System and method for automatically detecting agricultural activities for the curtailment of wind turbines

Abstract

The invention relates to a system (100) and to a method (300) for protecting birds by indirect detection for the curtailment of wind turbines, which system comprises at least one camera (110) acquiring images of a monitored area, a neural network detecting agricultural vehicles, and a computing unit (120) performing a contextual classification of agricultural activities by analysing a peripheral spatial environment extended beyond a bounding box (10) of the detected vehicle. The contextual classification analyses the condition of the soil, the presence of agricultural equipment, changes in terrain, and vegetation characteristics in order to distinguish between productive agricultural activities and mere traffic. A curtailment signal is generated in response to the detection of agricultural activities likely to attract birds, allowing the avian risk to be forecast 2 to 4 hours prior to the arrival of large numbers of birds. This indirect detection approach significantly increases the amount of advanced warning given (hours vs. seconds), reduces false positives from 35% to less than 2%, and simplifies the architecture compared to that of direct bird detection systems.

Description

SYSTEM AND METHOD FOR AUTOMATIC DETECTION OF AGRICULTURAL ACTIVITIES FOR THE CIRCULATION OF WIND TURBINES

[0001] The invention relates to the field of protecting bird biodiversity near wind energy production facilities. More specifically, it concerns a system and method for automatically detecting agricultural activities for the preventive curtailment of wind turbines to avoid collisions with birds.

[0002] It falls within the context of wildlife protection systems equipping onshore wind farms, with particular attention to applications requiring several hours' advance warning of bird risk.

[0003] Existing bird protection systems for wind turbines are mainly limited to two types of approaches: direct bird detection systems and active bird scaring systems.

[0004] Direct detection systems use cameras, radar, or acoustic sensors to identify birds flying in the immediate vicinity of wind turbines. When a bird is detected, a curb signal (reducing or temporarily stopping the rotation of the blades) is triggered to prevent a collision.

[0005] These systems, while effective in certain situations, have significant limitations. Reaction time is limited to just a few seconds before a potential collision. The false negative rate is high, especially in adverse weather conditions. Technical complexity necessitates precise calibration and regular maintenance. Installation and operating costs are substantial.

[0006] Despite the development of numerous direct detection systems over the past 15 years (with the first commercial systems appearing around 2010), the adoption rate of these solutions remains very low (estimated at less than 5% of wind farms in 2024). Installation costs range from €50,000 to €150,000 per wind turbine. Effectiveness is limited, with false negative rates of 15% to 30% depending on the conditions. Return on investment remains uncertain.

[0007] Prior art systems, such as those disclosed in US12273497B2, rely on a stereoscopic architecture using a pair of cameras arranged along a defined baseline, with parallel optical axes forming an angle with this baseline. These systems allow for the localization of birds in flight by calculating stereoscopic disparity.

[0008] However, these approaches have several limitations: - limited reaction time (detection a few seconds before collision), - complexity of stereoscopic calibration (rigid baseline maintenance), - restricted field of view requiring rotating cameras, and - sensitivity to weather conditions (fog, rain).

[0009] Active bird deterrent systems use sound, light signals, or visual changes in the propeller blades to discourage birds from approaching. These approaches suffer from limited effectiveness due to habituation.

[0010] This long-standing unmet need demonstrates the necessity of a radically different alternative approach.

[0011] Thus, there is a need for a bird protection system capable of operating with: - significantly improved time anticipation (hours vs seconds), - increased robustness in varied weather conditions, - architectural simplicity (without rigid stereoscopic constraints), - a reduction of false positives (unwanted throttling), and - an optimized cost / effectiveness ratio.

[0012] Such a system would address the operational constraints of wind farms, while improving the effectiveness of bird protection, particularly for raptor species especially vulnerable to collisions.

[0013] The invention aims to solve, at least partially, this need.

[0014] In particular, the invention relates to a method of bird protection by indirect detection for the curtailment of wind turbines, comprising: - the acquisition of images of a surveillance zone around at least one wind turbine by at least one camera having sufficient resolution to allow the detection of agricultural vehicles in the surveillance zone; - the detection of moving objects corresponding to vehicles by machine learning means for object detection; - the verification of classification of moving objects by a pre-trained machine learning means configured to extract visual features and produce a confidence measure in order to reduce false positives; - the temporal tracking of detected moving objects to maintain their identity through a sequence of images, the temporal tracking using an object association and trajectory prediction algorithm.- the contextual classification of at least one agricultural activity by means of a contextual analysis integrating an extended peripheral spatial and temporal environment of the detected mobile object and including: -- the extension of an analysis area beyond a bounding box of the detected mobile object, so as to include a surrounding peripheral area; -- the capture of an environmental context including at least one element selected from: soil surface condition, presence of agricultural equipment, terrain modifications, vegetation characteristics; -- temporal analysis by averaging information from an image sequence to improve classification robustness; -- the distinction between agricultural vehicles performing a productive activity and non-agricultural vehicles; and -- the identification of at least one agricultural activity.- the generation of a control signal to limit the speed of at least one wind turbine, based on the contextual classification obtained in the contextual classification step.

