Unmanned aerial vehicle forest damage monitoring method and system based on time perception attention

Abstract

The invention belongs to the technical field of data processing, and provides an unmanned aerial vehicle forest damage monitoring method and system based on time perception attention, and the method specifically comprises the steps: collecting images through an unmanned aerial vehicle, constructing a forest damage image sequence library, constructing a pest perception model based on time perception attention through the forest damage image sequence library, and obtaining a pest related vector; constructing a K-layer perceptron according to the insect pest correlation vector, extracting an image descriptor from the K-layer perceptron, obtaining an insect pest monitoring degree, collecting an insect pest image through the insect pest monitoring degree, marking an insect pest region, continuously obtaining the insect pest monitoring degree for the insect pest region, and calculating a forest pest occurrence track number; and finally, judging a forest damage monitoring result according to the forest damage occurrence track number. Modeling of insect population density fluctuation and control hysteresis effect on a micro scale is realized, so that hierarchical memory of historical intervention response and inflection point sensitivity is enhanced by using time sequence perception, power consumption and computing power are released, storage resources are saved, cruising ability is improved, and signal anxiety in a forest scene is prevented.

Description

Technical Field

[0001] The present invention belongs to the field of forest damage identification and UAV patrol technology, and specifically relates to a UAV forest damage monitoring method and system based on time-aware attention. Background Art

[0002] As drone technology matures, it's increasingly used as a tool for forest health monitoring. Thanks to its highly flexible and low-cost deployment, it significantly improves the efficiency of forest pest monitoring. However, current drone-based pest monitoring technology faces significant bottlenecks. One is geographical limitations. Common drone pest monitoring methods rely on image transmission. For example, the forestry pest monitoring drone with publication number CN114313260B requires transmitting camera-captured image data to a terminal control device to determine whether trees in the forest are infested with pests or diseases. Therefore, these drones are still constrained by signal range. From a forest pest monitoring perspective, large forests and limited signal infrastructure are common challenges. A second challenge is drone endurance, including battery life and storage capacity. These lightweight resources are the hardware foundation that effectively unleashes the monitoring capabilities of drones for forest pest monitoring. Directly capturing and storing image data often requires sacrificing image clarity or flight time to ensure adequate coverage of a specific area. This significantly reduces the efficiency of drone-based forest pest monitoring, both in terms of lower-resolution image data and more frequent monitoring missions. Third, the collected image data is typically single-point data without validity checks, resulting in a large amount of redundant data during drone monitoring, and low utilization of storage and computer resources. Fourth, current drone forest damage monitoring methods are based on forest damage detection without backtesting prevention and control measures. This means that the transformation trajectory of pests during the prevention and control process still relies on graphical comparisons between time nodes, while ignoring the consideration of the evolutionary process, making the monitoring function always incomplete in terms of prevention and control monitoring capabilities. Therefore, a drone forest damage monitoring method and system based on time-aware attention is urgently needed. Summary of the Invention

[0003] The purpose of the present invention is to propose a method and system for monitoring forest damage using a drone based on time-aware attention, so as to solve one or more technical problems existing in the prior art and at least provide a beneficial option or create conditions.

[0004] To achieve the above objectives, according to one aspect of the present invention, a method for monitoring forest damage using a drone based on time-aware attention is provided, the method comprising the following steps: S100, UAV collects images to build a forest damage image sequence library; S200, using the forest damage image sequence library to build a pest perception model based on time-aware attention to obtain pest related vectors; S300, constructs a K-layer perceptron based on pest-related vectors; S400, extracting image descriptors from the K-layer perceptron and obtaining the pest monitoring degree; S500, collects pest images and marks pest areas through pest monitoring; S501, continuously obtaining the pest monitoring degree of the pest-infested area and calculating the number of forest damage occurrence trajectories; S502, judging the forest damage monitoring result according to the number of forest damage occurrence trajectories.

[0005] Furthermore, in step S100, the method for collecting images by drones to construct a forest damage image sequence library is as follows: setting a flight route and a monitoring period according to the locations of known pests in the forest area, collecting images by drones at regular intervals every day to form an image sequence instance; a sequence composed of various pest images at the same location arranged in chronological order to form an image sequence instance, and the forest damage image sequence library is composed of several pest image sequence instances.

[0006] The forest pest image sequence library only corresponds to one type of forest pest, which needs to have obvious visual symptoms, clear development stages, and strong regional concentration. The types of forest pests may include pine brown beetles, gypsy moths, white poplar loopers, or pine caterpillars.

[0007] The image sequence example includes a start frame, an outbreak frame, and a prevention end frame. The start frame represents the time point image when pest symptoms or suspected symptoms first appear in the image sequence. The outbreak frame is the time point image when pest symptoms spread rapidly and reach a significant level in the image sequence. The prevention end frame is the time point image when pest symptoms are basically alleviated after human intervention, including spraying, insect nets, or forest management.

[0008] The flight path planning for the drone is set at an altitude of 30-60m (with a default value of 30m), a ground sampling interval of 3-5cm / pixel (with a default value of 5cm / pixel), a camera field of view of 84° horizontally and 63° vertically, a forward overlap of greater than or equal to 70% (with a default value of 70%), and a lateral overlap of greater than or equal to 60% (with a default value of 60%). The path mode is a zig-zag serpentine route.

[0009] The purpose of this step is to provide a sample database of forest pest image sequences with clear pest types and accurate temporal and spatial continuity for the subsequent training and deployment of temporal pest perception models. This database serves as the basic input for the training of temporal perception deep learning models, enabling the model to capture the temporal image evolution patterns of various types of pests at different development stages.

[0010] Furthermore, in step S200, a method for constructing a pest perception model based on time-aware attention using the forest damage image sequence library to obtain pest-related vectors is as follows: after preprocessing the forest damage image sequence library, the image sequence is trained using a deep learning model including a time-aware attention mechanism to obtain the pest perception model, wherein the time-aware attention mechanism adopts any one of the TimeSformer operator, the ConvLSTM operator, or the Temporal Attention Network operator; and the output result of the intermediate embedding layer of the pest perception model is extracted as a pest-related vector, which is mathematically expressed as: , where d is the feature dimension, and the value is 128 or 256.

[0011] Any image sequence instance needs to be preprocessed, including image alignment, uniform size adjustment and color normalization.

