Locust disaster prediction method and device, electronic equipment and storage medium

Abstract

The application provides a locust disaster prediction method and device, electronic equipment and storage medium, and relates to the technical field of intelligent agriculture. The application introduces historical locust disaster occurrence records into environmental data, so that a locust disaster prediction model can fully combine the characteristics of environmental data and historical locust disaster occurrence records to predict the locust disaster occurrence probability of each position point in a region, and the prediction accuracy of the locust disaster prediction model for the locust disaster occurrence probability under complex environmental data is improved. The application can explicitly consider the lagging effect of environmental data and historical locust disaster occurrence records by extracting the time sequence characteristics of a target feature image, and further generate a locust disaster occurrence probability that is more in line with the biological occurrence mechanism of locust disasters, thereby improving the prediction accuracy of the locust disaster occurrence probability.

Description

Technical Field

[0001] This invention relates to the field of intelligent agriculture technology, and in particular to a method, device, electronic device, and storage medium for predicting locust plagues. Background Technology

[0002] The desert locust is a major global agricultural pest characterized by its strong migratory ability, rapid reproduction, and wide diet. Due to its vast geographical reach and complex migration routes, traditional monitoring methods struggle to cover its entire life cycle and activity range. Therefore, accurately predicting and issuing early warnings about the occurrence and development of desert locusts using modern technologies to guide precise and efficient control efforts has become a critical technical challenge that urgently needs to be addressed in the global field of agricultural pest control.

[0003] To address these technological needs, existing technologies primarily rely on environmental factors and historical desert locust occurrence data to construct predictive models. These models fall into three categories: first, statistical analysis models, such as generalized additive models incorporating nonlinear functions and neural network models with high classification accuracy; second, niche models, such as maximum entropy models that still demonstrate high accuracy even with missing species distribution data; and third, traditional machine learning models, such as random forest models suitable for high-dimensional data and exhibiting stable predictive performance, as well as support vector machine models capable of handling nonlinear problems. Furthermore, some research has begun exploring integrated prediction methods, combining multiple models for density and trend prediction.

[0004] Existing prediction models have low accuracy in predicting locust plagues. Traditional machine learning methods are insufficient for building large-scale, uniform prediction models; their prediction accuracy is usually only acceptable in small, localized areas. When applied to vast locust-affected areas, models often need to be trained separately for different sub-regions, which easily leads to fragmented prediction results. This not only increases the complexity of modeling but also significantly increases the risk of missed reports, posing a potentially huge threat to agricultural production. Summary of the Invention

[0005] This invention provides a method, apparatus, electronic device, and storage medium for predicting locust plagues, thereby addressing the shortcomings of low prediction accuracy in existing technologies and improving the accuracy of locust plague prediction.

[0006] This invention provides a method for predicting locust plagues, comprising:

[0007] Based on environmental data of the predicted area within a set time period, and historical locust plague records, a target dataset is obtained; historical locust plague records include the time and number of historical locust plagues.

[0008] Generate target feature images of the predicted region based on the target dataset;

[0009] Input the target feature image into the locust disaster prediction model to obtain the probability of locust disaster occurrence in the predicted area output by the locust disaster prediction model;

[0010] The locust disaster prediction model extracts the temporal features, spatial distribution features of locusts, and contextual features of the target feature image; based on the temporal features, spatial distribution features of locusts, and contextual features, it determines the probability of locust disasters occurring; the locust disaster prediction model is trained on the basis of a preset model and based on sample feature images.

[0011] According to the locust plague prediction method provided by the present invention, the locust plague prediction model is trained in the following manner:

[0012] Obtain the first sampling point where the locust plague actually occurred and its sample locust plague occurrence record; the sample locust plague occurrence record includes the time and number of locust plagues occurring in the sample;

[0013] Based on the first sampling point, a second sampling point is generated where the locust plague did not actually occur.

[0014] Based on the first sampling point where the locust plague actually occurred and the second sampling point where the locust plague did not actually occur, we obtain the target sampling point carrying the label of whether the locust plague actually occurred and the locust plague occurrence record of the sample.

[0015] Acquire comprehensive sample environment images of the target sampling point within each time period; the comprehensive sample environment image within a time period is determined based on multiple sample environment data of the target sampling point within the time period.

[0016] For each target sampling point, the time period and the time of locust plague occurrence are aligned to merge the comprehensive sample environment image and the number of locust plague occurrences to obtain the merged sample environment image for each time period;

[0017] The merged sample environment images of at least one continuous time period are stitched together, and the resulting sample feature image is used as a training sample.

[0018] A pre-defined model is constructed based on the initial temporal convolutional attention module and the initial U-Net network.

[0019] A locust plague prediction model is obtained by training a preset model based on training samples.

[0020] According to the locust plague prediction method provided by the present invention, the locust plague occurrence records of the first sampling point and its samples are obtained based on the following method:

[0021] Obtain a set of recorded events of locust plagues that actually occurred within at least one historical time period. The set of recorded events includes multiple recorded events of locust plagues that actually occurred, and each recorded event carries the location coordinates and time points of the actual locust plagues.

[0022] Merge all recorded event sets and remove duplicate recorded events with the same location coordinates and time points to obtain a cleaned recorded event set;

[0023] Based on the cleaned event record set, obtain each target event record. Each target event record carries the target location coordinates and target time point of the locust plague.

[0024] By aggregating and statistically analyzing the events recorded for each target within the grid of location coordinates, the occurrence time and frequency of locust plagues within the grid of location coordinates can be obtained.

[0025] Use the grid at each location coordinate as the first sampling point;

[0026] Based on the location coordinates of the raster, the occurrence time and frequency of locust plagues are used to obtain sample locust plague occurrence records.

[0027] According to the locust plague prediction method provided by the present invention, the target dataset includes at least one type of environmental data from N consecutive time periods, and a target feature image of the prediction area is generated based on the target dataset, including:

[0028] Global standardization processing is performed on various environmental data for different time periods to obtain standardized environmental data for each time period.

[0029] The standardized environmental data from various time periods are converted into various environmental images from various time periods.

[0030] Based on historical locust plague records, locust plague density images of the predicted area are generated for each time period.

[0031] By overlaying locust plague density images and various environmental images at different time periods, a comprehensive feature image for each time period is obtained.

[0032] By stitching together the comprehensive feature images from N time periods, the target feature image is obtained.

[0033] According to the locust plague prediction method provided by the present invention, the locust plague prediction model includes a temporal convolutional attention module and a U-Net network, and the locust plague prediction model is used for:

[0034] Temporal features of the target feature image are extracted based on the temporal convolutional attention module to obtain a two-dimensional feature map containing temporal features;

[0035] Based on the U-Net network, spatial distribution features and contextual features of locusts are extracted from two-dimensional feature maps containing temporal features;

[0036] By fusing and activating temporal features, locust spatial distribution features, and contextual features, the probability of locust disasters can be obtained.