[0015] The invention also relates to a bird protection system for bridling wind turbines implementing the method according to the invention, as well as a use of the system allowing anticipation of the bird risk between 2 and 4 hours before a massive arrival of birds.

[0016] Other features and advantages of the invention will be better understood from the description that follows and with reference to the attached drawings, given for illustrative purposes only and not for limitation.

[0017] The diagram represents a schematic overview of a bird protection system 100 according to the invention installed on a wind turbine.

[0018] Lamontre presents a flowchart of the detection and classification process according to the invention

[0019] This presents an example of contextual classification with an extension of the analysis area beyond the bounding box of a detected tractor.

[0020] Context of the invention

[0021] The inventors observed that birds of prey (buzzards, kites, falcons) are massively attracted to agricultural activities (plowing, harvesting, mowing) that disturb their natural prey (rodents, insects). This well-known ecological correlation had never before been technically exploited for the protection of wind turbines from birds.

[0022] Prior art systems rely on a "direct detection" approach aimed at identifying the birds themselves before a collision. The present invention adopts a radically different "indirect detection" approach, which consists of identifying not the birds, but the human activities (agricultural activities) that attract birds to the vicinity of wind turbines.

[0023] This approach has several technical advantages: - early anticipation of risk (detection from the start of agricultural work, several hours before the massive arrival of birds), - increased robustness (large agricultural vehicles, moderate speed, predictable trajectories), - reliability in degraded weather conditions (tractors visible even in light fog).

[0024] This new technical solution significantly reduces collisions, decreases false positives, and improves the durability of installations, representing substantial progress compared to the existing state of the art.

[0025] Definition - Contextual classification

[0026] In the context of the present invention, "contextual classification" refers to an analysis method combining the visual characteristics of the detected object (agricultural vehicle) with the spatial and temporal data of its immediate environment, including in particular the state of the soil, the presence of agricultural equipment, changes in the terrain, and variations on a sequence of images.

[0027] This method allows for a more precise modeling of ongoing agricultural activity, distinguishing productive operations from simple circulations, thus significantly improving robustness and reliability compared to an isolated analysis of the object alone.

[0028] General system architecture

[0029] As illustrated in Figure 1, the bird protection system 100 according to the invention comprises: - at least one camera 110 configured to acquire images of a surveillance area around at least one wind turbine 200, the camera 110 having sufficient resolution, typically greater than or equal to 4K, to allow the reliable detection of agricultural vehicles in the surveillance area; - at least one computing unit 120 configured to implement the method according to the invention. It is preferably positioned at the base of the wind turbine 200's mast and includes processors dedicated to sequential processing and artificial intelligence calculations (detection, classification, and tracking processes). This computing unit 120 is equipped with sufficient working memory for real-time processing as well as storage memory for models and archived data.- several distinct functional modules, including a moving object detection module implemented by machine learning means, a contextual analysis module configured for the classification of agricultural activities taking into account the peripheral spatial and temporal environment, as well as a module for distinguishing between agricultural and non-agricultural vehicles by detecting specific characteristic elements: large diameter wheels, tracks, agricultural tools, raised cab. - at least one communication interface 130 with a wind turbine control system 200 to transmit a curtailment signal to the wind turbine control 200. This interface supports standard or proprietary protocols, and can transmit additional information such as the nature of the detected activity, the position, the detection confidence, and the timestamp.

[0030] Optionally, the system 100 can include local processing modules at each camera 110 to pre-analyze images, thereby reducing the load on the computing unit 120 and network bandwidth.

[0031] This composite architecture ensures effective monitoring and easy integration into existing wind farm infrastructures.

[0032] In one embodiment illustrated in Figure 1, a plurality of cameras 110, typically eight cameras 110, are distributed around the wind turbine 200 to provide 360° coverage of the surveillance area. The cameras 110 are mounted under the nacelle and equidistant, each camera 110 covering a defined angular sector, ensuring optimal stability and an unobstructed field of view. This configuration allows for complete omnidirectional surveillance of the capture area without requiring a rotation mechanism for the cameras 110.

[0033] The 110 cameras are configured to acquire high-resolution images of the surveillance area extending from the base of the mast. The surveillance distance can vary depending on application requirements and site constraints, typically between 200 and 1000 meters. In a preferred embodiment, the surveillance area extends to a distance of at least 400 meters, potentially reaching up to 600 meters under optimal visibility and hardware configuration conditions. This minimum distance of 400 meters in the preferred embodiment allows for sufficient advance warning of bird risk, compatible with the arrival times of raptors after the start of agricultural activities, as described below in the "Time Window Validation" section.

[0034] – Acquisition cameras

[0035] The 110 cameras have sufficient resolution to allow the detection and identification of agricultural vehicles at the minimum required distance.

[0036] The "sufficient resolution" of a 110 camera refers to its ability to acquire images with sufficient quality and detail to clearly identify agricultural vehicles within the surveillance area. This resolution depends on technical parameters, including the number of pixels, image sharpness, and acquisition frequency. Environmental conditions such as brightness and visibility also affect the effective resolution.