[0012] The model architecture of the time-aware deep learning model uses any of the TimeSformer, ConvLSTM, or Temporal Attention Network operators. The TimeSformer operator is used by default. This operator is a common operator specifically used to process video sequences or multi-temporal image sequences. Because the development process of insect pests is usually not linear and stable but contains mutation nodes, the TimeSformer operator has a long-range attention mechanism in the temporal dimension, which is conducive to capturing discontinuous but causally related insect pest dynamics across time.

[0013] The output pest perception model is a trained pest perception model, whose prediction features are occurrence confidence and prevention confidence; occurrence confidence refers to the percentile value of an image corresponding to a time point between the start frame and the outbreak frame, and prevention confidence refers to the percentile value of an image corresponding to a time point between the outbreak frame and the prevention end frame; the prevention confidence of each image between the start frame and the outbreak frame is 0, and the occurrence confidence of each image between the outbreak frame and the prevention end frame is 1.

[0014] During the training process of the pest perception model, the loss function includes cross-entropy loss, temporal consistency regularization term, and attention sparsity penalty term. The optimizer uses the Adam operator, the learning rate is set to 1e-4, and the number of training rounds is 100. In each round, the data of the forest damage image sequence library is divided into 70% for training, 15% for validation, and 15% for testing. Furthermore, in step S300, the method of constructing a K-layer perceptron based on the pest-related vector is: taking the pest-related vector as input to construct a feedforward neural network model comprising at least K fully connected layers as a K-layer perceptron, wherein each hidden layer adopts a ReLU activation function, and the output layer output includes occurrence confidence and control confidence, which are used to represent the temporal position state of the input image during the pest development process.

[0015] In the structural design process of the K-layer perceptron, the input layer dimension is consistent with the pest-related vector dimension, the number of hidden layers is K-2, K≥3, the default value is 5, each layer adopts a fully connected structure, and uses the ReLU activation function; the output layer dimension is 2, corresponding to the occurrence confidence and the control confidence; the loss function set for the K-layer perceptron adopts the mean square error loss, which supervises the occurrence confidence and the control confidence respectively, with a learning rate of 1e-4, and is optimized by the Adam operator. The generalization ability is enhanced by the Dropout and BatchNorm operators.

[0016] As a lightweight convolutional module, the K-layer perceptron can be efficiently applied to drone work scenarios where power consumption and computing power are constrained. The K-layer perceptron can implement lightweight and fast judgment on the drone side, which not only releases power consumption and computing power, but also saves storage resources and improves battery life. It can also prevent signal anxiety in forest scenes and can work normally even in forest scenes with poor communication conditions or no signal towers.

[0017] Furthermore, in step S400, the method for extracting image descriptors from the K-layer perceptron and obtaining the pest monitoring degree is: the drone inputs the collected image into the K-layer perceptron to obtain the image descriptor and obtains the occurrence confidence and prevention confidence of real-time monitoring through nonlinear mapping processing, which are defined as the occurrence monitoring degree and the prevention monitoring degree, respectively, and both are classified as pest monitoring degrees.

[0018] The obtained image descriptor has the same format as the pest-related vector and is used as the intermediate vector inside the perceptron. Its mathematical expression is: , where d is the feature dimension, and the value is 128 or 256; the ReLU activation function is used by default for nonlinear processing.

[0019] By unifying descriptor extraction and confidence rate calculation within a K-layer perceptron, a simplified edge deployment model structure is achieved, making it more suitable for real-time operation scenarios of drones. The K-layer perceptron structure ensures the independent operation of the entire model, making it more friendly to subsequent optimization or updates.

[0020] Furthermore, in step S500, the method of collecting pest images and marking pest areas through pest monitoring degrees is: when the UAV performs a forest monitoring task, the pest monitoring degrees are calculated for the images obtained in real time. If an area is not marked as a pest area, and the occurrence monitoring degree in the pest monitoring degree of the image corresponding to the area exceeds 10%, then the area is marked as a pest area, and the obtained image is stored and recorded as a pest image; otherwise, the image is not stored.

[0021] The computer activity of marking the area as an infested area includes appending the GPS location to the image metadata and subsequently storing it on a server and subsequently updating it to a GIS layer file in the infested area map.

[0022] This step dynamically determines the occurrence monitoring degree, controls the image storage and area marking behavior, and realizes the accurate collection of pest images and spatial modeling of pest areas. It is a key link in the leap from the UAV's time perception capability to spatial intelligence capability, avoids the storage of a large number of invalid or normal area images, saves equipment storage resources, and reduces the pressure of data return.

[0023] Furthermore, in step S501, the method for continuously acquiring the pest monitoring degree and calculating the number of forest pest occurrence trajectories in the pest-infested area is: During routine drone patrols, pest monitoring degrees are obtained from marked pest areas every natural day, including occurrence monitoring degrees and control monitoring degrees. The obtained pest monitoring degrees usually take values ​​between 0 and 1. The occurrence monitoring degree is the probability of determining whether a pest exists, and the control monitoring degree is the probability of the degree of improvement in control after human intervention. Both probabilities are determined based on the time perception model.