[0037] According to the locust plague prediction method provided by the present invention, the set time period includes N consecutive time periods, where N≥1, and the temporal convolutional attention module is used for:

[0038] Perform at least one 3D convolution on the target feature image to obtain the data features and their time weights for each time period;

[0039] The time series features are obtained by weighting and summing the data features and their time weights for each time period.

[0040] According to the locust plague prediction method provided by the present invention, the environmental data includes static environmental data and dynamic environmental data;

[0041] Dynamic environmental data includes at least one of the following: surface temperature, soil moisture content, surface wind speed, vegetation area, soil temperature, and precipitation.

[0042] Static environmental data includes at least one of the following: stratum elevation data, soil composition, and land use type.

[0043] The present invention also provides a locust plague prediction device, comprising:

[0044] The data acquisition module is used to obtain the target dataset based on environmental data of the predicted area within a set time period and historical locust plague occurrence records; the historical locust plague occurrence records include the time and number of historical locust plagues.

[0045] The preprocessing module is used to generate target feature images of the predicted region based on the target dataset;

[0046] The prediction module is used to input the target feature image into the locust disaster prediction model and obtain the probability of locust disaster occurrence in the predicted area output by the locust disaster prediction model. The locust disaster prediction model extracts the temporal features, locust spatial distribution features, and contextual features of the target feature image. Based on the temporal features, locust spatial distribution features, and contextual features, the probability of locust disaster occurrence is determined. The locust disaster prediction model is trained on the sample feature image based on the preset model.

[0047] The present invention also provides an electronic device, including 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 implement any of the locust plague prediction methods described above.

[0048] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the locust plague prediction methods described above.

[0049] The present invention provides a method, apparatus, electronic device, and storage medium for predicting locust plagues. Based on environmental data of the prediction area within a set time period and historical locust plague occurrence records, a target dataset is obtained. The historical locust plague occurrence records include the time and frequency of historical locust plague occurrences. A target feature image of the prediction area is generated based on the target dataset. The target feature image is input into a locust plague prediction model to obtain the probability of locust plague occurrence in the prediction area output by the model. The locust plague prediction model extracts temporal features, locust spatial distribution features, and contextual features from the target feature image. Based on these features, the probability of locust plague occurrence is determined. The locust plague prediction model is trained on a pre-defined model using sample feature images. This invention incorporates historical locust plague occurrence records into the environmental data, enabling the locust plague prediction model to fully combine the features of the environmental data with historical locust plague occurrence records to predict the probability of locust plague occurrence at various locations within the area, thus improving the accuracy of the locust plague prediction model in predicting the probability of locust plague occurrence under complex environmental data. This invention extracts the temporal features of target feature images, which can explicitly consider the lag effect of environmental data and historical locust disaster records, thereby generating a locust disaster probability that is more consistent with the biological mechanism of locust disasters and improving the prediction accuracy of locust disaster probability. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0051] Figure 1 This is one of the flowcharts illustrating the locust plague prediction method provided by the present invention.

[0052] Figure 2 This is the second flowchart illustrating the locust plague prediction method provided by this invention.

[0053] Figure 3 This is a schematic diagram of the process for obtaining a two-dimensional feature map containing temporal features provided by the present invention.

[0054] Figure 4 This is a schematic diagram of the locust plague prediction device provided by the present invention.

[0055] Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0057] The application areas of this invention include the occurrence and risk prediction of desert locust disasters. For ease of understanding, the relevant terminology of the locust disaster prediction model that may be involved in this invention will be introduced below.

[0058] The U-Net network employs a convolutional neural network architecture. The core idea of ​​U-Net lies in its clever combination of contextual and positional information to achieve accurate classification of each pixel in an image. The U-Net network mainly consists of three parts: contraction paths, expansion paths, and skip connections.

[0059] The shrinking path, also known as the encoder, is used to capture contextual information from an image. It consists of a series of convolutional and max-pooling layers. After the image is input, it undergoes repeated convolution-pooling operations. With each pooling operation, the size of the feature map is halved, but the number of channels increases. By continuously "shrinking" the image, the shrinking path extracts higher-level, more abstract semantic features to obtain contextual information.

[0060] The dilation path, also known as the decoder, is responsible for achieving precise localization and gradually restoring image details and resolution. It is symmetrical to the contraction path. The dilation path uses upsampling or transposed convolution operations to progressively enlarge the highly abstract feature map extracted from the bottom of the contraction path, restoring it to the original image size.

[0061] The role of skip connections is to reintroduce the location information lost by the encoder into the decoder. After each downsampling stage in the contraction path, the U-Net network directly copies its output feature map and concatenates it to the feature map of the corresponding upsampling stage in the expansion path. When the decoder recovers the image size through upsampling, it not only possesses the abstract semantic information from the bottom of the network but also directly obtains high-resolution, spatially detailed location information from the shallow layers of the encoder through skip connections. This combination of "high-level semantics" and "shallow details" enables the network to ultimately output very accurate segmentation results.

[0062] In recent years, with the development of remote sensing technology, acquiring large-scale, continuous temporal environmental and ecological observation data has become more convenient, providing important data support for pest and disease prediction. However, limited by the capabilities of existing prediction models, these multi-source remote sensing and meteorological data have not been fully and effectively utilized, making it difficult to comprehensively reflect the occurrence and migration patterns of desert locusts. To overcome these shortcomings, this invention proposes a desert locust migration simulation and risk early warning method based on spatiotemporal deep networks. This method fully integrates multi-source remote sensing data, meteorological data, and geographic information data, using a deep learning model to predict the occurrence of desert locust disasters, thereby achieving large-scale, refined risk assessment and early warning. This invention can effectively improve the timeliness and spatial generalization ability of locust disaster prediction, providing reliable technical support for green pest control and scientific decision-making.

[0063] The following is combined with Figures 1 to 5 The present invention describes a method, apparatus, and electronic device for predicting locust plagues.

[0064] Figure 1 This is one of the flowcharts illustrating the locust plague prediction method provided by the present invention. Figure 2 This is the second flowchart illustrating the locust plague prediction method provided by this invention. For example... Figure 1 and Figure 2 As shown, the method for predicting locust plagues includes steps S100 to S300, and the specific steps are as follows.

[0065] S100: Obtain the target dataset based on environmental data of the predicted area within a set time period and historical locust plague records.

[0066] Historical locust plague records include the time and number of historical locust plagues.

