Crop disaster monitoring, early warning and evaluation platform for main grain production area

Abstract

The invention discloses a crop disaster monitoring, early warning and evaluation platform for a main grain production area, and relates to the technical field of crop disaster monitoring and early warning. Interference areas such as agricultural implements and ridges are effectively stripped through the synergistic effect of a space attention mechanism and dynamic cutting, so that disaster recognition focuses on crop body characteristics, and the crop disaster monitoring and early warning effects are improved. The misjudgment risk caused by background noise is reduced; adaptive illumination correction eliminates overexposure / underexposure influences while leaf texture details are reserved through dynamic interval division and segmented mapping, and the quality tolerance of the model to crowdsourcing images is enhanced; the geographic position code embeds the spatial attributes of the production area into the feature channel, and the spatial constraint rule of the crop distribution database is combined to inhibit the misrecognition of the cross-regional crop diseases; the multi-time-scale gating network coordinates the long-term trend of meteorological prediction and the short-term fluctuation of real-time data of the sensor through dynamic weight distribution, and improves the early warning timeliness of composite disasters such as dry and hot air.

Description

Technical Field

[0001] The present invention relates to the technical field of crop disaster monitoring and early warning, and particularly to a crop disaster monitoring, early warning and evaluation platform in the main grain-producing areas. Background Art

[0002] Current grain disaster monitoring relies on satellite remote sensing, Internet of Things sensing nodes and mobile crowd-sourced data to achieve multi-source information fusion. Farmers upload real-time images of field disaster situations through intelligent terminals, which become an important data supplement for local sudden disaster monitoring. Mainstream systems generally adopt lightweight convolutional neural networks, such as EfficientNet-Lite, for end-to-end disaster recognition, and introduce a federated learning mechanism to optimize the model generalization ability.

[0003] However, crowd-sourced images are affected by the performance of shooting devices, the standardization of farmers' operations and the complexity of field environments, and there are noise problems such as subject offset, abnormal illumination, and background interference, resulting in misjudgment of the model for early disaster features. In response to the quality problems of crowd-sourced data, some solutions adopt a two-stage optimization strategy: in the data preprocessing stage, a method combining adaptive histogram equalization and semantic segmentation is used to strip background interference areas such as soil and agricultural machinery through a U-Net network; in the model training stage, a channel attention mechanism is introduced to strengthen the feature response of crop organ regions, and a generative adversarial network is used to synthesize training samples under extreme illumination conditions. However, background segmentation depends on high-precision annotation data and is difficult to adapt to the complex scenarios of more than a hundred crop combinations in the main producing areas. At the same time, the generation of adversarial samples is likely to cause feature distortion, resulting in a decrease in the robustness of the model to real noise. The process of end-side preprocessing and cloud inference is fragmented, and it is difficult to achieve the collaborative optimization of dynamic noise suppression and feature learning. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] The present invention provides a crop disaster monitoring, early warning and evaluation platform in the main grain-producing areas to solve the pain points faced by the existing disaster monitoring systems in the main grain-producing areas, such as serious noise interference in crowd-sourced images, temporal conflicts in multi-source data, and insufficient new disaster samples, and the problems of feature confusion and response delay easily caused by traditional preprocessing and model design in complex scenarios.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: An embodiment of the present invention provides a crop disaster monitoring, early warning and evaluation platform in the main grain-producing areas, which includes a mobile terminal data acquisition module configured to receive field disaster situation images uploaded by farmers through intelligent terminals and associated geographical location data; The cloud data processing module is communicatively connected to the mobile terminal data acquisition module and includes an image preprocessing sub-module, a disaster recognition model construction sub-module, and a multi-source data fusion sub-module; Among them, the image preprocessing sub-module integrates a spatial attention mechanism network for dynamically locating the main crop area in the image and performing background interference filtering; the disaster recognition model construction sub-module embeds a geographical location encoding unit to map geographical location data into a feature channel adjustment vector; the multi-source data fusion sub-module is configured to perform spatio-temporal alignment and feature-level fusion on the preprocessed image data with satellite remote sensing and ground sensor data.

[0007] As a preferred solution of the crop disaster monitoring, early warning and assessment platform for major grain-producing areas of the present invention, wherein: the spatial attention mechanism network is composed of 3 cascaded depthwise separable convolutional layers, and each convolutional layer is followed by a ReLU activation function and a max pooling operation, and finally outputs a saliency heat map with 1 channel; The dimension of the feature channel adjustment vector is equal to the number of feature channels of the disaster recognition model, and it acts on the output feature map of the convolutional layer through matrix dot multiplication.