[0024] Assume a time period as the monitoring period Qmet, Qmet∈[10,30] natural days. The pest monitoring degree is obtained by the drone during the scheduled patrol process of each natural day. The pest monitoring degree is stored as a binary array in the computer logic. The occurrence monitoring degree and control monitoring degree within the monitoring period are obtained. The distance between the maximum and minimum values ​​of the occurrence monitoring degree and the median is calculated respectively, and the maximum value is used as the oscillation limit number. The distances between the maximum and minimum values ​​and the median refer to the difference between the maximum and median, and the difference between the median and the minimum, respectively. The oscillation threshold is calculated based on sufficient data stability. The absolute difference between the median representing the intermediate trend and the difference between the maximum and minimum values ​​is calculated through a double-extreme difference comparison. Upward fluctuations and downward deviations are dynamically monitored, and the larger of the two is taken as the threshold to balance trends and fluctuations, avoid unilateral influences, and capture nonlinear mutations in pest population outbreaks. During the monitoring period, the absolute value of the difference between the control monitoring degrees of any two consecutive natural days is recorded as the control step number, and the average of all control step numbers is the step average; The calculation of the step average is to measure the average improvement of the prevention and control effect by calculating the average of the absolute value of the change in the prevention and control monitoring degree for two consecutive natural days, providing a reference for the effectiveness of the subsequent prevention and control measures. If the increase in the occurrence monitoring degree of a natural day compared with the first natural day in the reverse time direction is greater than the oscillation limit number or the decrease in the prevention and control monitoring degree is greater than the step average, then the natural day is recorded as a turning mark node; The increase in occurrence monitoring degree here refers to the difference between the occurrence monitoring degree of the current natural day and the occurrence monitoring degree of the previous natural day; The principle of determining the turning mark node is that when the monitoring degree increases too much, the risk of infestation increases; when the prevention and control monitoring degree decreases too much, it means that the improvement effect of prevention and control measures is poor. Based on the influence of these two variables, early warning of key time nodes is used to further enhance the influence of such key time nodes on the assessment of the number of forest damage occurrence trajectories, so as to accurately measure the number of forest damage occurrence trajectories in the current natural day. Set a time window pw, pw∈[5,7] natural days; the default value is 5 natural days, and calculate the difference between the monitoring degree of any natural day and pw natural days in the reverse time direction as the local change degree Lvdeg; A short-term window is selected to analyze recent trends in dynamic cumulative risk, eliminating long-term noise interference. A lower limit of 5 days for pw can cover periods of significant changes in pest activity, while an upper limit of 7 days can avoid trend smoothing caused by excessively long windows, thereby focusing on short-term dynamics. Local changes are used to quantify the pest situation over continuous time and the impact of control measures, thereby assessing short-term pest trends. By setting the penalty weight value of the turn-marked node to 1 and the rest to 0; For the current natural day, the number of forest damage occurrence trajectories Fsera is calculated based on the local change degree and penalty weighted function: ; Where t1 is the cumulative variable, Lvdeg t1 Opcmr is the difference between the monitoring degree of the current natural day and the natural day t1 in the reverse time direction. t1is the control monitoring degree of the t1th natural day in the reverse time direction of the current natural day, exp() is an exponential function with the natural constant e as the base; Opcmr represents the control monitoring degree of the current natural day; I(t1) is the penalty weighted value; the weight setting growth rate in the penalty term is lower than the linear growth rate, which is based on the ecological rhythm planning in which the larger the time window in natural evolution, the slower the weight of the single-day impact, so as to avoid the distortion or misjudgment of the total effect under long-term observation. If the penalty weighted value of the pest with a short control cycle is set to I(t1)=1, if the control cycle is long and takes more than one month, the penalty weighted value is calculated by the penalty weighted function: for the t1th natural day, the penalty weighted function is recorded as Here, t1 is the serial number of each natural day after the start of collecting pest monitoring data; The calculation principle of the number of forest damage occurrence trajectories is based on the local fluctuation characteristics and intervention response trends of the pest occurrence monitoring degree and the prevention and control monitoring degree in consecutive natural days within a short-term time window, simulating the interaction mechanism between "pest population dynamics-prevention and control intervention-residual rebound" in the development process of forest damage.

[0025] In natural forest ecosystems, the outbreak and decline of insect pests are often governed by highly time-sensitive ecological rhythms, including temperature, humidity, life cycle stage, and the delayed effects of control measures. Pest activity often exhibits nonlinear characteristics with short-term mutations and delayed responses. Therefore, by setting a proximal time window to cover microcyclic fluctuations in pest activity and the feedback cycle following intervention, we can focus on extracting dynamic trends in short-term pest risk. Simultaneously, we construct a cumulative interaction term between the degree of local change and the degree of daily control monitoring, and introduce a weighted accumulation mechanism to simulate the ecological process in which the accumulation of pest fluctuations drives the malignant expansion of forest pests.

[0026] It's important to note that in natural systems, if pest control fails within a short period of time, a reversal can easily occur. For example, pests like pine caterpillars and gypsy moths tend to repopulate within days after being dispersed. Therefore, this method uses reversal markers to identify these critical rebound nodes. Using a penalty-weighted function, it increases the contribution of these key time points to the number of occurrence trajectories, thereby more closely resembling the post-stress re-outbreak mechanisms of natural evolution.

[0027] Since the above-mentioned turning mark nodes are determined by the occurrence monitoring degree and prevention and control monitoring degree of the perception inversion at each time during the monitoring period, it can accurately capture the natural days that need to be paid attention to when the pest infestation is aggravated or the improvement effect of the prevention and control measures is poor. Therefore, the changing trend of pests and the improvement effect of the prevention and control measures in the short term can be targeted and weighted, and the role of the turning mark nodes can be amplified. Finally, the dynamic evaluation of the number of forest damage occurrence trajectories can be accurately performed. However, since the number of forest damage occurrence trajectories is also greatly correlated with the natural days with significant prevention and control effects, only amplifying the role of the turning mark nodes with greater risks will reduce the accuracy of calculating the number of forest damage occurrence trajectories. At the same time, relying solely on the static comparison of the threshold to highlight the key time nodes cannot dynamically reflect the residual impact of historical prevention and control effects on current risks. In order to improve the accuracy of evaluating the number of forest damage occurrence trajectories, the present invention proposes a more preferred solution. Furthermore, in step S501, the method for continuously acquiring the pest monitoring degree and calculating the number of forest pest occurrence trajectories in the pest-infested area is: Assume a time period as the monitoring period Qmet, Qmet∈[15,30] natural days. The pest monitoring degree, namely the occurrence monitoring degree Dgrcu and the control monitoring degree Opcmr, is obtained through the scheduled patrol of the drone on each natural day. On any natural day during the monitoring period, the ratio of the control monitoring degree to the occurrence monitoring degree is recorded as the pest control efficiency index Ipcoe. The relative efficiency of control is quantified by the ratio of the improvement degree of control effect to the probability of pest occurrence. It is used to evaluate the effectiveness of control measures. The higher the control efficiency index, the stronger the control effect. On the contrary, the control intensity is insufficient and the risk accumulates. The product of the remaining control monitoring degree and the occurrence monitoring degree is calculated as the relative occurrence product, Rmari, where the remaining control monitoring degree is the difference between 1 and the control monitoring degree. If the control monitoring degree is greater than 1, the data within the monitoring period needs to be normalized. Rmari = Dgrcu × (1 - Opcmr). By calculating the nonlinear interaction between pest occurrence and control monitoring effectiveness, the pest risk remaining after control is quantified, providing a basis for the subsequent calculation of the dynamic cumulative occurrence risk. Set the time window pw, whose value range is pw ∈ [5,7] natural days, and the default value is 5. For any natural day, calculate the dynamic cumulative occurrence risk based on the pest control efficiency index and the relative occurrence volume; Dymlr,

