction, cloud and fog removal processing, image fusion, and time series data alignment on the acquired raw data to generate a standardized remote sensing dataset. During data preprocessing, the planar position error of geometric fine correction is controlled within 3 meters. A multi-scale cloud removal algorithm is used to remove low-quality images, and a wavelet transform fusion algorithm is used to improve terrain identification under complex weather conditions. The cloud cover threshold of the multi-scale cloud removal algorithm is ≤10%. Step 2: Utilize the terrain intelligent recognition module to construct a land cover classification model based on a deep learning framework, and perform intelligent terrain category recognition on the standardized remote sensing dataset to improve edge recognition accuracy. The identified terrain categories include water areas, woodlands, farmland, built-up areas, bare land, and transition zones. Key elements of these terrain categories are extracted, including area proportion, spatial distribution pattern, and terrain factors. The land cover classification model adopts an improved U-Net++ architecture, combining transfer learning with data augmentation. The transfer learning uses ImageNet pre-trained ResNet50 as initial weights. Step 3: Based on the bird activity risk assessment module, receive refined terrain feature data, historical bird activity data, and spatiotemporal context data output from the terrain intelligent recognition module. Construct a risk assessment model using a deep spatiotemporal fusion network, taking terrain features as spatial input and historical bird activity and spatiotemporal context as temporal input. Calculate bird activity risk values ​​from three dimensions: bird aggregation risk, migration path overlap risk, and behavioral disturbance risk. Classify the risk values ​​into corresponding levels and generate a risk heatmap with the same resolution as the remote sensing data. The risk values ​​are calculated through weighted summation: aggregation risk weight 0.4, migration overlap risk weight 0.3, and disturbance risk weight 0.3. In mountainous scenarios, the migration overlap risk weight is reduced to 0.2, while the human activity intensity weight is increased to 0.4. Step 4: The early warning and prevention decision-making module generates corresponding early warning information based on the risk level. The early warning information includes the risk area, the dominant risk type, the bird species involved and their protection level, and the duration of the risk. At the same time, it provides a series of targeted prevention and control measures for each risk level, outputs a risk map and prevention and control suggestions, and sends the prevention and control effect data back to the bird activity risk assessment module to achieve closed-loop optimization.

[0015] Furthermore, in step 2, the U-Net++ network structure used enhances multi-scale feature fusion through nested skip connections and deep supervision mechanisms, thereby improving the recognition accuracy of terrain edges and small features. In the model training stage, transfer learning is introduced, using the weights pre-trained on a large-scale remote sensing land cover classification dataset as initial parameters, and data augmentation operations are combined to expand the diversity of training samples.

[0016] Furthermore, in step 3, the Bayesian network quantifies the correlation between the proportion of water area, forest canopy density, and bird aggregation probability by constructing a conditional probability model of terrain features, historical bird activity data, and bird activity. Fuzzy logic, targeting the uncertain information in the spatiotemporal context data, establishes a fuzzy rule base to transform qualitative descriptions into quantitative membership degrees. The two work together to first output the probability of birds appearing in the target area through the Bayesian network, and then combine the fuzzy logic to quantify the aggregation density, migration path overlap, and interference intensity, ultimately mapping the comprehensive result to a four-level risk level.

[0017] Furthermore, in step 4, the mapping relationship between risk level and five-level warning colors is as follows: low risk corresponds to blue warning, medium-low risk corresponds to green warning, medium risk corresponds to yellow warning, high risk corresponds to orange warning, and extremely high risk corresponds to red warning. The visualization interface integrates GIS map components, uses different color blocks to distinguish risk levels, displays risk heat maps in real time, marks the boundaries of high-risk areas and related bird activity information, and simultaneously displays a recommended list of targeted prevention and control measures. This allows airport air traffic controllers to view risk details, retrieve historical data, and confirm the implementation status of prevention and control measures through interface interaction, thereby improving decision-making efficiency.