[0037] In one embodiment, the 110 cameras feature 4K resolution (3840×2160 pixels, or 8 megapixels), enabling optimal detection up to 600 meters or more. For applications requiring a shorter range, typically between 200 and 400 meters, Full HD resolution (1920×1080 pixels, or 2 megapixels) may be sufficient, offering a cost-performance compromise suitable for certain site configurations.

[0038] In the preferred embodiment illustrated in Figure 1, the cameras 110 are positioned under the nacelle of the wind turbine 200, fixed to the tower structure, and oriented to cover equidistant angular sectors. For eight cameras 110, each camera 110 covers a 45° sector.

[0039] In one example, the 110 cameras are fixed, unlike prior art systems that used rotating 110 cameras. This configuration simplifies installation and eliminates maintenance issues related to rotating mechanisms.

[0040] The 110 cameras are connected to the 120 computing unit via a communication network ensuring sub-millisecond time synchronization, typically less than 10 milliseconds. This synchronization ensures temporal consistency of analyses from multiple 110 cameras.

[0041] The 100 camera system surrounding the 200 wind turbines is configured to ensure precise temporal synchronization of image acquisitions, via an internal communication protocol based on synchronized clocks (e.g., NTP or PTP protocol). This synchronization ensures that the 110 multi-camera analyses can be efficiently correlated in real time, facilitating data fusion and consistent detection of moving objects.

[0042] Redundancy is ensured by the configuration of several 110 cameras covering overlapping areas, allowing compensation in case of temporary or permanent failure of a 110 camera, with an automatic failover and data integration mechanism from the functional 110 cameras.

[0043] In one particular implementation, the system employs advanced time synchronization between cameras using a self-adaptive mesh network. This mesh network maintains sub-millisecond synchronization between the different cameras by automatically compensating for varying propagation delays in the communication network and dynamically recalibrating internal clock drifts. This approach ensures that images acquired simultaneously by multiple cameras are perfectly time-aligned, enabling consistent and reliable detection across the entire 360° surveillance system.

[0044] In an advanced embodiment, the system creates a dynamic 4D representation space in which the data from each camera is projected into a common volumetric spatial reference frame, incorporating the temporal dimension. This projection transforms the individual 2D images from each camera into a unified volumetric representation covering the entire surveillance area. A tensor fusion algorithm then combines the spatial (x, y, z coordinates) and temporal (t) information from all the cameras, enabling a consistent analysis of the objects detected in this unified space.

[0045] This 4D representation offers a significant technical advantage for occlusion resolution. When an agricultural vehicle is temporarily obscured from the field of view of one or more cameras by physical obstacles (buildings, vegetation, terrain), the predictive geometric reconstruction algorithm leverages the information accumulated in the volumetric space to estimate the probable position of the obscured object. This reconstruction relies on the object's historical trajectory, its speed, its direction of travel, as well as spatial data from adjacent cameras whose field of view is not obstructed. This mechanism considerably improves the robustness of temporal tracking and reduces identification losses due to occlusions, thus ensuring continuous detection even in complex environments.

[0046] – Distance calculation

[0047] Unlike earlier stereoscopic systems that calculate the distance to an object by disparity between two cameras 110 separated by a defined baseline, the present system 100 uses a monovision approach.

[0048] The distance to a detected object can be estimated by its apparent size in the image, by correlating the pixel size with the known actual distance for standard agricultural vehicles. Vertical position within the field of view can also be used, with more distant objects appearing higher in the image. Sharpness or contrast analysis provides an additional indicator, with distant objects exhibiting less sharpness.

[0049] This size-distance correlation model is established through prior on-site calibration, combining actual distance measurements with observed pixel dimensions for different types of agricultural vehicles. The model can be adjusted according to terrain specifics and local atmospheric conditions, thus ensuring a reliable and rapid estimation even in varied environments.

[0050] This single-vision approach has the advantage of not requiring complex stereoscopic calibration between pairs of 110 cameras, thus simplifying system installation and maintenance.

[0051] In one variant, several 110 cameras can be used redundantly for triangulation, but without the rigid baseline constraint found in conventional stereoscopic systems.

[0052] – Unit of calculation

[0053] In practice, the computing unit 120 includes one or more processors for sequential processing, one or more graphics processors for accelerating artificial intelligence calculations, sufficient working memory for real-time processing, and storage memory for artificial intelligence models and data recording.

[0054] In one embodiment, the computing unit 120 is positioned at the base of the wind turbine mast 200, in a weatherproof housing.

[0055] In one variant, each camera 110 can be equipped with a local processing unit performing a pre-analysis, the results of which are then aggregated by the computing unit 120 for final classification. This distributed architecture reduces the required network bandwidth.

[0056] The computing unit 120 includes dedicated processors capable of executing real-time detection and classification algorithms, leveraging pre-trained artificial intelligence models. The system 100 has sufficient storage space to host analysis programs, training databases, and collected data for archiving and auditing. Storage media may include SSDs, flash memory, or other technologies suitable for continuous operation in potentially harsh environments. Data is organized according to an architecture that allows for fast and secure access and retention in accordance with applicable standards.

[0057] Detection and classification method

[0058] As illustrated in the figure, the process 300 according to the invention comprises several successive steps implemented by the computing unit 120.