[0008] As a preferred solution of the crop disaster monitoring, early warning and assessment platform for major grain-producing areas of the present invention, wherein: the image preprocessing sub-module further includes: A spatial attention sub-module that generates an image region saliency heat map based on a lightweight convolutional neural network; A dynamic cropping sub-module that generates a minimum bounding rectangle according to the saliency heat map and performs adaptive region cropping on the original image; A lighting correction unit that uses a piecewise linear transformation algorithm to perform brightness equalization processing on the cropped image.

[0009] As a preferred solution of the crop disaster monitoring, early warning and assessment platform for major grain-producing areas of the present invention, wherein: in the lighting correction unit, using a piecewise linear transformation algorithm to perform brightness equalization processing on the cropped image, including: adaptively determining the segmentation thresholds of the low and high brightness intervals according to the peak of the brightness histogram of the cropped image: , wherein, represents the brightness level index, represents the number of pixels with brightness in the cropped image, represents rounding down, represents rounding up, represents the maximum possible brightness value of the image, represents the low brightness interval segmentation threshold, represents the high brightness interval segmentation threshold; Divide the brightness range into three intervals , , , and for each original pixel brightness value apply the following piecewise linear mapping: , wherein, represents the original brightness value of the pixel in the cropped image, represents the corrected brightness value, represents dividing the brightness range into three equal segments.

[0010] As a preferred solution of a crop disaster monitoring, early warning and assessment platform in the main grain-producing areas of the present invention, wherein: the geographical location coding unit includes: A longitude and latitude mapping layer that converts GPS coordinates into embedded vectors of a fixed dimension; A channel suppression layer that generates a feature channel mask matrix according to the embedded vector, and is used to suppress crop disease features irrelevant to the current geographical location; The feature channel mask matrix is generated by a sigmoid function, and its element values range between [0,1], and are used to attenuate the feature responses related to non-local crops in the channel dimension.

[0011] As a preferred solution of a crop disaster monitoring, early warning and assessment platform in the main grain-producing areas of the present invention, wherein: the disaster identification model construction sub-module further includes: A few-shot learning unit that integrates a meta-learning framework based on a prototype network, and is configured to construct a category prototype vector library using historical disaster images; An incremental training interface that responds to newly issued disaster type annotation data and triggers dynamic expansion of the prototype vector library and fine-tuning of model parameters.

[0012] As a preferred solution of a crop disaster monitoring, early warning and assessment platform in the main grain-producing areas of the present invention, wherein: the prototype vector library is communicatively connected to the model hot update module, and the hot update module is configured to: Receive a newly issued disaster feature description file pushed by the agricultural management department; Parse the key morphological parameters in the feature description file to generate corresponding prototype vector patches; Inject the prototype vector patches into the disaster identification model construction sub-module through a differential update protocol.

[0013] As a preferred solution of a crop disaster monitoring, early warning and assessment platform in the main grain-producing areas of the present invention, wherein: the disaster identification model construction sub-module is further connected to a crop feature confusion suppression module, and the crop feature confusion suppression module includes: Multi-crop feature decoupling layer, inserting a geographical location-related channel weight filter before the fully connected layer of the feature extraction network; Cross-crop false alarm correction unit, generating spatial constraint rules based on the main producing area crop planting distribution database, and post-processing and optimizing the disaster probability distribution output by the model.

[0014] As a preferred solution of the crop disaster monitoring, early warning and assessment platform in the main grain producing areas of the present invention, wherein: the multi-source data fusion sub-module includes: Multi-time scale gating network, configured to respectively extract temporal features of meteorological forecast data and real-time sensor data, and dynamically adjust the fusion ratio of the two types of data through a learnable weight matrix; Feature alignment unit, adopting a sliding window mechanism to match the spatial correspondence between data with different resolutions; The dimension of the learnable weight matrix is equal to the time step of the meteorological feature vector, and is jointly trained and updated with the disaster recognition model through the backpropagation algorithm.