[0018] The beneficial effects of this invention are as follows: Compared with the prior art, the system and method of this invention based on intelligent satellite terrain recognition and bird risk assessment address the problems of limited spatial coverage, low terrain recognition accuracy, single risk assessment dimension, and delayed decision response in traditional bird risk assessment methods. Specific beneficial effects are as follows: (1) In view of the limitations of traditional ground monitoring, such as the coverage radius of a single radar ≤ 5km and the spatial and temporal constraints of manual inspection, this invention adopts multi-source satellite remote sensing technology, which can realize normalized and large-scale dynamic monitoring of target areas (such as a 50km×50km range around an airport), and the monitoring cycle supports daily updates; at the same time, combined with time-series remote sensing data and migratory bird migration patterns, risk prediction can be completed 24-48 hours in advance, reserving a longer emergency response time for scenarios such as airport bird strike prevention, and solving the problem of passive response of traditional technology; (2) In view of the shortcomings of traditional remote sensing classification algorithms in identifying bird habitats such as narrow rivers and scattered shrublands, this invention uses an improved U-Net++ deep learning algorithm for pixel-level semantic segmentation and integrates channel attention and spatial attention mechanisms to enhance the extraction of fine terrain features. (3) In view of the problem that traditional risk assessment relies on only a single factor and cannot handle the uncertainty of complex environment (such as the ambiguity of bird activity time during seasonal transition and the fluctuation of wind force on bird behavior), this invention constructs a risk assessment model by integrating fuzzy logic and Bayesian network. Fuzzy logic is used to transform qualitative descriptions such as peak bird activity time and the impact of weak wind into quantitative membership degrees to handle data uncertainty; Bayesian network is used to quantify the correlation between terrain type, meteorological parameters and bird aggregation probability; the synergy of the two reduces the risk assessment error, reduces the deviation compared to single factor assessment, and improves the accuracy and reliability of risk prediction. (4) In view of the limitations of traditional technologies that rely solely on ground monitoring data, this invention integrates multi-dimensional data sources, including topographic and spatial data from satellite remote sensing, historical bird data from ground monitoring stations / infrared cameras (including population numbers and habitat behavior records of 18 common bird species), real-time temperature / wind / precipitation data from weather stations, and human activity data such as flight take-off and landing frequency and traffic flow on surrounding roads at airports, to construct a four-dimensional risk assessment system of space, time, environment and human activities. This system can comprehensively capture key factors affecting bird activity and avoid the one-sidedness of assessment caused by a single data source. (5) In response to the problems of traditional early warning relying solely on text notifications, difficulty for managers to quickly locate high-risk areas, and delayed prevention and control decisions, this invention generates a risk heat map with the same resolution as remote sensing data through an early warning and prevention and control decision-making module, and realizes the visualization and labeling of risk areas by combining it with GIS maps; at the same time, it automatically pushes specific prevention and control measures for different risk levels, which shortens the decision-making time of managers and significantly improves the scientific nature and execution efficiency of aviation safety management. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are only drawings of some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings. In addition, the drawings in the following description can be regarded as schematic diagrams and are not intended to limit the actual process of the system and method involved in the embodiments of the present invention.

[0020] Figure 1 This is a schematic diagram of the overall system structure of the present invention.

[0021] Figure 2 This is a schematic diagram of the working process of the terrain intelligent recognition module of the present invention.

[0022] Figure 3 This invention relates to a bird activity risk assessment model.

[0023] Figure 4 This is a schematic diagram of the data input and feature extraction module of the present invention.

[0024] Figure 5 This is a schematic diagram of the risk assessment engine module of the present invention.

[0025] Figure 6 This is a schematic diagram of the risk calculation and decision-making module of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] like Figures 1 to 6 As shown in the figure, an embodiment of the present invention provides a system for intelligent satellite terrain recognition and bird risk assessment, comprising: Satellite Data Acquisition Module: Developed in Python, this module integrates the GDAL and OGR libraries. By calling RESTful API interfaces from data platforms such as USGS and ESA, it automatically acquires multi-source remote sensing data and supporting auxiliary data (satellite orbit parameters, atmospheric correction parameters, digital elevation models) from optical and synthetic aperture radar satellites such as the Gaofen series, Landsat, and Sentinel-1 / 2. The module includes a built-in data preprocessing unit, which sequentially performs radiometric correction, converting DN values ​​to surface reflectance using the 6S model; geometric fine correction based on the RPC model and ground control points, ensuring a planar position error ≤3 meters; cloud and fog removal using a multi-scale cloud removal algorithm with a cloud cover threshold ≤10% for repair or removal; image fusion prioritizing wavelet transform algorithms to enhance spatial details of multi-source data to 5-meter resolution and align time-series data; interpolation based on satellite transit time; and finally, outputting a standardized GeoTIFF format remote sensing dataset through the OGC WMS interface.