[0059] – Step 310 - Image Acquisition

[0060] The 110 cameras continuously acquire images of the surveillance area at a rate of between 10 and 30 frames per second. A rate of 15 frames per second offers a compromise between detection speed and processing load.

[0061] – Step 320 - Primary detection of moving objects

[0062] The acquired images are analyzed using machine learning object detection methods to identify vehicles within the surveillance area. Machine learning methods used for object detection can include, but are not limited to, convolutional neural networks, hybrid methods combining classical image processing and deep learning, or any other supervised algorithm capable of identifying and locating objects in images. This flexibility allows the System 100 to be adapted to the hardware constraints and specific requirements of each deployment.

[0063] For each vehicle detected, the detection module provides a bounding box 10() defined by its coordinates in the image, a confidence score representing the probability that the object is indeed a vehicle, and a predicted class.

[0064] In one embodiment, a YOLO (You Only Look Once) convolutional neural network is used, with additional Feature Pyramid Network (FPN) layers to enhance multi-scale detection. However, the invention is not limited to this specific architecture and can utilize any object detection neural network such as Faster R-CNN, EfficientDet, or models based on Autoencoder and Transformer architectures.

[0065] Depending on the available computing power and the required detection range, different levels of network complexity can be configured, with the number of parameters typically ranging from 3 million to 26 million.

[0066] In one example, the neural networks used for detection and classification are trained on a large dataset of manually annotated images representing a wide variety of agricultural vehicle types (tractors, combine harvesters, sprayers, etc.) under different environmental conditions, such as varying lighting (day, night), weather conditions (rain, fog), and diverse geographical and environmental contexts, ensuring robust generalization. These annotations include not only the location of the vehicles in each image (bounding box edges) but also precise labels describing their activity (farm work, simple movement), enabling effective supervised learning.Data augmentation techniques are applied, such as rotations, zooms, brightness fluctuations, and additional noise, to increase the robustness of the model against natural and unforeseen variations during on-site implementation.

[0067] The learning process includes a cross-validation phase to fine-tune hyperparameters such as the learning rate, layer depth, and convolutional filter size. The training process also incorporates image segmentation and the extraction of specific feature elements (large-diameter wheels, agricultural implements) to enhance the model's ability to distinguish between different types of vehicles and agricultural activities. A continuous evaluation phase on separate validation sets optimizes the model while preventing overfitting, thus ensuring a good balance between accuracy and generalization.

[0068] Metrics such as accuracy, recall, and F1 score are monitored to evaluate and ensure model performance, with detection rates exceeding 95% and a false positive rate below 2% in real-world test environments. Specific tests under challenging conditions (fog, low light) are conducted to adapt the models and maintain high operational reliability.

[0069] The artificial intelligence algorithms used are described according to their principle and function. They rely on adaptive methodologies to ensure high robustness in various environments.

[0070] – Step 330 - Secondary Verification

[0071] To improve the robustness of detection and reduce false positives, a secondary verification step is implemented.

[0072] In a preferred embodiment, classification verification uses a pre-trained neural network designed to extract discriminating visual features from detected objects. This network is coupled with a supervised learning binary classification algorithm, which validates whether the object is an agricultural vehicle. Any type of neural network capable of extracting visual features, trained on a large annotated dataset, can be used, without limitation to a particular architecture. The network is trained on a large, manually annotated dataset with diverse lighting, weather, and environmental conditions, thus ensuring the robustness and generalizability of the model. The binary classification algorithm applies a threshold to a confidence metric derived from the neural network, such as the probability associated with the presence of an agricultural vehicle.This approach allows a significant reduction in false positives in detection, improving the reliability of the 100 system to avoid untimely throttling.

[0073] The pre-trained machine learning model used can take several forms, including but not limited to convolutional neural networks, support vector machines, Bayesian classifiers, or other types of binary supervised classifiers. The binary supervised classifier can be implemented in various algorithmic forms, including threshold classifiers, support vector machines (SVMs), random forests, or deep neural networks. This algorithmic flexibility allows the 300 process to be adapted to specific site constraints and environmental variations.

[0074] The main technical effect consists of a significant reduction in false positives, thus decreasing untimely restraints and optimizing the effectiveness of the avian protection system 100.

[0075] Secondary verification also includes the detection of characteristic features that distinguish agricultural vehicles from other vehicles. These characteristic features include, in particular, large diameter wheels (typically over 1 meter), rubber or metal tracks, towed or mounted agricultural implements, the raised cab characteristic of agricultural machinery, and specific color schemes.

[0076] This distinction helps to avoid false positives triggered by the circulation of non-agricultural vehicles such as maintenance trucks or light vehicles.

[0077] The combination of these methods ensures a robust, fast verification adapted to the operational constraints of the system.

[0078] – Step 340 - Multi-object time tracking

[0079] The detected and verified objects are tracked across successive images to maintain their temporal identity. This tracking allows for the accumulation of observations on the same entity even if it moves or undergoes temporary occlusions.