[0028]

[0029] Where t2 is the cumulative variable, Ipcoe t2 、Rmari t2 are the pest control efficiency index and relative occurrence cumulative amount on the t2th natural day in the reverse time direction of the current natural day, respectively. exp() is an exponential function with the natural constant e as the base; The calculation principle of the dynamic cumulative occurrence risk is as follows: by considering the impact of the historical prevention and control efficiency index and using exponential decay weighting, the natural day with a higher prevention and control efficiency index has a faster decay of its subsequent impact, and therefore has a smaller contribution to the current cumulative occurrence risk. As the time span increases, the cumulative impact on the risk of the current natural day decreases, so the faster the corresponding weight decays, the more dynamically it can reflect the residual impact of historical prevention and control effects on the current risk; For any time window, a line graph of the dynamic cumulative risk is drawn. The natural day corresponding to the point where the slope changes from negative to positive is recorded as a defect mark point, and the natural day corresponding to the point where the slope changes from positive to negative is recorded as a positive mark point. The difference between the defect mark and the slope of the previous day is recorded as the positive mark quantity Ramdu, and the difference between the positive mark point and the slope of the previous day is recorded as the defect mark quantity Ratco. The difference in the dynamic cumulative risk of adjacent natural days is calculated by taking the first-order difference of the dynamic cumulative risk, providing decision-makers with a time-sensitive signal for identifying the turning point of the risk trend. The change from negative to positive in the line slope indicates accelerated risk accumulation. The difference between the positive and negative slopes is used to quantify the intensity of risk acceleration. Conversely, when the slope changes from positive to negative, it indicates that the risk trend has slowed down or preventive measures have taken effect. Therefore, the difference between the negative and positive slopes can be used to quantify the intensity of risk mitigation and evaluate the actual effect of preventive measures. For any natural day k2, record the slope sign weight function : ; Among them Ramdu k2 、Ratco k2 are the positive marking amount and defect marking amount corresponding to natural day k2 respectively; the calculation purpose of the slope sign weight function is to highlight the impact of risk turning points by dynamically adjusting the contribution of key time nodes, using defect marking points to mark accelerated risk accumulation, and at the same time amplifying its impact range by increasing the weight; on the contrary, when prevention and control improves the natural environment, the risk trend slows down and its impact is reduced by reducing the weight; here k2 is the natural day serial number value; For the current natural day, the number of forest damage occurrence trajectories Fsera is calculated based on the dynamic cumulative occurrence risk and slope sign weight function: ; Among them, t3 is the cumulative variable, Dymlr t3 、 They represent the dynamic cumulative occurrence risk and slope sign weight function value of the t3th natural day counterclockwise from the current natural day, and Dymlr0 represents the dynamic cumulative occurrence risk and slope sign weight function value of the current natural day; The calculation principle of the forest pest occurrence trajectory number is to simulate the natural evolution trend of a typical nonlinear process of local pest outbreak in forest ecosystems, namely, prevention and control intervention, delayed response, re-outbreak or relief, by integrating the dynamic cumulative occurrence risk within the proximal time window, the trend slope weight function, and the time series weight decay mechanism. Its physical semantics are highly consistent with the ecological inertia in natural phenomena. That is, after pest control, pest risk often has a slow decline or rebound effect over time due to factors such as residual insect eggs, delayed host plant recovery, and climate warming. This model effectively captures this biological hysteresis response characteristic by assigning faster decay weights to historical points with high prevention and control efficiency. The model also introduces a marking mechanism based on the direction of changes in the slope of dynamic cumulative risk, and applies a slope sign weighting function to key inflection points to characterize sensitive time nodes of changes in pest risk trends; this mechanism is essentially an abstract mathematical reconstruction of the turning point outbreak or recession threshold turning point in the natural system, which can more realistically reflect the fluctuation amplification effect of risks in space and time, and form a comprehensive trend judgment of the current stage of pest development.

[0030] Beneficial Effects: Because the number of forest pest occurrence trajectories is a continuous assessment of pest monitoring intensity over time, it can effectively quantify the difficulty in objectively determining the monitoring endpoint, which arises from the fuzzy changes in residual pest risk during the post-control intervention phase. This reduces the inefficiency and risk of misoperation in traditional pest control monitoring, which relies heavily on manual operations. It effectively explains the occurrence degree of the control system endpoint from the perspective of trajectory rationality, thereby forming a closed-loop UAV monitoring process and improving the autonomy and automation of UAV forest pest monitoring. By introducing a pest control efficiency index and relative occurrence volume, and integrating time-series attenuation, slope mutation labeling, and weight reconstruction mechanisms, microscale modeling of insect population density fluctuations and control lag effects is achieved. This leverages time-series perception to enhance the hierarchical memory of historical intervention responses and inflection point sensitivity. At the macro level, a structural trajectory quantification of forest pest evolution trends is formed, providing decision-making self-evolution capabilities for ecological risk management. This model is endowed with a time-driven evolutionary understanding, enabling dynamic prediction and response discrimination rather than static estimation, thus establishing an intelligent forest pest monitoring paradigm for a closed-loop control loop.

[0031] Furthermore, in step S502, the method for judging the forest damage monitoring result based on the number of forest damage occurrence trajectories is: when the number of forest damage occurrence trajectories reaches a minimum value and the number of forest damage occurrence trajectories is higher than the minimum value within at least 15 natural days thereafter, the forest damage monitoring result of the area is judged to be forest damage elimination, and the insect pest area mark of the area is cancelled.

[0032] Preferably, all undefined variables in the present invention, if not clearly defined, can be manually set thresholds.