[0028] The terrain intelligent recognition module is based on an improved U-Net++ deep learning architecture to build a land cover classification model. Its encoder backbone uses a ResNet-50 pre-trained on ImageNet, and integrates a convolutional attention module (CBAM) in skip connections to enhance the extraction of fine habitat features. Nested skip connections and deep supervision mechanisms are also employed to optimize multi-scale feature fusion. The model training Dice Loss formula is as follows: (This formula is a classic loss function commonly used in semantic segmentation to solve the class imbalance problem. It is the standard configuration for the U-Net series models. In this invention, it is directly applied to the training of the terrain classification model by increasing the weights.) The innovation of this invention is to enhance the model's focus on bird habitats and micro-topography. The main loss function is LoveDA and other remote sensing datasets are used for transfer learning. Combined with data augmentation such as rotation and flipping, the model is trained for 100 rounds with a batch size of 16. Finally, pixel-level semantic segmentation is achieved with an overall classification accuracy of 94.2%. It can accurately identify terrain categories such as water (depth ≤ 2m), woodland (canopy closure ≥ 0.6), farmland (crops are grasses), built-up areas, bare land and transition zones, and output classification results in vector (SHP) and raster (GeoTIFF) formats that include area, spatial distribution and terrain factors.

[0029] Bird Activity Risk Assessment Module: This module, centered on a deep spatiotemporal fusion network, constructs a three-layer computational architecture: data quantification, probabilistic reasoning, and risk fusion. The first layer involves multi-source data quantification: topographic features (such as the proportion of water area). Forest canopy density (Input directly); The number of bird occurrences N in historical bird data is statistically analyzed through time windows, and then normalized using an expected probability model based on Poisson distribution to calculate the probability of bird occurrence per unit time. (This formula is a classic formula in statistics describing the probability of an event occurring per unit of time. This invention is used to quantify the probability of bird occurrences in historical birdwatching data, and the formula form is an existing formula.) Estimated from historical observation frequency; for uncertain data such as meteorological and human activities, a fuzzy logic system is used for processing, for example, "wind force level" is obtained through triangular membership functions. The function transforms the qualitative indicator (such as "weak wind") into a membership degree based on the most basic membership function form in fuzzy logic. This function converts qualitative indicators into quantitative membership degrees, and its structure follows a textbook standard. The second layer employs dual-engine probabilistic inference: in the spatial dimension, a Bayesian network establishes the conditional dependency between terrain features and the probability of bird appearance, calculating the posterior probability. (This formula is an improvement specifically for birdwatching, and the conditional dependence of bird appearance probability is a unique design of this invention); In the time dimension, the LSTM network learns the long-term dependency relationship between birdwatching and time-series data such as weather, predicting risk trends for future periods. The third layer is multi-dimensional risk fusion: clustered risk. It is obtained by weighting the posterior probability output by the Bayesian network with the habitat suitability index. (This formula, which calculates cluster risk by weighting the posterior probability output by a Bayesian network with a habitat suitability index, is specifically designed for birds and has no universally applicable textbook form); Migration route overlap risk This is obtained by calculating the spatiotemporal overlap between the real-time monitoring area and the known migration path. This formula employs customized quantitative logic, combining the "spatial overlap area ratio" (migration route and monitoring area) with "seasonal factors" to accurately quantify migration risks. It is a proprietary design tailored to the characteristics of bird migration, without a universally applicable textbook format; behavioral disturbance risk. Calculated based on the human activity intensity index (a normalized composite of flight frequency F, road traffic volume T, etc.). (This formula is an innovative multi-source indicator fusion: integrating two core human activity indicators, "flight frequency" and "road traffic flow," and quantifying interference risk through normalized weighting, it is adapted to the correlation analysis of "human activity-bird behavior," and is a unique design of this invention.) Finally, the comprehensive risk value R is fused using dynamic weighting: (Non-fixed values, achieving "terrain adaptation and historical feedback" in risk assessment, without a textbook-standard format), where weights The attention mechanism is based on real-time terrain complexity (e.g., the rate of change of slope calculated via DEM). Dynamically generated, (The weighted calculation method addresses the shortcomings of traditional "one-size-fits-all" assessments and is a unique design of this invention.) The risk value R is mapped to four levels and a geocoded heatmap is generated.