[0080] In one embodiment, a multi-object tracking algorithm is used, combining visual feature extraction (e.g., using a neural network), an association algorithm to match detections between successive images, and a prediction filter to estimate the future position of objects. Temporal tracking includes management of temporary occlusions. When a vehicle temporarily disappears from the field of view, the system maintains its identity for a configurable duration by predicting its probable trajectory. This configurable duration can be adjusted from a few seconds to several minutes depending on installation conditions and robustness requirements, ensuring optimal adaptation to the various scenarios encountered in the field. Trajectory prediction relies on probabilistic models that estimate the object's future position based on its speed, direction, and movement history.

[0081] This mechanism significantly improves the continuity of identification data, ensuring a more robust and reliable analysis of detected agricultural activities even in the presence of visual disturbances.

[0082] – Step 350 - Contextual classification of agricultural activity

[0083] In this step, instead of only analyzing the detected vehicle, the 100 system analyzes the environmental context around the vehicle to determine the ongoing activity.

[0084] Extended contextual analysis integrates spatial and temporal information beyond simple object detection, including a peripheral zone adapted to the speed, size, and type of vehicle, as well as environmental characteristics such as soil surface condition, the presence of agricultural equipment, terrain modifications, and vegetation characteristics. This multi-dimensional approach significantly improves the robustness and accuracy of agricultural activity classification.

[0085] –– Analysis area extension

[0086] As illustrated in the figure, contextual classification is based on extending the area of ​​interest beyond the bounding box 10 of the detected moving object, so as to include a surrounding peripheral area 20.

[0087] This extension can be uniform, with a constant margin added to all sides of the bounding box 10, or adaptive, with a variable extension depending on object characteristics such as its speed, direction of movement, size, or type. The area extension can be modulated according to the type of object detected, its speed, and environmental conditions, ensuring an optimal compromise between accuracy and computational load.

[0088] The extension of the analysis area beyond the bounding box 10 of the detected object represents between 20% and 200% of the surface area of ​​this box, depending on whether the extension is uniform or adaptive. This extension aims to include a surrounding peripheral zone 20, allowing for a better capture of the environmental context and characteristic elements related to ongoing agricultural activity. Experimental studies have demonstrated that the extension significantly impacts the sensitivity and accuracy of contextual classification, by enabling the integration of additional visual cues such as soil texture, the presence of equipment, or changes in terrain. The adaptive methods include consideration of vehicle size, speed, the nature of the activity, and specific terrain conditions, thus optimizing the trade-off between accuracy and computational load.

[0089] –– Capture of the environmental context

[0090] The environmental context analyzed includes the visual elements present in the extended peripheral zone 20.

[0091] The condition of the soil surface constitutes a first type of element analyzed. This condition includes turned soil with a brown earth color and irregular texture, the presence of parallel furrows characteristic of recent plowing, flattened vegetation resulting from mowing or harvesting, freshly disturbed soil during sowing or stubble cultivation, or the change in soil color following spreading.

[0092] The presence of agricultural equipment constitutes a second type of element analyzed. This equipment includes plows hitched behind a tractor, combine harvesters in working configuration, visible spreaders, sprayers or seed drills, or any identifiable towed tool.

[0093] Changes to the terrain constitute a third type of element analyzed. These changes include traces of recent passage forming ruts, mowed areas contrasting with unmowed areas, hay windrows characteristic of haymaking, or visible crop residues.

[0094] The characteristics of the surrounding vegetation constitute a fourth type of element analyzed. These characteristics include the type of crop (cereals, corn, grassland), the height and density of vegetation, or the presence of mowed or cut vegetation.

[0095] The analysis of the environmental and behavioral context includes various criteria to distinguish a productive agricultural operation from simple traffic. These criteria include: the presence of specific agricultural equipment (plows, harvesters), the vehicle's speed and trajectory, physical modifications to the soil (furrows, flattened vegetation), and the location in cultivated areas versus roads. An operation is considered productive if several of these criteria are simultaneously detected and meet predefined thresholds established by machine learning. To improve accuracy, the system can use additional historical and contextual data specific to the installation site.

[0096] Masking irrelevant areas allows for the automatic filtering of detections located on geographical or artificial features unrelated to agricultural activities. These areas include, in particular, paved roads, uncultivated dirt tracks, agricultural buildings, forested areas, and bodies of water. Irrelevant areas can be defined either by prior mapping integrated into the system during installation (for example, cadastral plans or topographic surveys), or by machine learning, which automatically identifies these areas based on the visual and geospatial characteristics detected in the images. This dual approach ensures optimal implementation flexibility adapted to the specific characteristics of each site.

[0097] This approach significantly reduces the number of false positives related to simple movement or inert elements, thus improving the overall accuracy of the contextual classification system. Tests carried out demonstrate a notable decrease in false alerts in the complex and varied contexts encountered in the field.

[0098] This combination of masking and precise criteria for the characterization of agricultural activities ensures a reliable and robust distinction between productive activity and simple circulation.

[0099] –– Time analysis by averaging

[0100] To improve the robustness of the classification and reduce random variations between successive images, contextual classification performs a temporal analysis by averaging information from a sequence of consecutive images.

[0101] This averaging can be implemented by pixel-by-pixel averaging of images, by averaging of features extracted by neural network, or by voting between successive predictions.