[0015] As a preferred solution of the crop disaster monitoring, early warning and assessment platform in the main grain producing areas of the present invention, wherein: in the multi-time scale gating network, the steps of temporal feature extraction and dynamic fusion include: For the meteorological forecast data sequence And the sensor data sequence , respectively construct two symmetric causal one-dimensional convolutional branches, , Among them, Represents the meteorological forecast input vector at the th moment, Represents the meteorological forecast data sequence, Represents the ground sensor data sequence, Represents the time offset of the convolutional kernel, Represents the convolutional kernel length, Represents the weight of the th convolutional kernel in the first layer of the meteorological branch, Represents the dimension of the input feature vector at each moment, Represents the output feature dimension of the convolutional layer, Represents the bias, Represents the output feature of the first layer, Represents the rectified linear unit function; , Among them, Represents the weight of the th convolutional kernel in the second layer of the meteorological branch, Represents the bias, Represents the output features of the second layer of the meteorological branch; The sensor branch mirrors the above structure and generates and , and the corresponding parameters are: , , , , Subsequently, perform temporal average pooling on the deep outputs of the two branches: , , Among them, represents the total sequence length, are the global temporal feature vectors of the meteorological and sensor branches respectively; Concatenate the features of the two branches and input them into the fusion layer to generate dynamic fusion weights through a learnable matrix: , Among them, represents feature concatenation, represents the fusion weight matrix, represents the fusion bias, represents the normalization function, satisfies and ; Finally, output the fused features: , Among them, represents the multi-source temporal features after dynamic weighting.

[0016] The beneficial effects of the present invention are as follows: Through multi-dimensional technological innovation, the present invention significantly improves the practicability and reliability of the disaster monitoring system in the main grain-producing areas: The synergistic effect of the spatial attention mechanism and dynamic cropping effectively strips interference areas such as agricultural machinery and field ridges, enabling disaster recognition to focus on the crop body features and reducing the risk of misjudgment caused by background noise. Adaptive light correction eliminates overexposure / underexposure effects while retaining leaf texture details through dynamic interval division and segmented mapping, enhancing the model's tolerance to the quality of crowdsourced images.

[0017] Geographical location encoding embeds the spatial attributes of the production area into the feature channels, and combines the spatial constraint rules of the crop distribution database to suppress the misrecognition of cross-regional crop diseases. The multi-time-scale gated network coordinates the long-term trend of meteorological prediction and the short-term fluctuations of sensor real-time data through dynamic weight allocation, improving the early warning timeliness of compound disasters such as dry hot winds.

[0018] An incremental learning framework based on prototype networks supports the rapid adaptation of the features of newly emerging migratory pests, breaking through the limitation of traditional models relying on large-scale labeled data. The differential update mechanism ensures that the disaster feature patches pushed by provincial management departments can be seamlessly integrated, enabling minute-level model iteration.

[0019] The lightweight convolution design and dynamic feature compression reduce the computational overhead, enabling the system to process tens of thousands of crowdsourced images in parallel under standard server configurations and meet the high-concurrency requirements of sudden disasters. Brief Description of the Drawings

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0021] Figure 1 It is a schematic diagram of the framework of the crop disaster monitoring, early warning and assessment platform in the main grain-producing areas in Embodiment 1. Detailed Embodiments

[0022] To make the above objects, features, and advantages of the present invention more obvious and understandable, the detailed embodiments of the present invention will be described in detail below with reference to the drawings of the specification.

[0023] Many specific details are set forth in the following description to facilitate a thorough understanding of the present invention. However, the present invention can also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0024] Secondly, the so-called "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that can be included in at least one implementation manner of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that excludes other embodiments.