[0030] Early Warning and Prevention Decision Module: Based on real-time received risk level and heat map data, the early warning and prevention decision module automatically triggers multi-channel early warnings through a built-in rule engine. According to the mapping relationship of "low risk - blue, low-medium risk - green, medium risk - yellow, high risk - orange, extremely high risk - red," it generates early warning information including the coordinates of the risk area (accuracy 10m x 10m), the dominant risk type, the involved bird species and their protection level, and the duration of the risk. This information is simultaneously sent to on-site terminals (audible and visual alarms), the management platform (MQTT protocol push), and responsible persons (SMS API). Prevention and Control Measures Adaptation Logic: Based on a knowledge base of prevention and control measures related to risk levels, dominant risk types, and terrain features, it automatically matches and recommends specific measures with priority ranking (e.g., for extremely high-risk water areas, it recommends "hourly drone inspections + activation of intelligent bird deterrence system + temporary traffic restrictions"). The system achieves closed-loop optimization: After the prevention and control measures are implemented, the effect data such as equipment logs, scheduling records and risk change rates are sent back to the risk assessment module through a dedicated interface. After timeliness and rationality verification, the data is used for incremental learning and parameter optimization of the Bayesian network conditional probability table and fuzzy rule base, thus forming an adaptive optimization process of "assessment-early warning-prevention and control-feedback".

[0031] Regarding system hardware and data environment configuration, this system can be deployed on a cloud platform or a local server. Recommended hardware includes: a high-performance GPU server for deep learning model inference, large-capacity storage devices for satellite imagery and historical data storage, and network equipment. Satellite data can be obtained regularly from commercial satellite data providers or public data sources such as the USGS and ESA data portals.

[0032] Regarding software implementation and module configuration, the system of this invention adopts a modular software architecture design. Each module works collaboratively through standardized data interfaces. The core technology configuration and implementation methods are as follows: The satellite data acquisition module is developed using Python as its core language, integrating a geographic data abstraction library and a spatial data processing extension library. This enables unified reading and conversion of mainstream remote sensing data formats, supporting the parsing of raster and raster-vector hybrid data types. Coordinate matching is completed through the library's geometric fine-correction function, and parameters are injected using radiometric correction tools to ensure preprocessing accuracy meets standards. Automated data processing scripts are developed, calling interfaces of mainstream satellite data platforms to achieve automatic downloading and time-series filtering of multiple satellite data series. The scripts include a built-in data quality detection function, automatically removing low-quality images based on cloud cover thresholds to ensure the validity and reliability of the input data.

[0033] The terrain intelligent recognition module constructs an improved semantic segmentation model based on a deep learning framework. It employs a pre-trained backbone network integrating an attention mechanism and nested skip connections to enhance the extraction capability of fine terrain features. Distributed training tools are used to achieve parallel computing across multiple devices, improving model training efficiency. The pixel-level classification results output by the model are processed by morphological filtering tools from a computer vision library to remove isolated noise points, and the terrain boundary continuity is optimized through boundary extraction. Finally, the data is converted into vector format to provide structured terrain information for subsequent risk assessment.

[0034] The bird activity risk assessment module utilizes a fuzzy inference system written in Python. Based on fuzzy logic programming tools, it constructs membership functions for input variables, converting qualitative indicators such as distance to water sources, vegetation cover, and intensity of human activity into quantitative values. A Bayesian network is built using probabilistic graphical modeling tools, with network nodes encompassing terrain type, season, weather conditions, bird appearance probability, and risk level. The conditional probability table between nodes is generated through algorithms learned from historical bird activity data and refined using ornithological expert knowledge. Risk value calculation employs a numerical computation library to achieve multi-dimensional data weighted fusion, with weight coefficients stored in independent configuration files, allowing for dynamic adjustment based on application scenarios. Finally, a data processing library generates a correlation table between risk levels and corresponding regions, providing data support for risk heatmap creation.

[0035] The early warning and prevention decision-making module adopts a browser / server architecture to develop a visual interactive system. The front end is based on web development technology, integrating geographic information visualization tools to load base maps, and rendering risk heat maps through a drawing interface. High-risk areas are marked with red semi-transparent polygons, and pop-up windows display information such as bird species, protection levels, and risk periods. The early warning logic is implemented through a rule engine tool. Core rules include: if the risk level is extremely high, triggering audible and visual alarms, management terminal push notifications, and electronic fence marking; if the dominant risk is overlapping migration routes, prioritizing the recommendation of temporary flight route adjustments. The prevention and control measures library is stored in a database, associating corresponding prevention and control measures according to risk levels, supporting real-time viewing and status synchronization by management personnel.

[0036] The modules interact with each other through interfaces, and use a common data exchange format to encapsulate data such as terrain classification results, risk values ​​and early warning information. The interface call latency is controlled to be within 500ms to ensure the overall response efficiency of the system.