[0033] The present invention also provides a time-aware attention-based drone forest damage monitoring system, which includes: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the time-aware attention-based drone forest damage monitoring method are implemented. The time-aware attention-based drone forest damage monitoring system can be run on computing devices such as desktop computers, laptop computers, PDAs, and cloud data centers. The executable system may include, but is not limited to, a processor, a memory, and a server cluster. The processor executes the computer program in the following system units: A forest damage image sequence library construction unit is used to collect images from drones to construct a forest damage image sequence library; A time-aware modeling unit is used to construct a pest perception model based on time-aware attention using a forest damage image sequence library to obtain pest-related vectors; A perceptron construction unit, used to construct a K-layer perceptron based on pest-related vectors; Perception data inversion unit, used to extract image descriptors from the K-layer perceptron and obtain the pest monitoring degree; The pest area marking unit is used to collect pest images and mark the pest area through pest monitoring; A forest damage occurrence trajectory number calculation unit is used to continuously obtain the pest monitoring degree of the pest-infested area and calculate the number of forest damage occurrence trajectories; The prevention and control feedback unit is used to judge the forest damage monitoring results based on the number of forest damage occurrence trajectories; The beneficial effects of the present invention are: the present invention provides a method and system for monitoring forest damage by a drone based on time-aware attention, which realizes lightweight and rapid judgment on the drone side through the K-layer perception function, releases power consumption and computing power, saves storage resources and improves endurance, prevents signal anxiety in forest scenes, effectively quantifies the problem of difficulty in objectively judging the monitoring termination node caused by the fuzzy changes in the risk of residual pests in the forest damage monitoring stage after prevention and control intervention, reduces the inefficient execution and misoperation risks of traditional pest control monitoring that is highly dependent on manual operations, effectively explains the degree of occurrence of the endpoint of the prevention and control system from the rationality of the trajectory, and thus forms a closed loop of the drone monitoring process, thereby improving the autonomy and automation capabilities of drone forest damage monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The above and other features of the present invention will become more apparent through a detailed description of the embodiments shown in conjunction with the accompanying drawings. In the drawings of the present invention, the same reference numerals represent the same or similar elements. Obviously, the drawings described below are only some embodiments of the present invention. It is possible for a person skilled in the art to derive other drawings based on these drawings without inventive effort. In the drawings: Figure 1 The figure shows a flow chart of a method for UAV forest damage monitoring based on time-aware attention; Figure 2 Shown is a structural diagram of a UAV forest damage monitoring system based on time-aware attention. DETAILED DESCRIPTION

[0035] The following will be combined with the embodiments and drawings to clearly and completely describe the concept, specific structure and technical effects of the present invention so as to fully understand the purpose, scheme and effect of the present invention. It should be noted that the embodiments and features in the embodiments of this application can be combined with each other unless there is a conflict.

[0036] like Figure 1 The following is a flow chart of a method for monitoring forest damage using drones based on time-aware attention. Figure 1 A method for monitoring forest damage using a drone based on time-aware attention according to an embodiment of the present invention is described below. The method comprises the following steps: S100, UAV collects images to build a forest damage image sequence library; S200, using the forest damage image sequence library to build a pest perception model based on time-aware attention to obtain pest related vectors; S300, constructs a K-layer perceptron based on pest-related vectors; S400, extracting image descriptors from the K-layer perceptron and obtaining the pest monitoring degree; S500, collects pest images and marks pest areas through pest monitoring; S501, continuously obtaining the pest monitoring degree of the pest-infested area and calculating the number of forest damage occurrence trajectories; S502, judging the forest damage monitoring result according to the number of forest damage occurrence trajectories.

[0037] Furthermore, in step S100, the method for collecting images by drones to construct a forest damage image sequence library is as follows: setting a flight route and a monitoring period according to the locations of known pests in the forest area, collecting images by drones at regular intervals every day to form an image sequence instance; a sequence composed of various pest images at the same location arranged in chronological order to form an image sequence instance, and the forest damage image sequence library is composed of several pest image sequence instances.

[0038] The forest pest image sequence library only corresponds to one type of forest pest, which needs to have obvious visual symptoms, clear development stages, and strong regional concentration. The types of forest pests may include pine brown beetles, gypsy moths, white poplar loopers, or pine caterpillars.

[0039] The image sequence example includes a start frame, an outbreak frame, and a prevention end frame. The start frame represents the time point image when pest symptoms or suspected symptoms first appear in the image sequence. The outbreak frame is the time point image when pest symptoms spread rapidly and reach a significant level in the image sequence. The prevention end frame is the time point image when pest symptoms are basically alleviated after human intervention, including spraying, insect nets, or forest management.

[0040] The flight path planning for the drone is set at an altitude of 30-60m (with a default value of 30m), a ground sampling interval of 3-5cm / pixel (with a default value of 5cm / pixel), a camera field of view of 84° horizontally and 63° vertically, a forward overlap of greater than or equal to 70% (with a default value of 70%), and a lateral overlap of greater than or equal to 60% (with a default value of 60%). The path mode is a zig-zag serpentine route.

[0041] Furthermore, in step S200, a method for constructing a pest perception model based on time-aware attention using the forest damage image sequence library to obtain pest-related vectors is as follows: after preprocessing the forest damage image sequence library, the image sequence is trained using a deep learning model including a time-aware attention mechanism to obtain the pest perception model, wherein the time-aware attention mechanism adopts any one of the TimeSformer operator, the ConvLSTM operator, or the Temporal Attention Network operator; and the output result of the intermediate embedding layer of the pest perception model is extracted as a pest-related vector, which is mathematically expressed as: , where d is the feature dimension, and the value is 128 or 256.

[0042] Any image sequence instance needs to be preprocessed, including image alignment, uniform size adjustment and color normalization.

[0043] The model architecture of the time-aware deep learning model uses any of the TimeSformer, ConvLSTM, or Temporal Attention Network operators. The TimeSformer operator is used by default. This operator is a common operator specifically used to process video sequences or multi-temporal image sequences. Because the development process of insect pests is usually not linear and stable but contains mutation nodes, the TimeSformer operator has a long-range attention mechanism in the temporal dimension, which is conducive to capturing discontinuous but causally related insect pest dynamics across time.

[0044] The output pest perception model is a trained pest perception model, whose prediction features are occurrence confidence and prevention confidence; occurrence confidence refers to the percentile value of an image corresponding to a time point between the start frame and the outbreak frame, and prevention confidence refers to the percentile value of an image corresponding to a time point between the outbreak frame and the prevention end frame; the prevention confidence of each image between the start frame and the outbreak frame is 0, and the occurrence confidence of each image between the outbreak frame and the prevention end frame is 1.

[0045] During the training process of the pest perception model, the loss function includes cross-entropy loss, temporal consistency regularization term, and attention sparsity penalty term. The optimizer uses the Adam operator, the learning rate is set to 1e-4, and the number of training rounds is 100. In each round, the data of the forest damage image sequence library is divided into 70% for training, 15% for validation, and 15% for testing. Furthermore, in step S300, the method of constructing a K-layer perceptron based on the pest-related vector is: taking the pest-related vector as input to construct a feedforward neural network model comprising at least K fully connected layers as a K-layer perceptron, wherein each hidden layer adopts a ReLU activation function, and the output layer output includes occurrence confidence and control confidence, which are used to represent the temporal position state of the input image during the pest development process.