[0067] The prediction region is the area where it is necessary to predict whether a locust plague (e.g., desert locust) will occur. For example, the prediction region could be global, a country, a province, or a farm. The prediction region consists of multiple raster grids with location coordinates (raster grids divided according to location coordinates). Each raster grid with a location coordinate serves as a data collection point. Environmental data for each raster grid within the prediction region over a set time period, along with historical locust plague occurrence records, are acquired to obtain the target dataset.

[0068] Environmental data refers to data on environmental factors that influence locust plagues. When an environmental factor affects one or more growth stages of locusts, that factor is considered environmental data. Environmental data includes at least one type of dynamic environmental data and static environmental data. For example, there are 17 types of environmental data.

[0069] Historical records of locust plagues include records of whether locust plagues actually occurred in the predicted area within a specified time period.

[0070] The target dataset is obtained by collecting environmental data of the predicted area within a specified time period, as well as the actual occurrence time and frequency of historical locust plagues within that time period. The specified time period includes at least one continuous time cycle. For example, a one-month time cycle can be used, and the specified time period includes three consecutive months. The target dataset is obtained based on the environmental data of the predicted area over the past three months, as well as the actual occurrence time and frequency of historical locust plagues within the past three months.

[0071] S200: Generate the target feature image of the predicted region based on the target dataset.

[0072] Before inputting the target dataset into the locust plague prediction model, the dataset needs to be preprocessed, including standardization and image processing. Standardization includes global standardization. Image processing involves converting the target dataset into image data.

[0073] Target feature images of the prediction area are generated based on the target dataset. All environmental data and historical locust plague occurrence records are integrated into a single target feature image. Multiple (e.g., 16 types) of environmental images and locust plague occurrence density images for each month are generated from the target dataset, including all data collection points. Optionally, the 16 types of environmental images and locust plague occurrence density images for each month are overlaid to obtain a comprehensive feature image for each month. The comprehensive feature images of three consecutive months are then stitched together to obtain the target feature image.

[0074] S300: Input the target feature image into the locust disaster prediction model to obtain the probability of locust disaster occurrence in the predicted area output by the locust disaster prediction model.

[0075] The locust disaster prediction model extracts the temporal features, spatial distribution features of locusts, and contextual features of the target feature image; based on the temporal features, spatial distribution features of locusts, and contextual features, it determines the probability of locust disasters occurring; the locust disaster prediction model is trained on the basis of a preset model and based on sample feature images.

[0076] A pre-defined model is constructed based on the initial temporal convolutional attention module and the initial U-Net network. Sample feature images are acquired, and the pre-defined model is trained based on these images to obtain a locust plague prediction model.

[0077] The target feature image is input into the locust plague prediction model. The locust plague prediction model includes a temporal convolutional attention module, a U-Net network, and an output module. The temporal convolutional attention module extracts features from the target feature image to obtain temporal features. The extracted two-dimensional feature map containing temporal features is then input into the U-Net network. The U-Net network further extracts features from the two-dimensional feature map containing temporal features to obtain the spatial distribution features of locusts and the contextual features of the locust spatial distribution. Based on the temporal features, the spatial distribution features of locusts, and the contextual features, the output module generates a logits vector of length 2 for each pixel of the target feature image (corresponding to the prediction of locust plague occurrence versus non-occurrence). The logits vector is transformed using an activation function (e.g., softmax) to obtain an occurrence probability map. The occurrence probability map includes the probability of locust plague occurrence for each pixel in the prediction region.

[0078] Furthermore, it displays the probability of locust plagues occurring at each pixel.

[0079] The locust plague prediction method provided in this invention obtains a target dataset based on environmental data of the prediction area within a set time period and historical locust plague occurrence records. The historical locust plague occurrence records include the time and frequency of historical locust plague occurrences. A target feature image of the prediction area is generated based on the target dataset. The target feature image is input into a locust plague prediction model to obtain the probability of locust plague occurrence in the prediction area output by the model. The locust plague prediction model extracts temporal features, locust spatial distribution features, and contextual features from the target feature image. Based on these features, the probability of locust plague occurrence is determined. The locust plague prediction model is trained on a pre-defined model using sample feature images. This invention introduces historical locust plague occurrence records into the environmental data, enabling the locust plague prediction model to fully combine the features of the environmental data with historical locust plague occurrence records to predict the probability of locust plague occurrence at various locations within the area, thus improving the accuracy of the locust plague prediction model in predicting the probability of locust plague occurrence under complex environmental data. This invention extracts the temporal features of target feature images, which can explicitly consider the lag effect of environmental data and historical locust disaster records, thereby generating a locust disaster occurrence probability that is more consistent with the biological mechanism of locust disasters and improving the accuracy of locust disaster occurrence probability.

[0080] Based on the above embodiments, environmental data includes static environmental data and dynamic environmental data;

[0081] Dynamic environmental data includes at least one of the following: surface temperature, soil moisture content, surface wind speed, vegetation area, soil temperature, and precipitation.

[0082] Static environmental data includes at least one of the following: stratum elevation data, soil composition, and land use type.

[0083] Dynamic environmental data consists of multiple dynamic environmental factors. These factors include at least one of the following: surface temperature, soil moisture content, surface wind speed, vegetation area, soil temperature, and precipitation. These dynamic environmental factors change rapidly over time, typically fluctuating on monthly, ten-day, or daily scales, and directly impact locust reproduction, development, and migration. For example, post-precipitation vegetation growth provides food for desert locusts. Monthly aggregation (MONTHLY AGGR) was performed on remote sensing images, including the land surface reanalysis dataset (ERA5_LAND) released by weather forecasting platforms, to obtain the various dynamic environmental factors.

[0084] Surface temperature includes average surface temperature and maximum surface temperature. Soil moisture content includes the volumetric moisture content of the first and second soil layers. Surface wind speed includes the eastward and northward components of the 10-meter wind. Vegetation area includes high and low leaf area index. Soil temperature includes the first and second soil layers. Precipitation includes the sum of total precipitation. The dynamic environmental factors are shown in Table 1.

[0085] Table 1

[0086]

[0087]

[0088] Static environmental data consists of multiple static environmental factors. These factors include at least one of the following: stratigraphic elevation data, soil composition, and land use type. Stratigraphic elevation data can be derived from a Digital Elevation Model (DEM). Soil composition can be derived from the Soil Grid. Soil type can be derived from remote sensing imagery such as medium-resolution imaging spectrometers of land cover types.

[0089] Soil composition includes the average content of coarse soil debris (cfvo_5-15cm_mean), sand content (sand_5-15cm_mean), clay content (clay_5-15cm_mean), and silt content (silt_5-15cm_mean) at a depth of 5-15cm below the surface.