[0037] In one embodiment, the satellite data acquisition module supports access to multiple satellite data sources, including optical satellites and synthetic aperture radar satellites, and can perform registration and fusion operations on data of different resolutions and different time phases; among them, Gaofen-3 C-band SAR eliminates ground object interference through polarization calibration, Sentinel-1 achieves time series data alignment through orbit parameter correction, and Sentinel-2 adapts to multispectral analysis through band calibration.

[0038] In one embodiment, the terrain intelligent recognition module uses a convolutional neural network and attention mechanism to achieve pixel-level terrain classification. For typical bird habitats, the terrain intelligent recognition module constructs a habitat suitability feature set to achieve accurate identification of key habitats such as wetlands, grasslands, and orchards. The habitat suitability feature set includes water depth, forest canopy closure, and crop type. The quantification standard for the habitat suitability feature set is: water depth ≤ 2m, forest canopy closure ≥ 0.6, and crop type is Gramineae. Gramineae includes wheat / rice.

[0039] In one embodiment, the bird activity risk assessment module employs a risk assessment model that combines fuzzy logic and Bayesian networks. Input features include terrain type, season, meteorological data, proximity to water sources, and vegetation cover.

[0040] In one embodiment, the early warning and prevention decision module can trigger different levels of early warning signals based on the risk level, and recommend ecological governance, sound and light bird deterrence, or flight scheduling prevention and control measures for identified high-risk areas.

[0041] A method for intelligent satellite terrain recognition and bird risk assessment includes the following steps: Step 1: Data Acquisition and Preprocessing Under the premise of clearly defining the latitude and longitude range of the target area, opening access permissions to satellite data sources (optical satellites, synthetic aperture radar satellites, etc.), ensuring that hardware storage resources meet the data storage requirements of multi-source satellites, and completing parameter configuration of the data preprocessing unit, such as radiometric correction coefficients and geometric correction coordinate systems, the satellite data acquisition module connects to data platform APIs such as USGS and ESA to automatically pull multi-source satellite remote sensing data from Gaofen-3, Sentinel-1 / 2, etc., including high-resolution optical satellite imagery, multispectral remote sensing imagery, and supporting auxiliary data such as satellite orbit parameters, atmospheric correction parameters, and digital elevation model data. During the data acquisition process, if a download failure occurs, a three-retry mechanism will be automatically triggered (retry interval of 5 minutes). If it still fails, it will switch to a backup commercial satellite data interface and record the fault log. If there is a delay in the acquisition of supporting auxiliary data, the latest available parameters will be cached through the database connection pool to ensure the process progresses.

[0042] The acquired raw data undergoes radiometric correction, geometric fine correction, cloud and fog removal, image fusion, and time-series data alignment in sequence: Radiometric correction eliminates sensor response errors and atmospheric scattering effects, ensuring accurate correspondence between image grayscale values ​​and surface reflectance; Geometric fine correction uses the WGS84 coordinate system, strictly controlling the planar position error to ≤3 meters. If this limit is exceeded, the GDAL geometric correction tool is invoked again, supplementing ground control point data for secondary correction; Cloud removal employs a multi-scale cloud removal algorithm with a cloud cover threshold of ≤10%. If the current image has excessive cloud cover, it is automatically removed and matched with a contemporary image that meets the standard. If no contemporary image is available, spatiotemporal interpolation is used to fill in the missing areas; Image fusion prioritizes wavelet transform fusion algorithms to improve terrain recognition under complex weather conditions. If fusion fails, it switches to the PCA fusion backup algorithm and marks the fusion method. Finally, a standardized remote sensing dataset is output in GeoTIFF format via the OGC WMS standard interface and pushed to the terrain intelligent recognition module.

[0043] Step 2: Intelligent extraction of terrain features

[0044] With the standardized remote sensing dataset meeting quality standards (cloud cover ≤10%, correction error ≤3 meters), the improved U-Net++ model loaded with ImageNet pre-trained ResNet50 weights and initialized, GPU computing resources ready (single card memory ≥16GB), and habitat suitability feature set quantification standards (water depth ≤2m, forest canopy closure ≥0.6, etc.) configured, the terrain intelligent recognition module receives the GeoTIFF format standardized dataset output by the satellite data acquisition module through batch data fragmentation transmission. It then uses the improved U-Net++ architecture, which integrates channel attention and spatial attention mechanisms, to perform pixel-level terrain category recognition on the data.