[0025] Embodiment 1, referring to Figure 1 , this embodiment provides a crop disaster monitoring, early warning and assessment platform in the main grain-producing areas, including: A mobile terminal data collection module configured to receive the field disaster images and associated geographical location data uploaded by farmers through intelligent terminals; A cloud data processing module communicatively connected to the mobile terminal data collection module, including an image preprocessing sub-module, a disaster recognition model construction sub-module, and a multi-source data fusion sub-module; Among them, the image preprocessing sub-module integrates a spatial attention mechanism network, which is used to dynamically locate the main crop area in the image and perform background interference filtering; The spatial attention mechanism network consists of 3 serially connected depthwise separable convolutional layers. After each convolutional layer, a ReLU activation function and a max-pooling operation are connected, and finally a saliency heat map with 1 output channel is obtained; The dimension of the feature channel adjustment vector is equal to the number of feature channels of the disaster recognition model, and it acts on the output feature map of the convolutional layer through matrix dot multiplication; The image preprocessing sub-module further includes: A spatial attention sub-module, which generates a saliency heat map of the image area based on a lightweight convolutional neural network; A dynamic cropping sub-module, which generates a minimum bounding rectangle according to the saliency heat map and performs adaptive region cropping on the original image; A lighting correction unit, which uses a piecewise linear transformation algorithm to perform brightness equalization processing on the cropped image; The generation condition of the minimum bounding rectangle is that the pixel saliency of the continuous area in the heat map exceeds a preset threshold, and the area of the area is not less than 10%-20% of the total image area; The piecewise linear transformation algorithm dynamically divides the low, medium, and high brightness intervals according to the peak distribution of the image brightness histogram, and performs contrast stretching on each interval using independent linear coefficients; In the lighting correction unit, using the piecewise linear transformation algorithm to perform brightness equalization processing on the cropped image includes: adaptively determining the segmentation thresholds of the low and high brightness intervals according to the peak of the brightness histogram of the cropped image: , Among them, represents the brightness level index, represents the number of pixels with brightness in the cropped image, represents rounding down, represents rounding up, represents the maximum possible brightness value of the image, represents the low brightness interval segmentation threshold, represents the high brightness interval segmentation threshold; The brightness range is divided into three intervals , , , and for each pixel's original brightness value the following piecewise linear mapping is applied: , Among them, represents the original brightness value of the pixel in the cropped image, Indicates the corrected brightness value. Indicates that the brightness range is equally divided into three segments.

[0026] Specifically, here, the local peak of the brightness histogram is used to accurately locate the segmentation point between the low-light and high-light regions, enabling the threshold to adapt to different scene lighting conditions, avoiding the failure of a fixed threshold. The segmented mapping equally divides the brightness interval into three segments, and different linear mapping coefficients are used for the dark part, the middle part, and the high light respectively. This can not only stretch the details in the dark part but also suppress the overexposure of the high light, enhancing the overall contrast. The coefficient form is inversely proportional to the interval width, ensuring smooth mapping and no introduced artifacts. This adaptive equalization method takes into account both detail retention and global brightness balance, improving the robustness and accuracy of the disaster recognition model under extreme lighting conditions. The disaster recognition model construction sub-module embeds a geographical location encoding unit, which maps geographical location data into a feature channel adjustment vector. The geographical location encoding unit includes: A latitude and longitude mapping layer that converts GPS coordinates into an embedded vector of a fixed dimension. A channel suppression layer that generates a feature channel mask matrix according to the embedded vector, which is used to suppress crop disease features unrelated to the current geographical location. The feature channel mask matrix is generated through a sigmoid function, and its element values range between [0, 1], which is used to attenuate the feature responses related to non-local crops in the channel dimension.

[0027] The disaster recognition model construction sub-module also includes: A few-shot learning unit that integrates a meta-learning framework based on a prototype network and is configured to construct a category prototype vector library using historical disaster images. An incremental training interface that responds to newly issued disaster type annotation data and triggers the dynamic expansion of the prototype vector library and the fine-tuning of model parameters. The prototype vector library is communicatively connected to a model hot update module, and the hot update module is configured to: Receive a newly issued disaster feature description file pushed by the agricultural management department. Parse the key morphological parameters in the feature description file to generate corresponding prototype vector patches. Inject the prototype vector patches into the disaster recognition model construction sub-module through a differential update protocol. The disaster recognition model construction sub-module is also connected to a crop feature confusion suppression module, which includes: A multi-crop feature decoupling layer that inserts a geographical location-related channel weight filter before the fully connected layer of the feature extraction network. A cross-crop false alarm correction unit that generates a spatial constraint rule based on the main production area crop planting distribution database and post-processes and optimizes the disaster probability distribution output by the model. The multi-source data fusion sub-module is configured to perform spatio-temporal alignment and feature-level fusion on the preprocessed image data with satellite remote sensing and ground sensor data; The multi-source data fusion sub-module includes: The multi-time scale gating network is configured to extract temporal sequence features from meteorological forecast data and real-time sensor data respectively, and dynamically adjust the fusion ratio of the two types of data through a learnable weight matrix; The feature alignment unit uses a sliding window mechanism to match the spatial correspondence between data with different resolutions; The dimension of the learnable weight matrix is equal to the time step of the meteorological feature vector, and it is jointly trained and updated with the disaster recognition model through the backpropagation algorithm; In the multi-time scale gating network, the steps of temporal sequence feature extraction and dynamic fusion include: For the meteorological forecast data sequence and the sensor data sequence , two symmetric causal one-dimensional convolutional branches are respectively constructed, , where, represents the meteorological forecast input vector at the th moment, represents the meteorological forecast data sequence, represents the ground sensor data sequence, represents the time offset of the convolutional kernel, represents the length of the convolutional kernel, represents the weight of the th convolutional kernel in the first layer of the meteorological branch, represents the dimension of the input feature vector at each moment, represents the output feature dimension of the convolutional layer, represents the bias, represents the output feature of the first layer, represents the rectified linear unit function; , where, represents the weight of the th convolutional kernel in the second layer of the meteorological branch, represents the bias, represents the output feature of the second layer of the meteorological branch; The sensor branch mirrors the above structure to generate and , and the corresponding parameters are respectively: , , , , Subsequently, perform temporal average pooling on the deep outputs of the two branches: , , wherein, represents the total length of the sequence, are the global temporal feature vectors of the meteorological and sensor branches respectively; Concatenate the features of the two branches and input them into the fusion layer to generate dynamic fusion weights through a learnable matrix: , wherein, represents feature concatenation, represents the fusion weight matrix, represents the fusion bias, represents the normalization function, satisfies and ; Finally, output the fused features: , wherein, represents the multi-source temporal features after dynamic weighting.