[0090] This invention selects land use types that are prone to locust plagues. Various land use types that are prone to locust plagues are numbered, with each number representing a different land use type. The land use types of this invention include: 6 Dense shrubland, 7 Sparse shrubland, 8 Woody savanna, 9 Savanna, 10 Grassland, 12 Farmland, 14 Mixed farmland and vegetation, 16 Desert, wasteland, etc.

[0091] The environmental data of this invention fully incorporates multi-source environmental factors such as surface temperature, soil moisture content, surface wind speed, vegetation area, soil temperature, precipitation, stratum elevation data, soil composition, and land use type, comprehensively collecting environmental data that affect locust plagues, which is conducive to improving the accuracy of subsequent predictions of locust plague probability.

[0092] Based on the above embodiments, the locust plague prediction model was trained in the following manner:

[0093] Obtain the first sampling point where the locust plague actually occurred and its sample locust plague occurrence record; the sample locust plague occurrence record includes the time and number of locust plagues occurring in the sample;

[0094] Based on the first sampling point, a second sampling point is generated where the locust plague did not actually occur.

[0095] Based on the first sampling point where the locust plague actually occurred and the second sampling point where the locust plague did not actually occur, we obtain the target sampling point carrying the label of whether the locust plague actually occurred and the locust plague occurrence record of the sample.

[0096] Acquire comprehensive sample environment images of the target sampling point within each time period; the comprehensive sample environment image within a time period is determined based on multiple sample environment data of the target sampling point within the time period.

[0097] For each target sampling point, the time period and the time of locust plague occurrence are aligned to merge the comprehensive sample environment image and the number of locust plague occurrences to obtain the merged sample environment image for each time period;

[0098] The merged sample environment images of at least one continuous time period are stitched together, and the resulting sample feature image is used as a training sample.

[0099] A pre-defined model is constructed based on the initial temporal convolutional attention module and the initial U-Net network.

[0100] A locust plague prediction model is obtained by training a preset model based on training samples.

[0101] Data from all desert locust ground survey sites from January 1985 to June 2025 were collected. The study area was defined as 26.04000°W–85.35000°E, 4.27417°S–37.38333°N, extending 10 km outward from the boundary of each survey site. Based on the collected data from all desert locust ground survey sites, at least one ground point dataset showing a desert locust infestation was obtained. All ground point datasets showing desert locust infestations were statistically analyzed to obtain the first sampling point where the locust infestation actually occurred and its sample locust infestation records. The spatial resolution of the first sampling point and its sample locust infestation records can be 10 km, and the temporal resolution can be monthly. Analysis of comprehensive ecological and historical distribution data reveals that the core occurrence area of ​​the desert locust plague (the core distribution point of the first sampling site) is mainly located in tropical to subtropical arid and semi-arid deserts and savannas. These areas are mostly open and flat, with sandy soil, characterized by a climate of year-round warmth (e.g., average annual temperature above 20°C) and concentrated, short-lived rainfall (e.g., annual precipitation of approximately 50–300 mm, with a distinct rainy season). Vegetation is dominated by drought-resistant grasses with medium to low coverage. Typical distribution areas include North Africa and the Sahel region, the Arabian Peninsula, West Asia (and the southern edge of Central Asia).

[0102] Based on the first sampling points, second sampling points are generated where the locust plague did not actually occur. These second sampling points are also called pseudo-absence (PA) points. The collection of first sampling points for desert locusts exhibits spatial bias, favoring suitable and susceptible areas. Therefore, this invention uses random generation with exclusion buffer (RGEB) to generate pseudo-absence points, representing areas rarely visited by desert locusts over those years, thus enhancing the representativeness of PA points. Optionally, multiple (e.g., 30,604) second sampling points are randomly generated outside a 20km buffer zone of each first sampling point. These second sampling points are randomly allocated to each month in a 1:1 ratio. For example, if there are 1,000 first sampling points in January 2000, then the number of second sampling points allocated to January 2000 is also 1,000, ensuring a 1:1 ratio between the number of first and second sampling points in January 2000.

[0103] Based on the first sampling point where a locust plague actually occurred and the second sampling point where a locust plague did not actually occur, a label carrying whether a locust plague actually occurred is obtained, along with the target sampling point for the locust plague occurrence record. Optionally, the label for whether a locust plague actually occurred includes 1 and 0, where 1 indicates that a locust plague occurred and 0 indicates that a locust plague did not occur.

[0104] Acquire comprehensive sample environment images of target sampling points within various time periods. The time period includes one month, one year, etc. This embodiment of the invention uses a one-month time period as an example. Download multiple sample environment data for all target sampling points in each month. The spatial resolution of the sample environment data is the same as that of the first sampling point, both being 10km. Perform global standardization on the multiple sample environment data for each month, and save the standard deviation and mean to prepare for subsequent prediction using future environmental factors. Convert the globally standardized multiple sample environment data for each month into multiple sample environment images for each month. For example, the multiple sample environment images include 11 (channel) dynamic sample environment images and 5 static sample environment images. The format of the sample environment images includes Tag Image File Format (TIFF). Overlay the multiple sample environment images according to the time period (month) to obtain a comprehensive sample environment image for each time period. For example, overlaying 11 dynamic sample environment images and 5 static sample environment images from January 2000 yields a 16-channel comprehensive sample environment image for January 2000.

[0105] For each target sampling point, the time period and the locust plague occurrence time are aligned to merge the comprehensive sample environment image and the number of locust plague occurrences, obtaining merged sample environment images (including 17 channels) for each time period. For example, the number of locust plague occurrences for each target sampling point in January 2000 is obtained. The number of locust plague occurrences for each target sampling point in January 2000 is merged into the 16-channel comprehensive sample environment image for January 2000, resulting in a 17-channel merged sample environment image for January 2000. Following the above steps, merged sample environment images for each time period (month) are obtained.

[0106] Considering that the occurrence of desert locust plagues is influenced by environmental data from at least one continuous time period (e.g., three consecutive months), the merged sample environmental data from at least one continuous time period are stitched together to obtain sample feature images. For example, merged sample environmental images from three consecutive months are stitched together to obtain sample feature images. Based on all sample feature images, training samples are obtained. The training samples are then divided into training and test sets according to a preset ratio (e.g., 8:2).