[0045] The model enhances multi-scale feature fusion through nested skip connections and deep supervision mechanisms, focusing on identifying water bodies (rivers, lakes, marshes, artificial ponds), woodlands (coniferous forests, broad-leaved forests, mixed forests, shrublands), farmland, built-up areas, bare land, and transition zones. It simultaneously extracts the area proportion, spatial distribution pattern, and topographic factors of various terrain types. To address the lack of identification of key habitats such as narrow rivers and scattered shrublands, it automatically calls a high-resolution image supplementation interface to acquire 0.5-meter resolution images for secondary identification. During model training, data augmentation (rotation, flipping, brightness adjustment) expands sample diversity. If the overall classification accuracy is <90%, an incremental retraining process is triggered to optimize model parameters. After classification, morphological opening operations remove isolated noise points with an area >100㎡, optimizing terrain boundary continuity. Finally, the terrain classification results are pushed to the terrain feature input port of the bird activity risk assessment module in both raster (GeoTIFF) and vector (SHP) formats.

[0046] Step 3: Bird Activity Risk Assessment

[0047] Provided that the terrain classification results pass the quality checks of boundary continuity and category integrity, the historical bird information database (containing population numbers and habitat behavior records of 18 common bird species) is accessible (data coverage ≥ 95%), the spatiotemporal context data is acquired in real time and the Bayesian network conditional probability table and fuzzy logic rule base are loaded, the bird activity risk assessment module receives refined terrain feature data (area proportion, spatial distribution, slope, aspect, etc.) output by the terrain intelligent recognition module through a dedicated interface, and obtains real-time data from the weather station and traffic flow statistics through JDBC connection to the historical bird information database and RESTful interface, thus constructing a spatial (terrain features) - temporal (historical bird information, spatiotemporal context) dual input system.

[0048] A risk assessment model is constructed using a deep spatiotemporal fusion network integrating LSTM and CNN architectures. First, a Bayesian network quantifies the correlation between topographic elements such as water area ratio and forest canopy density and bird aggregation probability, outputting the probability of bird appearance. Then, a fuzzy logic rule base is used to transform qualitative descriptions such as peak bird activity times and the impact of weak winds into quantitative membership degrees, accurately quantifying aggregation density, migration path overlap, and disturbance intensity. Risk values ​​are calculated through weighted summation, with basic weights configured as aggregation risk 0.4, migration overlap risk 0.3, and disturbance risk 0.3. If the terrain slope is identified as >25° (indicating a mountainous scene) by the digital elevation model (DEM), the weights are adjusted to aggregation risk 0.4, migration overlap risk 0.2, and human activity intensity 0.4. During the calculation process, if a single data source is missing (such as interrupted meteorological data), the historical average value for the same period is used to fill the gap and the missing data status is marked; if the Bayesian network calculation is abnormal, the backup conditional probability table is loaded and the inference engine is restarted; if the risk value exceeds the limit (<0 or >1), the calculation is recalculated after checking for anomalies in terrain factors or spatiotemporal data. Finally, the comprehensive results are mapped to four risk levels (low risk <0.3, medium risk 0.3-0.6, high risk 0.6-0.8, and extremely high risk ≥0.8), generating a risk heatmap (GeoTIFF format) with the same resolution as the remote sensing data and risk level JSON data, which are then pushed to the early warning and prevention decision-making module.

[0049] Step 4: Risk Warning and Prevention Decisions

[0050] Under the premise that the risk heat map and risk level data are fully received, the communication links of the early warning terminals (on-site audio-visual equipment, management platform, and SMS gateway) are unobstructed, the prevention and control measures library is updated to the latest version, and the GIS visualization component has completed its initial configuration, the early warning and prevention and control decision module analyzes the received risk data and generates differentiated early warning information according to the mapping relationship of "low risk - blue, medium-low risk - green, medium risk - yellow, high risk - orange, and extremely high risk - red". The content includes the risk area coordinates with a precision of 10 meters × 10 meters, the dominant risk type, the bird species involved and their protection level, and the duration of the risk.

[0051] Early warning information is pushed out simultaneously through multiple channels: on-site terminals trigger audible and visual alarms, the management platform receives data via the MQTT protocol, and designated contacts receive notifications via SMS API. If a channel fails to push, it automatically switches to a backup communication method (such as SMS forwarding to platform push) and records the information of the failed terminal to trigger manual reminders. The GIS visualization interface integrates map components, displaying a risk heat map in real time with different colored blocks, marking the boundaries of high-risk areas and related bird activity information, and simultaneously displaying a list of recommended prevention and control measures. Precise prevention and control plans are pushed out for different risk levels: emergency response is activated in extremely high-risk areas, recommending hourly drone inspections, deployment of 2-20kHz variable frequency sound waves and 532nm green laser intelligent bird deterrence systems, and temporary restrictions on human activities; daily inspections are recommended for high-risk areas, setting up bird nets and warning signs, and regulating water levels; weekly inspections, ecological protection publicity, and dynamic recording of bird activity are carried out in medium-risk areas; and monthly routine monitoring and observation equipment maintenance are conducted in low-risk areas.