[0046] In the structural design process of the K-layer perceptron, the input layer dimension is consistent with the pest-related vector dimension, the number of hidden layers is K-2, K≥3, the default value is 5, each layer adopts a fully connected structure, and uses the ReLU activation function; the output layer dimension is 2, corresponding to the occurrence confidence and the control confidence; the loss function set for the K-layer perceptron adopts the mean square error loss, which supervises the occurrence confidence and the control confidence respectively, with a learning rate of 1e-4, and is optimized by the Adam operator. The generalization ability is enhanced by the Dropout and BatchNorm operators.

[0047] Furthermore, in step S400, the method for extracting image descriptors from the K-layer perceptron and obtaining the pest monitoring degree is: the drone inputs the collected image into the K-layer perceptron to obtain the image descriptor and obtains the occurrence confidence and prevention confidence of real-time monitoring through nonlinear mapping processing, which are defined as the occurrence monitoring degree and the prevention monitoring degree, respectively, and both are classified as pest monitoring degrees.

[0048] The obtained image descriptor has the same format as the pest-related vector and is used as the intermediate vector inside the perceptron. Its mathematical expression is: , where d is the feature dimension, and the value is 128 or 256; the ReLU activation function is used by default for nonlinear processing.

[0049] Furthermore, in step S500, the method of collecting pest images and marking pest areas through pest monitoring degrees is: when the UAV performs a forest monitoring task, the pest monitoring degrees are calculated for the images obtained in real time. If an area is not marked as a pest area, and the occurrence monitoring degree in the pest monitoring degree of the image corresponding to the area exceeds 10%, then the area is marked as a pest area, and the obtained image is stored and recorded as a pest image; otherwise, the image is not stored.

[0050] The computer activity of marking the area as an infested area includes appending the GPS location to the image metadata and subsequently storing it on a server and subsequently updating it to a GIS layer file in the infested area map.

[0051] Furthermore, in step S501, the method for continuously acquiring the pest monitoring degree and calculating the number of forest pest occurrence trajectories in the pest-infested area is: During routine drone patrols, pest monitoring degrees are obtained from marked pest areas every natural day, including occurrence monitoring degrees and control monitoring degrees. The obtained pest monitoring degrees usually take values ​​between 0 and 1. The occurrence monitoring degree is the probability of determining whether a pest exists, and the control monitoring degree is the probability of the degree of improvement in control after human intervention. Both probabilities are determined based on the time perception model.

[0052] Assume a time period as the monitoring period Qmet, Qmet∈[10,30] natural days. The pest monitoring degree is obtained by the drone during the scheduled patrol process of each natural day. The pest monitoring degree is stored as a binary array in the computer logic. The occurrence monitoring degree and control monitoring degree within the monitoring period are obtained. The distance between the maximum and minimum values ​​of the occurrence monitoring degree and the median is calculated respectively, and the maximum value is used as the oscillation limit number. The distances between the maximum and minimum values ​​and the median refer to the difference between the maximum and median, and the difference between the median and the minimum value, respectively. During the monitoring period, the absolute value of the difference between the control monitoring degrees of any two consecutive natural days is recorded as the control step number, and the average of all control step numbers is the step average; If the increase in the occurrence monitoring degree of a natural day compared with the first natural day in the reverse time direction is greater than the oscillation limit number or the decrease in the prevention and control monitoring degree is greater than the step average, then the natural day is recorded as a turning mark node; The increase in occurrence monitoring degree here refers to the difference between the occurrence monitoring degree of the current natural day and the occurrence monitoring degree of the previous natural day; Set a time window pw, pw∈[5,7] natural days; the default value is 5 natural days, and calculate the difference between the monitoring degree of any natural day and pw natural days in the reverse time direction as the local change degree Lvdeg; By setting the penalty weight value of the turn-marked node to 1 and the rest to 0; For the current natural day, the number of forest damage occurrence trajectories Fsera is calculated based on the local change degree and penalty weighted function: ; Where t1 is the cumulative variable, Lvdeg t1 Opcmr is the difference between the monitoring degree of the current natural day and the natural day t1 in the reverse time direction. t1 is the control monitoring degree of the t1th natural day in the reverse time direction of the current natural day, exp() is an exponential function with the natural constant e as the base; Opcmr represents the control monitoring degree of the current natural day; I(t1) is the penalty weighted value; the weight setting growth rate in the penalty term is lower than the linear growth rate, which is based on the ecological rhythm planning in which the larger the time window in natural evolution, the slower the weight of the single-day impact, so as to avoid the distortion or misjudgment of the total effect under long-term observation. If the penalty weighted value of the pest with a short control cycle is set to I(t1)=1, if the control cycle is long and takes more than one month, the penalty weighted value is calculated by the penalty weighted function: for the t1th natural day, the penalty weighted function is recorded as Here, t1 is the serial number of each natural day after the start of collecting pest monitoring data; The calculation principle of the number of forest damage occurrence trajectories is based on the local fluctuation characteristics and intervention response trends of the pest occurrence monitoring degree and the prevention and control monitoring degree in consecutive natural days within a short-term time window, simulating the interaction mechanism between "pest population dynamics-prevention and control intervention-residual rebound" in the development process of forest damage.

[0053] In natural forest ecosystems, the outbreak and decline of insect pests are often governed by highly time-sensitive ecological rhythms, including temperature, humidity, life cycle stage, and the delayed effects of control measures. Pest activity often exhibits nonlinear characteristics with short-term mutations and delayed responses. Therefore, by setting a proximal time window to cover microcyclic fluctuations in pest activity and the feedback cycle following intervention, we can focus on extracting dynamic trends in short-term pest risk. Simultaneously, we construct a cumulative interaction term between the degree of local change and the degree of daily control monitoring, and introduce a weighted accumulation mechanism to simulate the ecological process in which the accumulation of pest fluctuations drives the malignant expansion of forest pests.

[0054] It's important to note that in natural systems, if pest control fails within a short period of time, a reversal can easily occur. For example, pests like pine caterpillars and gypsy moths tend to repopulate within days after being dispersed. Therefore, this method uses reversal markers to identify these critical rebound nodes. Using a penalty-weighted function, it increases the contribution of these key time points to the number of occurrence trajectories, thereby more closely resembling the post-stress re-outbreak mechanisms of natural evolution.