[0107] Furthermore, the study obtains typical time periods of locust plague occurrences, such as the time periods of full-scale locust outbreaks, the time periods of locust plagues spiraling out of control, the time periods of rapid locust population growth, and the time periods with potential locust plague outbreak risks. Training samples within the typical locust plague occurrence time periods are used as the training set, while time periods of atypical locust plague occurrences are used as the test set. For example, this invention fully covers the temporal, spatial, and geographical distribution of all target sampling points from January 1985 to June 2025 when dividing the dataset, ensuring that the pre-set model can learn the long-term occurrence characteristics of desert locusts in typical distribution areas. The research period totals 492 months. Historically recorded periods of full-scale locust outbreaks and periods of uncontrolled locust outbreaks include 1986–1989 and 2019–2022. Rising / warning periods (e.g., periods of rapid locust population growth and periods with potential locust outbreak risk) include 1992–1994, 1994–1996, 1996–1998, 2003–2005, and 2006–2018. All data from January to December covered in these periods are included in the training set, totaling 31 years (372 months). Data from January to June for the remaining years (1985, 1990, 1991, 1999, 2000, 2001, 2002, 2023, 2024, and 2025) are used as the test set, totaling 10 years (114 months). This division ensures that the training set contains complete temporal information on multiple key outbreak and warning phases, and also provides independent test samples for the generalization ability of the preset model under different years and climate backgrounds.

[0108] A pre-defined model is constructed based on the initial temporal convolutional attention module and the initial U-Net network. The pre-defined model is trained using the training set and tested using the test set to obtain a locust plague prediction model.

[0109] To obtain evaluation metrics, a confusion matrix (f1) was used to statistically analyze the predicted locust plague probability output by the pre-defined model (referred to as the prediction probability). A threshold was then set to convert the prediction probability into a binary category. After model training, the final model weight parameters were obtained, and predictions were made on the test set and throughout the entire study area to evaluate the generalization performance of the pre-defined model. The prediction effect was also verified using known ground survey points (target sampling points).

[0110] The trained model weight parameters were tested on a test set, and the results were obtained. The test results include an accuracy of 0.8734, a precision of 0.9753, a recall of 0.7528, and an F1 score of 0.8497. The test results demonstrate that the locust plague prediction model obtained in this invention significantly outperforms traditional models in terms of test metrics.

[0111] Furthermore, all years and months in the test set were visualized, focusing not only on the predicted probability of the sample area, but also on the explicit desert locust ground survey points (first sampling points) and pseudo-absent points (second sampling points), to compare their rationality.

[0112] This invention generates a second sampling point based on a first sampling point, indicating that a locust plague has not actually occurred. This allows the pre-set model to learn the environmental characteristics when a locust plague has not actually occurred, enabling the model to accurately distinguish between environmental characteristics when a locust plague has occurred and when it has not, thus predicting a locust plague probability that is more consistent with ecological logic. Based on the comprehensive sample environmental images, this invention incorporates the number of locust plague occurrences, overcoming the information loss and limitations of traditional binary labels (occurred / not occurred), and improving the accuracy of the pre-set model in predicting the probability of locust plague occurrence under complex ecological conditions. Based on multiple sample environmental data, this invention determines comprehensive sample environmental data, fully considering multi-source sample environmental data affecting locust plague occurrence, which helps improve the accuracy of subsequent prediction model outputs predicting the probability of locust plague occurrence. This invention stitches together merged sample environmental images from at least one continuous time period to obtain sample feature images, considering the occurrence of locust plagues and the influence of environmental factors affecting locust growth and development within at least one continuous time period, which helps obtain a locust plague occurrence probability that is more consistent with the biological mechanism of locust plague occurrence.

[0113] Based on the above embodiments, the locust plague occurrence records of the first sampling point and its samples were obtained in the following manner:

[0114] Obtain a set of recorded events of locust plagues that actually occurred within at least one historical time period. The set of recorded events includes multiple recorded events of locust plagues that actually occurred, and each recorded event carries the location coordinates and time points of the actual locust plagues.

[0115] Merge all recorded event sets and remove duplicate recorded events with the same location coordinates and time points to obtain a cleaned recorded event set;

[0116] Based on the cleaned event record set, obtain each target event record. Each target event record carries the target location coordinates and target time point of the locust plague.

[0117] By aggregating and statistically analyzing the events recorded for each target within the grid of location coordinates, the occurrence time and frequency of locust plagues within the grid of location coordinates can be obtained.

[0118] Use the grid at each location coordinate as the first sampling point;

[0119] Based on the location coordinates of the raster, the occurrence time and frequency of locust plagues are used to obtain sample locust plague occurrence records.

[0120] Obtain a set of recorded events from at least one historical time period. For example, the set of recorded events includes the locust plague record set from 1985 to 2021 (a total of 50,000 records, with 36,853 records remaining after deleting the ControlREV class) and the locust plague record set from 2001 to 2025 (a total of 50,000 records, with 13,101 records remaining after deleting the No Locust class).

[0121] All recorded event sets are merged, and duplicate events with the same location coordinates and time points are removed to obtain a cleaned set of recorded events. For example, merging the recorded event sets of locust plagues from 1985-2021 and 2001-2025 yields the recorded event set of locust plagues from 1985-2025 (the merged set), which includes a total of 49,952 recorded events (occurrence point data). The merged set undergoes a first cleaning process, removing duplicate events with the same location coordinates (e.g., identical latitude and longitude) and time points (identical collection dates), resulting in 46,140 cleaned recorded events. This first cleaning essentially merges identical recorded events.

[0122] Based on the cleaned event log set, obtain the individual target event logs. Each target event log carries the actual location coordinates and time of the locust plague.

[0123] A second cleaning process is performed on all target recorded events. A grid of location coordinates is constructed, and all target recorded events are assigned to the corresponding grid. The target recorded events within each grid are aggregated and statistically analyzed to obtain the occurrence time and frequency of locust plagues within that grid. For example, if grid 1 contains three target recorded events: January 2000, February 2000, and March 2010, then the locust plague occurrence times within grid 1 are January 2000, February 2000, and March 2010, and the locust plague frequency within grid 1 is three.

[0124] Each location coordinate grid is used as the first sampling point. Based on the time and frequency of locust infestation occurrences within the grid at each location coordinate, sample locust infestation occurrence records are obtained.

[0125] This invention removes duplicate recorded events with the same location coordinates and time points to obtain a cleaned set of recorded events, eliminating redundant and repetitive events and ensuring the simplicity of the target recorded events. Furthermore, by aggregating and statistically analyzing each target recorded event within a grid of location coordinates, this invention obtains the occurrence time and frequency of locust plagues within that grid. It creatively accumulates and statistically analyzes duplicate target recorded events within the same grid of location coordinates, extracting locust plague occurrence time and frequency features that reflect historical locust plague density. This allows the pre-set model to learn the non-linear relationship between the probability of locust plague occurrence and historical locust plague density, thus overcoming the information loss and limitations of traditional binary labels (occurred / not occurred), and improving the accuracy of the pre-set model in predicting the probability of locust plague occurrence under complex ecological conditions.