[0052] After the prevention and control measures are implemented, the system transmits data on the prevention and control effect, such as equipment operation logs, flight scheduling records, and risk level change rates, back to the bird activity risk assessment module through a dedicated interface to achieve closed-loop optimization of "assessment-early warning-prevention-feedback". If the feedback data is delayed for more than 2 hours, the system sends a reminder notification. If it is delayed for more than 4 hours, it is marked as abnormal and manual verification is initiated. If the closed-loop data shows a sudden change in risk level without a reasonable reason, the abnormal data is isolated and the original risk record is retained. The model parameters are updated after verification.

[0053] In one embodiment, in step 2, the U-Net++ network structure used enhances multi-scale feature fusion through nested skip connections and deep supervision mechanisms, thereby improving the recognition accuracy of terrain edges and small features. During the model training stage, transfer learning is introduced, using the weights pre-trained on a large-scale remote sensing land cover classification dataset as initial parameters. Combined with data augmentation operations, the diversity of training samples is expanded, effectively reducing the risk of overfitting in small sample scenarios and improving the overall accuracy of terrain classification by 8%-12%.

[0054] In one embodiment, in step 3, the Bayesian network quantifies the correlation between the proportion of water area, forest canopy density, and bird aggregation probability by constructing a conditional probability model of terrain features, historical bird data, and bird activity. Fuzzy logic establishes a fuzzy rule base to transform qualitative descriptions into quantitative membership degrees for uncertain information in the spatiotemporal context data. The two work together to first output the probability of birds appearing in the target area through the Bayesian network, and then combine the fuzzy logic to quantify the aggregation density, migration path overlap, and interference intensity, and finally map the comprehensive result to a four-level risk level.

[0055] In one embodiment, in step 4, the mapping relationship between risk level and five warning colors is as follows: low risk corresponds to blue warning, medium-low risk corresponds to green warning, medium risk corresponds to yellow warning, high risk corresponds to orange warning, and extremely high risk corresponds to red warning. The visualization interface integrates GIS map components, uses different color blocks to distinguish risk levels, displays risk heat maps in real time, marks the boundaries of high-risk areas and related bird activity information, and simultaneously displays a recommended list of targeted prevention and control measures. This allows airport air traffic controllers to view risk details, retrieve historical data, and confirm the implementation status of prevention and control measures through interface interaction, thereby improving decision-making efficiency.

[0056] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A system for satellite terrain intelligent identification and bird risk assessment, characterized in that, include: Satellite data acquisition module: Configured with a data preprocessing unit, it acquires multi-source satellite remote sensing data and supporting auxiliary data, performs radiometric correction, geometric fine correction, cloud and fog removal, image fusion and time series data alignment on the raw data, and outputs a standardized remote sensing dataset; the multi-source satellite remote sensing data includes high-resolution optical satellite imagery and multispectral remote sensing imagery, and the supporting auxiliary data includes satellite orbital parameters, atmospheric correction parameters and digital elevation model data; The terrain intelligent recognition module: Based on a deep learning framework, a land cover classification model is constructed to perform terrain category recognition on a standardized remote sensing dataset; the terrain categories include water areas, woodlands, farmland, built-up areas, bare land and transition zones, where water areas include rivers, lakes, marshes and wetlands and artificial ponds, and woodlands include coniferous forests, broad-leaved forests, mixed forests and shrublands; Bird activity risk assessment module: A risk assessment model is constructed using a deep spatiotemporal fusion network. It receives terrain feature data, historical bird activity data and spatiotemporal context data output by the terrain intelligent recognition module. It calculates risk values ​​and classifies them into levels from three dimensions: risk of bird gathering, risk of overlapping migration routes and risk of behavioral interference. It outputs a risk heat map with the same resolution as the remote sensing data. Early warning and prevention decision-making module: Generates early warning information based on risk level and outputs data suggesting targeted prevention and control measures. 2.The system for satellite-based terrain intelligent identification and bird risk assessment according to claim 1, characterized in that: The satellite data acquisition module connects to data sources from optical satellites and synthetic aperture radar satellites, and performs registration and fusion on time phase data of different resolutions. Among them, Gaofen-3 C-band SAR is processed by polarization calibration, Sentinel-1 achieves time alignment by orbit parameter correction, and Sentinel-2 is adapted for multispectral analysis by band calibration.