[0055] Furthermore, in step S501, the method for continuously acquiring the pest monitoring degree and calculating the number of forest pest occurrence trajectories in the pest-infested area is: Assume a time period as the monitoring period Qmet, Qmet∈[15,30] natural days, and obtain the pest monitoring degree, namely the occurrence monitoring degree Dgrcu and the control monitoring degree Opcmr, through the scheduled patrol process of the drone on each natural day. On any natural day during the monitoring period, the ratio of the control monitoring degree to the occurrence monitoring degree is recorded as the pest control efficiency index Ipcoe. The product of the remaining control monitoring degree and the occurrence monitoring degree is calculated as the relative occurrence product Rmari, where the remaining control monitoring degree is the difference between 1 and the control monitoring degree. If the control monitoring degree is greater than 1, the data within the monitoring period needs to be normalized; Rmari = Dgrcu × (1-Opcmr); Set the time window pw, whose value range is pw ∈ [5,7] natural days, and the default value is 5. For any natural day, calculate the dynamic cumulative occurrence risk based on the pest control efficiency index and the relative occurrence volume; Dymlr,

[0056]

[0057] Where t2 is the cumulative variable, Ipcoe t2 、Rmari t2 are the pest control efficiency index and relative occurrence cumulative amount on the t2th natural day in the reverse time direction of the current natural day, respectively. exp() is an exponential function with the natural constant e as the base; Draw a line graph of the dynamic cumulative risk for any time window. The natural day corresponding to the point where the slope changes from negative to positive is recorded as the defect mark point, and the natural day corresponding to the point where the slope changes from positive to negative is recorded as the positive mark point. The difference between the defect mark and the slope of the previous day is recorded as the positive mark quantity Ramdu, and the difference between the positive mark point and the slope of the previous day is recorded as the defect mark quantity Ratco. For any natural day k2, record the slope sign weight function : ; Among them Ramdu k2 、Ratco k2 are the positive marking amount and defect marking amount corresponding to natural day k2 respectively; For the current natural day, the number of forest damage occurrence trajectories Fsera is calculated based on the dynamic cumulative occurrence risk and slope sign weight function: ; Among them, t3 is the cumulative variable, Dymlr t3 、 They represent the dynamic cumulative occurrence risk and slope sign weight function value of the t3th natural day counterclockwise from the current natural day, and Dymlr0 represents the dynamic cumulative occurrence risk and slope sign weight function value of the current natural day; Furthermore, in step S502, the method for judging the forest damage monitoring result based on the number of forest damage occurrence trajectories is: when the number of forest damage occurrence trajectories reaches a minimum value and the number of forest damage occurrence trajectories is higher than the minimum value within at least 15 natural days thereafter, the forest damage monitoring result of the area is judged to be forest damage elimination, and the insect pest area mark of the area is cancelled.

[0058] The embodiment of the present invention provides a UAV forest damage monitoring system based on time-aware attention, such as Figure 2 The figure shows a structural diagram of a time-aware attention-based drone forest damage monitoring system of the present invention. The time-aware attention-based drone forest damage monitoring system of this embodiment includes: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps in the above-mentioned embodiment of the time-aware attention-based drone forest damage monitoring method are implemented.

[0059] The system includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to run in the following units of the system: A forest damage image sequence library construction unit is used to collect images from drones to construct a forest damage image sequence library; A time-aware modeling unit is used to construct a pest perception model based on time-aware attention using a forest damage image sequence library to obtain pest-related vectors; A perceptron construction unit, used to construct a K-layer perceptron based on pest-related vectors; Perception data inversion unit, used to extract image descriptors from the K-layer perceptron and obtain the pest monitoring degree; The pest area marking unit is used to collect pest images and mark the pest area through pest monitoring; A forest damage occurrence trajectory number calculation unit is used to continuously obtain the pest monitoring degree of the pest-infested area and calculate the number of forest damage occurrence trajectories; The prevention and control feedback unit is used to judge the forest damage monitoring results based on the number of forest damage occurrence trajectories.

[0060] The time-aware attention-based drone forest damage monitoring system can be run on computing devices such as desktop computers, laptops, PDAs, and cloud servers. The system that can run the time-aware attention-based drone forest damage monitoring system may include, but is not limited to, a processor and a memory. Those skilled in the art will understand that the example is merely an example of a time-aware attention-based drone forest damage monitoring system and does not constitute a limitation on a time-aware attention-based drone forest damage monitoring system. It may include more or fewer components than the example, or a combination of certain components, or different components. For example, the time-aware attention-based drone forest damage monitoring system may also include input and output devices, network access devices, buses, etc.

[0061] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the operation system of the time-aware attention-based drone forest damage monitoring system, and utilizes various interfaces and lines to connect the various parts of the entire time-aware attention-based drone forest damage monitoring system operation system.

[0062] The memory can be used to store the computer programs and / or modules. The processor implements the various functions of the time-aware attention-based drone forest damage monitoring system by running or executing the computer programs and / or modules stored in the memory and accessing the data stored in the memory. The memory may primarily include a program storage area and a data storage area. The program storage area may store an operating system and at least one application required for a function (such as sound playback or image playback); the data storage area may store data generated based on the use of the mobile phone (such as audio data and a phone book). Furthermore, the memory may include high-speed random access memory (RAM) and non-volatile memory, such as a hard disk, internal memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other volatile solid-state storage device.

[0063] Although the present invention has been described in considerable detail and with particularity with respect to several embodiments, it is not intended to limit the present invention to any of these details or embodiments or any particular embodiment, so as to effectively encompass the intended scope of the present invention. In addition, the present invention has been described above with respect to embodiments foreseen by the inventors for the purpose of providing a useful description, and those insubstantial modifications of the present invention that are not currently foreseen may still represent equivalent modifications of the present invention.