[0126] Based on the above embodiments, the target dataset includes at least one type of environmental data from N consecutive time periods. Generating a target feature image of the prediction region based on the target dataset includes the following steps:

[0127] Global standardization processing is performed on various environmental data for different time periods to obtain standardized environmental data for each time period.

[0128] The standardized environmental data from various time periods are converted into various environmental images from various time periods.

[0129] Based on historical locust plague records, locust plague density images of the predicted area are generated for each time period.

[0130] By overlaying locust plague density images and various environmental images at different time periods, a comprehensive feature image for each time period is obtained.

[0131] By stitching together the comprehensive feature images from N time periods, the target feature image is obtained.

[0132] The process of generating target feature images of the prediction region based on the target dataset during model application is the same as the process of generating sample feature maps during model training.

[0133] The defined time period includes N consecutive time cycles. For example, the defined time period includes three consecutive months. The target dataset includes environmental data of the prediction area over three consecutive months, as well as historical records of locust plagues.

[0134] The target dataset includes at least one type of environmental data spanning N consecutive time periods (e.g., three consecutive months). Environmental data can be categorized into static and dynamic environmental data based on major environmental types. Dynamic environmental data includes surface temperature, soil moisture content, surface wind speed, vegetation area, soil temperature, and precipitation. Static environmental data includes geological elevation data, soil composition, and land use type.

[0135] Specifically, environmental data can be categorized into 17 environmental subcategories based on environmental type, including surface temperature, maximum surface temperature, first-layer soil volumetric water content, second-layer soil volumetric water content, eastward component of 10-meter wind, northward component of 10-meter wind, high vegetation leaf area index, low vegetation leaf area index, first-layer soil temperature, second-layer soil temperature, total precipitation, stratum elevation data, average content of coarse soil debris at a depth of 5-15cm below the surface, average content of soil sand at a depth of 5-15cm below the surface, average content of soil clay at a depth of 5-15cm below the surface, average content of soil silt at a depth of 5-15cm below the surface, and land use type.

[0136] Land use type is used separately. Preprocessing is performed on 16 other types of environmental data. Global standardization is applied to the 16 types of environmental data over three consecutive months to obtain standardized environmental data for each time period (month). The calculation formulas for the standardized environmental data are as follows.

[0137] ;

[0138] in, This is a standardized environmental data. As a type of environmental data, This is the mean of environmental data. This represents the standard deviation of environmental data.

[0139] The standardized environmental data from various time periods are converted into various environmental images for those time periods. Furthermore, the 16 standardized environmental data sets from each month are converted into 16 different environmental images for that month.

[0140] Based on historical locust plague records, locust plague density images of the predicted area are generated for each time period (e.g., month).

[0141] By overlaying locust plague density images and various environmental images (16 types of environmental images) at different time periods, a comprehensive feature image (including 17 channels) for each time period is obtained.

[0142] By stitching together the comprehensive feature images from N time periods, the target feature image is obtained. By stitching together the comprehensive feature images from three consecutive months, the target feature image (including 17×3=51 channels) is obtained.

[0143] The target dataset of this invention comprehensively considers multi-source environmental data influencing locust plagues, as well as historical locust plague records that significantly impact their occurrence. This comprehensive collection of factors influencing locust plague occurrence improves the accuracy of subsequent locust plague prediction models in outputting locust plague probability. This invention stitches together comprehensive feature images from N time periods to obtain the target feature image. It considers that locust plague occurrence is influenced by environmental factors affecting locust growth and development over at least one continuous time period, thus facilitating the acquisition of locust plague probability figures that better reflect the biological mechanisms of locust plagues.

[0144] Based on the above embodiments, the locust plague prediction model includes a temporal convolutional attention module and a U-Net network. The locust plague prediction model is used for:

[0145] Temporal features of the target feature image are extracted based on the temporal convolutional attention module to obtain a two-dimensional feature map containing temporal features;

[0146] Based on the U-Net network, spatial distribution features and contextual features of locusts are extracted from two-dimensional feature maps containing temporal features;

[0147] By fusing and activating temporal features, locust spatial distribution features, and contextual features, the probability of locust disasters can be obtained.

[0148] The spatial distribution characteristics of locusts refer to the spatial distribution of locations where locust plagues occur. Contextual characteristics include the correlations between locust plague locations.

[0149] The target feature image is obtained by stitching together composite feature images from N time periods (e.g., 3 months). Figure 3 As shown, a 51-channel target feature image is input into the locust plague prediction model. The locust plague prediction model extracts the comprehensive feature image (including 17 channels) for month T, month T+1 (including 17 channels), and month T+2 (including 17 channels) from the target feature image. The size of the target feature image is adjusted to T×C×H×W=3×17×1241×465. Here, T represents the number of time periods (time steps), C represents the number of channels (including 16 channels of environmental data and 1 channel of locust plague occurrence density), H represents the height of the target feature image, and W represents the width of the target feature image.

[0150] Considering the varying contributions of environmental data from different time periods and historical locust plague records to the prediction of locust plague probability, and considering that the U-Net network input requires 2D feature maps, this invention designs a temporal convolutional attention module.

[0151] The temporal convolutional attention module is used to extract temporal features from the target feature image, resulting in a two-dimensional feature map containing these temporal features. For example, the size of the two-dimensional feature map containing temporal features is 32×3×H×W. This two-dimensional feature map containing temporal features is then input into the U-Net network.

[0152] The U-Net network extracts spatial distribution features and contextual features of locusts from a 2D feature map containing temporal features. The target feature image containing temporal features, locust spatial distribution features, and contextual features is input to the output module.

[0153] The output module generates a logits vector of length 2 for each pixel of the feature image (corresponding to the predicted occurrence of a locust plague or the absence of a locust plague) based on temporal features, locust spatial distribution features, and contextual features. The logits vector is transformed by an activation function (e.g., softmax) to obtain an occurrence probability map. The occurrence probability map includes the probability of a locust plague occurring at each pixel in the prediction region.

[0154] This invention extracts the temporal features of target feature images, which can explicitly consider the lag effect of environmental data, better conforms to the biological mechanism of locust plagues, and helps to improve the accuracy of predicting the probability of locust plagues.

[0155] Based on the above embodiments, the set time period includes N consecutive time periods, where N≥1, and the temporal convolutional attention module is used for:

[0156] Perform at least one 3D convolution on the target feature image to obtain the data features and their time weights for each time period;

[0157] The time series features are obtained by weighting and summing the data features and their time weights for each time period.

[0158] Perform at least one 3D convolution on the target feature image to obtain the data features and their temporal weights for each time period. Optionally, encode the target feature image as a feature tensor of (batch, 32, T, H, W). Subsequently, use a convolution with a kernel size of 1×1×1 to compute the attention logits for each spatial pixel at each time step, and then perform softmax (attention pooling and activation) to obtain the data features and their temporal weights for each time period.