3. The system for intelligent satellite terrain recognition and bird risk assessment according to claim 1, characterized in that: The terrain intelligent recognition module uses a convolutional neural network and attention mechanism to construct a land cover classification model, performing pixel-level terrain classification. The terrain intelligent recognition module constructs a habitat suitability feature set, and identifies wetlands, grasslands, and orchards based on the habitat suitability feature set. The habitat suitability feature set includes water depth, forest canopy closure, and farmland crop type, with the following quantification standards: water depth ≤ 2m, forest canopy closure ≥ 0.6, and farmland crop type is Gramineae. Gramineae includes wheat and rice.

4. The system for intelligent satellite terrain recognition and bird risk assessment according to claim 1, characterized in that: The risk assessment model of the bird activity risk assessment module is constructed using a combination of fuzzy logic and Bayesian networks. The input features include terrain type, season, meteorological data, proximity to water sources, and vegetation coverage.

5. The system for intelligent satellite terrain recognition and bird risk assessment according to claim 1, characterized in that: The early warning and prevention decision-making module triggers different levels of early warning signals based on the risk level, and outputs suggested prevention and control measures such as ecological governance, sound and light bird deterrence, or flight scheduling for high-risk areas.

6. A method for intelligent satellite terrain recognition and bird risk assessment, characterized in that: Includes the following steps: Step 1: Acquire multi-source satellite remote sensing data and supporting auxiliary data for the target area, perform radiometric correction, geometric fine correction, cloud and fog removal, image fusion, and time series data alignment, and output a standardized remote sensing dataset; the geometric fine correction plane position error is ≤3 meters, the cloud removal adopts a multi-scale cloud removal algorithm with a cloud cover threshold of ≤10%, and the image fusion adopts a wavelet transform fusion algorithm. Step 2: Use a land cover classification model based on a deep learning framework to perform terrain category identification on the standardized remote sensing dataset, and extract the area proportion, spatial distribution pattern and terrain factors of each terrain category; the land cover classification model is an improved U-Net++ architecture, which uses transfer learning with initial weights of ImageNet pre-trained ResNet50 combined with data augmentation. Step 3: Construct a risk assessment model using a deep spatiotemporal fusion network. Input the terrain feature data, historical bird activity data, and spatiotemporal context data output by the terrain intelligent recognition module. Use terrain features as spatial input and historical bird activity and spatiotemporal context as temporal input to calculate bird activity risk values ​​and classify them into levels, and output a risk heat map. The risk values ​​are calculated by weighted summation, with a cluster risk weight of 0.4, a migration overlap risk weight of 0.3, and a disturbance risk weight of 0.

3. In mountainous scenarios, the migration overlap risk weight is adjusted to 0.2, and the human activity intensity weight is adjusted to 0.

4. Step 4: Generate early warning information based on the risk level, output risk map and prevention and control measure suggestions, and send the prevention and control effect data back to the risk assessment model; the early warning information includes the risk area, the dominant risk type, the bird species involved and their protection level, and the duration of the risk.

7. The method for intelligent satellite terrain recognition and bird risk assessment according to claim 6, characterized in that: In step 2, the U-Net++ network structure uses nested skip connections and deep supervision mechanisms. During the model training phase, transfer learning combined with data augmentation is used, with the weights pre-trained on a large-scale remote sensing ground object classification dataset as initial parameters.

8. The method for intelligent satellite terrain recognition and bird risk assessment according to claim 6, characterized in that: In step 3, the Bayesian network constructs a conditional probability model of terrain features, historical bird activity data, and bird activity to quantify the correlation between the proportion of water area, forest canopy density, and bird aggregation probability; fuzzy logic constructs a fuzzy rule base to perform qualitative-to-quantitative membership degree processing on the uncertain information in the spatiotemporal context data. The risk assessment model first outputs the probability of bird appearance through a Bayesian network, and then combines the quantitative results of cluster density, migration path overlap and interference intensity with fuzzy logic to map the comprehensive results into a four-level risk level.

9. The method for intelligent satellite terrain recognition and bird risk assessment according to claim 6, characterized in that: In step 4, the risk level is divided into multiple levels and corresponds to different warning levels; the visualization interface integrates GIS map components, displays risk heat maps with differentiated color blocks, marks the boundaries of high-risk areas and related bird activity information, displays a list of recommended prevention and control measures, and supports interactive operations such as viewing risk details, retrieving historical data, and confirming the implementation status of prevention and control measures.

Images