[0102] In a preferred embodiment, the analyzed sequence comprises between 10 and 50 consecutive images. This sequence allows for the aggregation of collected visual and contextual information to smooth out fluctuations and significantly improve the accuracy of agricultural activity classification. A sequence of 25 consecutive images represents an optimal compromise, which can be adjusted according to operational constraints. The image sequence is acquired at a predefined acquisition rate, typically between 10 and 30 frames per second, ensuring smooth detection while maintaining an acceptable computational load. For example, at a rate of 15 frames per second, a 25-image sequence corresponds to a time window of approximately 1.7 seconds of observation, sufficient to capture stable activity characteristics while filtering out random variations.

[0103] These numerical ranges were determined following empirical studies and real-world tests to optimize the sensitivity and stability of the classification while ensuring real-time processing adapted to the available computing power. Specific adjustments can be made to adapt to the conditions and requirements specific to each implementation site, thus guaranteeing the robustness and reliability of the 300 process.

[0104] The graduated curtailment thresholds have been optimized based on field studies and ornithological data, designed to maximize biodiversity protection while minimizing the impact on energy production. These thresholds can be dynamically adapted according to weather conditions, crop type, and the specific configuration of wind farms, thus allowing for optimal responsiveness and efficiency of the bird protection system.

[0105] –– Identification of agricultural activity

[0106] Based on environmental and temporal contextual analysis, the 100 system classifies ongoing agricultural activity into a set of predefined classes.

[0107] Agricultural activities that may be classified include: - plowing (deep turning of the soil), - stubble cultivation (shallow tillage of the soil after harvest), - harvesting (cereal harvest), - mowing (cutting grass or fodder), - haymaking (drying and conditioning hay), - sowing (planting seeds), - spreading (fertilizing, soil amendment), - application of plant protection products, - harrowing (fine soil preparation), or any other agricultural operation involving the movement of motorized vehicles on cultivated plots.

[0108] The selection of these activities is based on their ecological importance in attracting birds of prey and their potential impact on the risk of collision with wind turbines. Each activity is identified through the integration of specific visual signals extracted from the image sequence, including the presence of agricultural tools, soil conditions, and characteristic machinery movements. The identification criteria are calibrated to maximize classification accuracy while minimizing the risk of confusion with non-productive movements or other non-agricultural activities.

[0109] In a simplified embodiment, system 100 can be limited to a binary classification distinguishing a detected productive agricultural activity from an absence of productive activity corresponding to simple circulation.

[0110] Step 360 - Generation of the bridging signal

[0111] Based on the contextual classification obtained in the contextual classification step, system 100 generates a control signal to curtail the power output of wind turbine 200. The control signal emitted for curtailment can be binary (full activation or deactivation) or graduated (gradual power reduction), with the intensity and duration of the curtailment modulated according to the detected activity parameters (type, distance, intensity) and the operational requirements of the wind turbines. This flexibility optimizes the balance between biodiversity protection and continuous energy production.

[0112] In one embodiment, the suppression signal is graduated according to distance. The suppression thresholds (e.g., 50% at 600 m, 80% at 400 m, 100% at 200 m) are chosen based on field studies and analyses of bird behavior, aiming to optimize the balance between biodiversity protection and production continuity. This gradation makes it possible to adapt the suppression power to the intensity of the bird risk associated with the proximity of detected agricultural activities.

[0113] To prevent excessively frequent switching that could cause mechanical fatigue in the blades, the 100 system incorporates a timer. A delay of between 30 seconds and 2 minutes after activity detection can be configured for activation of the clipping mechanism. A deactivation delay of between 10 and 30 minutes after the end of detection allows protection to be maintained during the period when the activity may have attracted birds of prey that remain temporarily in the area.

[0114] –– Control interface

[0115] The curtailment signal is transmitted to the control system of the 200 wind turbine via the communication interface 130. This interface can use a standardized protocol such as Modbus, OPC UA, or a proprietary interface of the 200 wind turbine manufacturer.

[0116] The interface can transmit additional information such as the type of activity detected, the distance and direction of the vehicle, the level of confidence of the classification, or the timestamp of the detection.

[0117] The System 100 is equipped with a wireless or wired communication module enabling real-time transmission of alerts, images, and analyzed data to a remote control station or operator, using secure protocols (e.g., TLS, VPN). An intuitive user interface is available, offering graphical visualizations of monitoring data, throttling alerts, and a configurable dashboard for managing parameters and monitoring system status. The interface is designed for access via mobile devices, tablets, or computers, ensuring flexibility and responsiveness in monitoring bird control activities.

[0118] –– Operating modes

[0119] The System 100 can operate in several configurable modes. An automatic mode triggers the speed limiter automatically upon detection of activity. A supervised mode notifies a human operator who then decides when to apply the limiter. A learning mode allows data to be recorded without any speed limiter action, for calibration purposes.

[0120] –– Time window validation

[0121] An optimal time window of 2 to 4 hours is envisaged according to one embodiment of the invention. This is based on a rigorous ornithological analysis corresponding to the peak periods of raptor presence, allowing for effective preventive curb control that minimizes the impact on energy production.