[0159] The data features and their time weights for each time period are weighted and summed to obtain the temporal features. Through 1×1 two-dimensional convolution, the temporal features and other channel features of the target feature map are mixed and compressed in the spatial dimension to obtain a two-dimensional feature map containing the temporal features (size 32×H×W).

[0160] Furthermore, a two-dimensional feature map containing temporal features is input into the U-Net network.

[0161] This invention, by weighting and summing the data features and their time weights for each time period, can fully extract the differences in the contribution of environmental data and historical locust disaster records to the probability of locust disasters in different time periods, thus achieving accurate extraction of time-series features.

[0162] This invention addresses the challenges of predicting desert locust outbreaks, particularly the low accuracy of traditional models and the difficulty in establishing unified and effective prediction models for large areas. It proposes a temporal convolutional deep learning method that combines multi-temporal environmental factors with desert locust density images. The key features of this locust disaster prediction method are as follows.

[0163] (1) Fully consider the driving role of multi-source environmental data on the occurrence of desert locusts. The occurrence and outbreak of desert locusts is the result of the combined effects of multiple factors such as climate, hydrology, soil and vegetation. This invention fully incorporates multi-source environmental data such as precipitation, surface temperature, soil moisture and temperature, and vegetation index into the model input. (2) Pay attention to the scarcity and importance of ground survey data. In long-term monitoring data, some historical real survey points are deleted during data preprocessing because they are in the same grid. However, this information is still very valuable for the model to capture the spatiotemporal distribution pattern of desert locust occurrence. To this end, this invention innovatively proposes to accumulate and statistically analyze the repeated real survey point data (target record events) in the same grid coordinates to form a density distribution map reflecting the historical occurrence intensity, rather than a simple binary label of "occurrence / non-occurrence". This enables the locust disaster prediction model (or simply the model) to learn the nonlinear relationship between the occurrence probability and the historical intensity, thereby breaking through the information loss and limitations brought about by the traditional binary label and improving the locust disaster prediction model's ability to characterize and predict locust disaster risk under complex ecological conditions. (3) In the early stage of locust disaster prediction model input, 3D convolution (temporal convolution) is used to extract temporal features to explicitly consider the lag effect of environmental factors, which is more in line with the biological mechanism of locusts.

[0164] Compared to existing methods for predicting desert locust outbreaks, this invention proposes a U-Net network model (locust outbreak prediction model) based on temporal convolution. This model fully utilizes multi-temporal environmental data and density distribution maps of historical desert locust outbreaks (generated from historical locust outbreak records). By introducing a temporal convolutional attention module, it effectively extracts the dynamic features of environmental data over time, enhancing the locust outbreak prediction model's ability to characterize desert locust occurrence patterns. Compared to traditional machine learning models, this invention exhibits stronger generalization ability, significantly improved prediction accuracy, and the ability to achieve large-scale, continuous spatial prediction in major desert locust outbreak areas, rather than being limited to local regions. In summary, this invention provides more efficient and accurate technical support for early warning and macro-level control of desert locust outbreaks.

[0165] The locust disaster prediction device provided by the present invention will be described below. The locust disaster prediction device described below can be referred to in correspondence with the locust disaster prediction method described above.

[0166] like Figure 4 As shown, a locust plague prediction device includes:

[0167] The data acquisition module 401 is used to acquire a target dataset based on environmental data of the predicted area within a set time period and historical locust plague occurrence records; the historical locust plague occurrence records include the time and number of historical locust plague occurrences.

[0168] Preprocessing module 402 is used to generate target feature images of the prediction region based on the target dataset;

[0169] The prediction module 403 is used to input the target feature image into the locust disaster prediction model and obtain the probability of locust disaster occurrence in the predicted area output by the locust disaster prediction model. The locust disaster prediction model extracts the temporal features, locust spatial distribution features and contextual features of the target feature image. Based on the temporal features, locust spatial distribution features and contextual features, the probability of locust disaster occurrence is determined. The locust disaster prediction model is trained based on the sample feature image on the basis of the preset model.

[0170] The locust plague prediction device provided in this invention obtains a target dataset based on environmental data of the prediction area within a set time period and historical locust plague occurrence records. The historical locust plague occurrence records include the time and frequency of historical locust plague occurrences. A target feature image of the prediction area is generated based on the target dataset. The target feature image is input into a locust plague prediction model to obtain the probability of locust plague occurrence in the prediction area output by the model. The locust plague prediction model extracts temporal features, locust spatial distribution features, and contextual features from the target feature image. Based on these features, the probability of locust plague occurrence is determined. The locust plague prediction model is trained on a preset model using sample feature images. This invention introduces historical locust plague occurrence records into the environmental data, enabling the locust plague prediction model to fully combine the features of the environmental data with historical locust plague occurrence records to predict the probability of locust plague occurrence at various locations within the area, thus improving the accuracy of the locust plague prediction model in predicting the probability of locust plague occurrence under complex environmental data. This invention extracts the temporal features of target feature images, which can explicitly consider the lag effect of environmental data and historical locust disaster records, thereby generating a locust disaster probability that is more consistent with the biological mechanism of locust disasters and improving the prediction accuracy of locust disaster probability.

[0171] All relevant content of each step involved in the above method embodiments can be referenced from the functional description of the corresponding functional module, and will not be repeated here.

[0172] Figure 5 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 5As shown, the electronic device may include: a processor 510, a communication interface 520, a memory 530, and a communication bus 540, wherein the processor 510, the communication interface 520, and the memory 530 communicate with each other through the communication bus 540. The processor 510 can call logical instructions in the memory 530 to execute a locust plague prediction method, which includes: obtaining a target dataset based on environmental data of the prediction area within a set time period and historical locust plague occurrence records; the historical locust plague occurrence records include the time and frequency of historical locust plague occurrences; generating a target feature image of the prediction area based on the target dataset; inputting the target feature image into a locust plague prediction model to obtain the probability of locust plague occurrence in the prediction area output by the locust plague prediction model; wherein the locust plague prediction model extracts temporal features, locust spatial distribution features, and contextual features from the target feature image; and determines the probability of locust plague occurrence based on the temporal features, locust spatial distribution features, and contextual features; the locust plague prediction model is trained based on sample feature images on the basis of a preset model.

[0173] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0174] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements a method for predicting locust plagues provided by the methods described above. This method includes: obtaining a target dataset based on environmental data of the prediction area within a set time period and historical locust plague occurrence records; the historical locust plague occurrence records include the occurrence time and frequency of historical locust plagues; generating a target feature image of the prediction area based on the target dataset; inputting the target feature image into a locust plague prediction model to obtain the probability of locust plague occurrence in the prediction area output by the locust plague prediction model; wherein the locust plague prediction model extracts temporal features, locust spatial distribution features, and contextual features from the target feature image; determining the probability of locust plague occurrence based on the temporal features, locust spatial distribution features, and contextual features; the locust plague prediction model is trained based on sample feature images on the basis of a preset model.