Claims

1. A method for monitoring forest damage using a drone based on time-aware attention, characterized in that: The method comprises the following steps: S100, UAV collects images to build a forest damage image sequence library; S200, using the forest damage image sequence library to build a pest perception model based on time-aware attention to obtain pest related vectors; S300, constructs a K-layer perceptron based on pest-related vectors; S400, extracting image descriptors from the K-layer perceptron and obtaining the pest monitoring degree; S500, collects pest images and marks pest areas through pest monitoring; S501, continuously obtaining the pest monitoring degree of the pest-infested area and calculating the number of forest damage occurrence trajectories; S502, judging the forest damage monitoring result according to the number of forest damage occurrence trajectories; The method for continuously obtaining the pest monitoring degree in the pest-infested area and calculating the number of forest damage occurrence trajectories is to obtain the oscillation boundary number through the horizontal comparison of the occurrence monitoring degree in the pest monitoring degree; mark the turning mark node through the oscillation boundary number, thereby setting the penalty weighted value, and calculating the change degree according to the occurrence monitoring degree.

2. The method for monitoring forest damage using a drone based on time-aware attention according to claim 1, characterized in that: In step S100, the method for collecting images by drones to construct a forest damage image sequence library is as follows: setting a flight route and a monitoring period according to the locations of known pests in the forest area, collecting images by drones at regular intervals every day to form an image sequence instance; and arranging the pest images at the same location in chronological order to form an image sequence instance. The forest damage image sequence library is composed of a number of pest image sequence instances.

3. The method for monitoring forest damage using a drone based on time-aware attention according to claim 1, characterized in that: In step S200, a method for constructing a pest perception model based on time-aware attention using a forest damage image sequence library to obtain pest-related vectors is as follows: after preprocessing the forest damage image sequence library, the image sequence is trained using a deep learning model including a time-aware attention mechanism to obtain a pest perception model, wherein the time-aware attention mechanism adopts any one of the TimeSformer operator, the ConvLSTM operator or the Temporal Attention Network operator; and the output result of the intermediate embedding layer of the pest perception model is extracted as a pest-related vector.

4. The method for monitoring forest damage using a drone based on time-aware attention according to claim 1, characterized in that: In step S300, the method for constructing a K-layer perceptron based on the pest-related vector is: using the pest-related vector as input to construct a feedforward neural network model containing at least K fully connected layers as a K-layer perceptron, wherein each hidden layer adopts a ReLU activation function, and the output layer output includes occurrence confidence and control confidence, which are used to represent the time position state of the input image in the pest development process.

5. The method for monitoring forest damage using a drone based on time-aware attention according to claim 1, characterized in that: In step S400, the method for extracting image descriptors from the K-layer perceptron and obtaining the pest monitoring degree is: the drone inputs the collected image into the K-layer perceptron to obtain the image descriptor and obtains the occurrence confidence and prevention confidence of real-time monitoring through nonlinear mapping processing, which are defined as the occurrence monitoring degree and the prevention monitoring degree respectively, and both are classified as pest monitoring degrees.

6. The method for monitoring forest damage using a drone based on time-aware attention according to claim 1, characterized in that: In step S500, the method of collecting pest images and marking pest areas through pest monitoring degree is: when the drone performs a forest monitoring task, the pest monitoring degree is calculated for each image obtained in real time. If an area is not marked as a pest area, and the occurrence monitoring degree of the pest monitoring degree of the image corresponding to the area exceeds 10%, the area is marked as a pest area, and the obtained image is stored and recorded as a pest image; otherwise, the image is not stored.

7. The method for monitoring forest damage using a drone based on time-aware attention according to claim 1, characterized in that: In step S501, the method for continuously obtaining the pest monitoring degree and calculating the number of forest pest occurrence trajectories in the pest-infested area is as follows: obtaining the occurrence monitoring degree and the control monitoring degree within the monitoring period, calculating the distance between the maximum and minimum values ​​of the occurrence monitoring degree and the median, and taking the maximum value as the oscillation limit number; within the monitoring period, the absolute value of the difference between the control monitoring degrees of any two consecutive natural days is recorded as the control step number, and the average of all control step numbers is the step average; If the increase in the occurrence monitoring degree of a natural day compared with the first natural day in the reverse time direction is greater than the oscillation limit number or the decrease in the prevention and control monitoring degree is greater than the step average, then the natural day is recorded as a turning mark node; set a time window pw, and calculate the difference in the occurrence monitoring degree of any natural day and pw natural days in the reverse time direction, and record it as the local change degree Lvdeg; The penalty weight value of the turning mark node is set to 1 and the others are set to 0; for the current natural day, the number of forest damage occurrence trajectories is calculated based on the local change degree and the penalty weight value.

8. The method for monitoring forest damage using a drone based on time-aware attention according to claim 1, characterized in that: In step S501, the method for continuously obtaining the pest monitoring degree and calculating the number of forest pest occurrence trajectories in the pest-infested area is as follows: on any natural day within the monitoring period, the ratio of the control monitoring degree to the occurrence monitoring degree is recorded as the pest control efficiency index; Calculate the product of the remaining control monitoring degree and the occurrence monitoring degree as the relative occurrence cumulative amount; Set the time window pw and calculate the dynamic cumulative occurrence risk for any natural day based on the pest control efficiency index and the relative occurrence cumulative amount; Draw a line graph of the dynamic cumulative risk of occurrence for any time window, record the natural day corresponding to the point where the slope changes from negative to positive as the defect mark point, and the natural day corresponding to the point where the slope changes from positive to negative as the positive mark point; record the difference between the defect mark and its slope on the previous day as the positive mark quantity, and record the difference between the positive mark point and its slope on the previous day as the defect mark quantity; calculate the slope sign weight for any natural day and calculate the number of forest damage occurrence trajectories in combination with the dynamic cumulative risk of occurrence.

9. The method for monitoring forest damage using a drone based on time-aware attention according to claim 1, characterized in that: In step S502, the method for judging the forest damage monitoring result based on the number of forest damage occurrence trajectories is: when the number of forest damage occurrence trajectories reaches a minimum value and the number of forest damage occurrence trajectories is higher than the minimum value within at least 15 natural days thereafter, the forest damage monitoring result of the area is judged to be forest damage elimination, and the insect pest area mark of the area is cancelled.

10. A UAV forest damage monitoring system based on time-aware attention, characterized in that: The time-aware attention-based drone forest damage monitoring system includes: a processor, a memory, and a computer program stored in the memory and runnable on the processor. When the processor executes the computer program, the steps of the time-aware attention-based drone forest damage monitoring method described in any one of claims 1 to 9 are implemented. The time-aware attention-based drone forest damage monitoring system runs on computing devices such as desktop computers, laptop computers, PDAs, and cloud data centers.

Images