[0122] The average time it takes for the first birds of prey to arrive after plowing begins is between 45 and 90 minutes. The peak presence of birds of prey occurs between 2.5 and 3.5 hours after the start of agricultural activity. The average duration of an agricultural activity such as plowing or harvesting is between 2 and 6 hours. Birds of prey remain in the area for between 30 and 60 minutes after the activity has ended.

[0123] Preventive bridling started within the 2 to 4 hour window allows us to cover the period of maximum risk corresponding to the peak presence of birds of prey, while limiting production losses linked to bridling too early, and ensuring a safety margin with a start of bridling before the massive arrival of birds.

[0124] Thus, the optimization of the curtailment signal incorporates a careful balance between maximizing the protection of bird biodiversity and minimizing energy production losses. The safety margin built into process 300 ensures that curtailment begins sufficiently in advance to cover the period of maximum risk without starting prematurely, thereby avoiding excessive curtailment. This approach ensures precise and adaptive curtailment management, respecting the operational constraints of wind farms while maximizing environmental efficiency.

[0125] Application examples

[0126] – First example – Plough detection

[0127] A tractor equipped with a plow enters the monitoring zone 550 meters from wind turbine 200. System 100 detects the tractor with a confidence score of 0.96 and verifies it as an agricultural vehicle by detecting the attached plow. Time tracking maintains the tractor's identity for 25 seconds despite a temporary occlusion. Contextual classification analyzes the extended area and identifies turned soil with an irregular texture (score: 0.92), the presence of parallel furrows (score: 0.88), and a plow in working configuration (score: 0.94). The final classification identifies plowing in progress with an overall confidence of 0.91. System 100 generates a curtailment signal 60 seconds after the initial detection. Wind turbine 200 reduces its power and then shuts down. The tractor continues plowing for 3 hours and 15 minutes. Curtailment is maintained for up to 30 minutes after the end of detection. No birds were impacted during the risk period.

[0128] – Second example – Non-detection of simple traffic

[0129] A tractor is traveling on a paved access road 400 meters from wind turbine 200, without engaging in any agricultural activity. System 100 detects the tractor and verifies it as an agricultural vehicle. The contextual classification analyzes the extended area and identifies a paved road (masked as an irrelevant area), the absence of tilled soil, the absence of attached agricultural equipment, and a high speed of 25 km / h, incompatible with agricultural work. The final classification identifies simple traffic without productive activity. System 100 does not generate a speed restriction signal. No unintended speed restrictions are applied.

[0130] – Third example – Harvest detection

[0131] A combine harvester enters a wheat field 480 meters from wind turbine 200. After detection and verification, the contextual classification identifies flattened vegetation in the path, a difference between cut and uncut areas, and the combine's working configuration. The final classification identifies harvesting in progress with a confidence level of 0.89. A curtailment signal is generated. Wind turbine 200 is curtailed for the duration of the harvest (2 hours 30 minutes) plus a 30-minute safety margin.

[0132] Uses and practical applications

[0133] The invention is primarily applicable to the prevention and management of bird nuisance in wind farms, by enabling the anticipation of periods of high bird activity through real-time analysis of agricultural vehicle behavior in the monitored area. The System 100 detects activities likely to attract birds and dynamically adjusts the curb signal of the wind turbine equipment to reduce the risk of collisions, while maintaining optimal energy performance outside of critical periods. This application effectively balances environmental protection with energy production requirements, limiting negative impacts on operations.

[0134] Process 300 anticipates agricultural activity over a defined time window, for example, between 2 and 4 hours, allowing for optimized scheduling of wind turbine curtailment. This forecast is based on the analysis of successive image sequences and the history of detected activities, incorporating predictive models adapted to the local and seasonal context. Thus, system 100 anticipates periods of increased risk and proactively adjusts curtailment levels, ensuring a more effective and better-calibrated response to fluctuations in agricultural activity.

[0135] Multi-wind turbine extension

[0136] In one embodiment, a single centralized system 100 simultaneously monitors several wind turbines in a wind farm. The computing unit 120 receives video feeds from multiple camera arrays 110 and generates individualized curtailment signals for each wind turbine 200 based on its proximity to the detected activity.

[0137] Although specific algorithms have been mentioned as examples, the invention covers any implementation using object detection neural networks, machine learning methods, and temporal tracking algorithms. In particular, the invention covers emerging architectures such as Vision-Language models combining visual analysis and semantic understanding, LSTM networks for temporal sequential analysis, Transformers architectures for detection and classification, or any future technological developments in artificial intelligence.

[0138] The number of 110 cameras is not limited. The 100 system can operate with a single 110 camera providing partial coverage, with four to six 110 cameras representing a compromise between coverage and cost, or with twelve or more 110 cameras ensuring maximum redundancy. Resolutions can vary depending on the required range.

[0139] Although the invention is described in the context of onshore wind farms located near agricultural areas, it can be applied to any terrestrial environment where productive agricultural activities (plowing, harvesting, mowing, sowing, spreading) create a bird risk by attracting birds of prey, including ground-mounted solar installations located in agricultural areas or energy infrastructure located in similar rural contexts.

[0140] The invention also covers all technically possible combinations of the described features, even if these combinations are not explicitly listed.