[0175] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0176] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0177] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting locust plagues, characterized in that, include: Based on environmental data of the predicted area within a set time period and historical locust plague records, a target dataset is obtained; the historical locust plague records include the time and number of historical locust plagues. Generate a target feature image of the predicted region based on the target dataset; The target feature image is input into the locust disaster prediction model to obtain the probability of locust disaster occurrence in the predicted area output by the locust disaster prediction model. The locust disaster prediction model extracts the temporal features, locust spatial distribution features, and contextual features of the target feature image; based on the temporal features, the locust spatial distribution features, and the contextual features, it determines the probability of the locust disaster occurring; the locust disaster prediction model is trained based on sample feature images on the basis of a preset model. The sample feature image is determined based on the following method: Obtain the first sampling point where the locust plague actually occurred and its sample locust plague occurrence record; the sample locust plague occurrence record includes the time and number of locust plague occurrences; Based on the first sampling point, a second sampling point is generated where the locust plague did not actually occur. Based on the first sampling point where the locust plague actually occurred and the second sampling point where the locust plague did not actually occur, a label carrying whether the locust plague actually occurred and the target sampling point of the locust plague occurrence record are obtained. Obtain comprehensive sample environment images of the target sampling points in each time period; For each of the target sampling points, the time period and the time of occurrence of the sample locust plague are aligned to merge the comprehensive sample environment image and the number of times the sample locust plague occurs, thereby obtaining the merged sample environment image for each time period; The sample feature image is obtained by stitching together the merged sample environment images of at least one continuous time period.

2. The method for predicting locust plagues according to claim 1, characterized in that, The comprehensive sample environment image within the time period is determined based on multiple sample environment data of the target sampling point within the time period; the locust plague prediction model is trained in the following manner: The sample feature image is used as a training sample; A pre-defined model is constructed based on the initial temporal convolutional attention module and the initial U-Net network. The preset model is trained based on the training samples to obtain the locust disaster prediction model.

3. The method for predicting locust plagues according to claim 2, characterized in that, The locust plague records for the first sampling point and its samples were obtained in the following manner: Obtain a set of recorded events of locust plagues that actually occurred within at least one historical time period. The set of recorded events includes multiple recorded events of locust plagues that actually occurred, and each recorded event carries the location coordinates and time points of the actual locust plague occurrence. Merge all the recorded event sets and remove duplicate recorded events with the same location coordinates and the same time point to obtain the cleaned recorded event set; Based on the cleaned record event set, each target record event is obtained, and the target record event carries the target location coordinates and target time point of the locust plague. Aggregate and statistically analyze the events recorded for each target within the grid of location coordinates to obtain the time and frequency of locust plague occurrences within the grid of location coordinates; Each of the aforementioned location coordinates is used as a grid point for the first sampling point; Based on the time and frequency of locust infestations within the grid of the location coordinates, the locust infestation occurrence records of the sample are obtained.

4. The method for predicting locust plagues according to claim 1, characterized in that, The target dataset includes at least one type of environmental data from N consecutive time periods. Generating the target feature image of the prediction region based on the target dataset includes: Global standardization processing is performed on various environmental data for different time periods to obtain standardized environmental data for each time period. The standardized environmental data for each time period are converted into various environmental images for each time period. Based on the historical locust plague occurrence records, locust plague occurrence density images of the predicted area at various time periods are generated; By overlaying locust plague density images and various environmental images at different time periods, a comprehensive feature image for each time period is obtained. The target feature image is obtained by stitching together the comprehensive feature images from N time periods.

5. The method for predicting locust plagues according to claim 1, characterized in that, The locust plague prediction model includes a temporal convolutional attention module and a U-Net network. The locust plague prediction model is used for: Based on the temporal convolutional attention module, the temporal features of the target feature image are extracted to obtain a two-dimensional feature map containing the temporal features; Based on the U-Net network, the spatial distribution features of locusts and the context features of the two-dimensional feature map containing temporal features are extracted; The probability of locust disaster occurrence is obtained by fusing and activating the temporal features, the spatial distribution features of locusts, and the contextual features.

6. The method for predicting locust plagues according to claim 5, characterized in that, The defined time period includes N consecutive time periods, where N≥1, and the temporal convolutional attention module is used for: Perform at least one 3D convolution on the target feature image to obtain the data features and their time weights for each time period; The time series features are obtained by weighted summation of the data features and their time weights for each time period.

7. The method for predicting locust plagues according to claim 1, characterized in that, The environmental data includes static environmental data and dynamic environmental data; The dynamic environmental data includes at least one of the following: surface temperature, soil moisture content, surface wind speed, vegetation area, soil temperature, and precipitation. The static environmental data includes at least one of the following: stratum elevation data, soil composition, and land use type.

8. A device for predicting locust plagues, characterized in that, include: The data acquisition module is used to obtain a target dataset based on environmental data of the predicted area within a set time period and historical locust plague occurrence records; the historical locust plague occurrence records include the time and number of historical locust plague occurrences. The preprocessing module is used to generate a target feature image of the predicted region based on the target dataset; The prediction module is used to input the target feature image into a locust disaster prediction model and obtain the probability of locust disaster occurrence in the predicted area output by the locust disaster prediction model; wherein, the locust disaster prediction model extracts the temporal features, locust spatial distribution features, and contextual features of the target feature image; and determines the probability of locust disaster occurrence based on the temporal features, the locust spatial distribution features, and the contextual features; the locust disaster prediction model is trained based on sample feature images on the basis of a preset model; The sample feature image is determined based on the following method: Obtain the first sampling point where the locust plague actually occurred and its sample locust plague occurrence record; Based on the first sampling point, a second sampling point is generated where the locust plague did not actually occur. Based on the first sampling point where the locust plague actually occurred and the second sampling point where the locust plague did not actually occur, a label carrying whether the locust plague actually occurred and the target sampling point of the locust plague occurrence record are obtained. Obtain comprehensive sample environment images of the target sampling points in each time period; For each of the target sampling points, the time period and the time of occurrence of the sample locust plague are aligned to merge the comprehensive sample environment image and the number of times the sample locust plague occurs, thereby obtaining the merged sample environment image for each time period; The sample feature image is obtained by stitching together the merged sample environment images of at least one continuous time period.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the locust plague prediction method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the locust plague prediction method as described in any one of claims 1 to 7.

Images