Claims

A method (300) for bird protection by indirect detection for wind turbine curtailment, comprising: - the acquisition (310) of images of a surveillance area around at least one wind turbine (200) by at least one camera (110) having sufficient resolution to allow the detection of agricultural vehicles in the surveillance area; - the detection (320) of moving objects corresponding to vehicles by machine learning means for object detection; - the verification (330) of classification of moving objects by a pre-trained machine learning means configured to extract visual features and produce a confidence measure in order to reduce false positives; - the temporal tracking (340) of detected moving objects to maintain their identity through a sequence of images, the temporal tracking using an object association and trajectory prediction algorithm.- the contextual classification (350) of at least one agricultural activity by means of a contextual analysis integrating an extended peripheral spatial and temporal environment of the detected mobile object and including: -- the extension of an analysis area beyond a bounding box (10) of the detected mobile object, so as to include a surrounding peripheral area (20); -- the capture of an environmental context including at least one element selected from: soil surface condition, presence of agricultural equipment, terrain modifications, vegetation characteristics; -- temporal analysis by averaging information from an image sequence to improve classification robustness; -- the distinction between agricultural vehicles performing a productive activity and non-agricultural vehicles; -- the identification of at least one agricultural activity.- the generation (360) of a control signal for curtailing at least one wind turbine (200) according to the contextual classification obtained in the contextual classification step. Method (300) according to claim 1, wherein the classification verification at the verification step uses a pre-trained machine learning model for visual feature extraction coupled with a binary supervised classifier, including but not limited to threshold classifiers, support vector machines or neural networks. Method (300) according to any one of claims 1 or 2, wherein the time tracking at the time tracking stage includes management of temporary occlusions of moving objects for a configurable duration and by means of trajectory prediction. Method (300) according to any one of claims 1 to 3, wherein the contextual classification at the contextual classification step includes masking irrelevant areas, defined by mapping and / or by machine learning, and selected from: roads, buildings, forest areas, aquatic areas. Method (300) according to any one of claims 1 to 4, wherein the extension of the analysis area at the contextual classification step represents between 20% and (200)% of the surface of the bounding box (10) of the detected moving object. Method (300) according to any one of claims 1 to 5, wherein the temporal analysis at the contextual classification stage is performed on a sequence comprising between 10 and 50 consecutive images acquired at a predefined acquisition rate. Method (300) according to claim 6, wherein:- the sequence comprises 25 consecutive images acquired at a predefined acquisition rate, and- the agricultural activity is selected from a group comprising at least: plowing, stubble cultivation, harvesting, mowing, haymaking, application of plant protection products, sowing, spreading. Bird protection system (100) for the curb control of wind turbines, comprising: - at least one camera (110) configured to acquire images of a surveillance zone around at least one wind turbine (200), the at least one camera (110) having sufficient resolution to allow the detection of agricultural vehicles in the surveillance zone; - a computing unit (120) configured to implement the method (300) according to any one of claims 1 to 7, the computing unit (120) comprising: - a moving object detection module implemented by machine learning means; - a contextual analysis module configured for the classification of agricultural activities; - a module for distinguishing between agricultural and non-agricultural vehicles by detecting characteristic elements selected from: large diameter wheels, tracks, towed implements, raised cab,- a communication interface (130) with a control system for at least one wind turbine (200) to transmit a curtailment signal. System (100) according to claim 8, comprising,- a plurality of cameras (110) with a resolution greater than or equal to 4K distributed to ensure (360)° coverage of the surveillance area,- a centralized computing unit (120) positioned at the base of the wind turbine mast (200). System (100) according to claim 9, comprising exactly 8 cameras (110) distributed equidistantly around the wind turbine (200). System (100) according to any one of claims 8 to 10, wherein the cameras (110) are positioned under the nacelle of the wind turbine (200) and oriented so as to cover an area extending from the base of the mast over a distance allowing the temporal anticipation of the bird risk, typically between (400) meters and (1000) meters. System (100) according to any one of claims 9 to 11, wherein the communication network between the cameras (120) and the computing unit (110) is configured as a self-adaptive mesh network ensuring sub-millisecond temporal synchronization between the cameras (110), and wherein the system (100) is configured to automatically compensate for propagation delays and dynamically recalibrate the temporal drifts of the camera clocks (12(10)), creates a dynamic 4D representation space in which the data from each camera (110) are projected into a common volumetric reference frame, and implements a tensor fusion algorithm combining spatial and temporal information, enabling predictive geometric reconstruction to resolve occlusions of detected moving objects. Use of a system (100) for detecting and analyzing the context of agricultural activities for the automatic curtailment of at least one wind turbine (200) based on the risk of bird collision, characterized in that: - the system (100) includes at least one camera (110) and a computing unit (120) configured to detect agricultural vehicles and classify agricultural activities by contextual analysis of a peripheral spatial environment of the vehicles, - the curtailment is triggered in response to the detection of at least one productive agricultural activity selected from: plowing, harvesting, mowing, sowing, spreading, application of plant protection products, - the use allows an anticipation of the bird risk of between 2 and 4 hours before a massive arrival of birds attracted by the agricultural activity. Use according to claim 13, implementing a system (100) according to any one of claims 8 to 12.

Images