[0028] Specifically, two parallel causal convolutional branches are used to specifically extract the time-dependent features of weather forecasts and on-site sensors, retain the time-domain information of their respective signals, and the average pooling operation compresses the sequence into a global feature vector, reducing the subsequent computational complexity. The fusion layer uses a learnable matrix and softmax normalization, which can dynamically allocate the importance ratio of the two according to the content of the input data, achieving adaptive emphasis on different scenarios (such as extreme weather or local emergencies). This dynamic fusion design takes into account the complementarity of the two types of data, improves the response rate of the model to mutation and delay information. At the same time, the weight matrix is jointly trained with the disaster recognition model to ensure that the fusion strategy is consistent with the downstream task objectives, thus significantly improving the overall recognition accuracy and robustness.

[0029] 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 preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A crop disaster monitoring, early warning and assessment platform in the main grain-producing areas, characterized in that, including, a mobile terminal data acquisition module configured to receive field disaster images and associated geographical location data uploaded by farmers through intelligent terminals; a cloud data processing module communicatively connected to the mobile terminal data acquisition module, including an image preprocessing sub-module, a disaster recognition model construction sub-module, and a multi-source data fusion sub-module; wherein, the image preprocessing sub-module integrates a spatial attention mechanism network for dynamically locating the main crop area in the image and performing background interference filtering; the disaster recognition model construction sub-module embeds a geographical location encoding unit for mapping geographical location data into a feature channel adjustment vector; the multi-source data fusion sub-module is configured to perform spatio-temporal alignment and feature-level fusion on the preprocessed image data with satellite remote sensing and ground sensor data.

2. The crop disaster monitoring, early warning and assessment platform in the main grain-producing area according to claim 1, wherein The spatial attention mechanism network is composed of 3 cascaded depthwise separable convolutional layers, and each convolutional layer is followed by a ReLU activation function and a max pooling operation, and finally outputs a saliency heat map with 1 channel; The dimension of the feature channel adjustment vector is equal to the number of feature channels of the disaster recognition model, and it acts on the output feature map of the convolutional layer through matrix dot multiplication.

3. The crop disaster monitoring, early warning and assessment platform in the main grain producing areas according to claim 2, characterized in that, The image preprocessing sub-module further includes: a spatial attention sub-module for generating an image region saliency heat map based on a lightweight convolutional neural network; a dynamic cropping sub-module for generating a minimum bounding rectangle according to the saliency heat map and performing adaptive region cropping on the original image; a lighting correction unit for performing brightness equalization processing on the cropped image using a piecewise linear transformation algorithm.

4. The crop disaster monitoring, early warning and assessment platform in the main grain-producing area according to claim 3, characterized in that, In the lighting correction unit, the piecewise linear transformation algorithm is used to perform brightness equalization processing on the cropped image, including: adaptively determining the segmentation thresholds of the low and high brightness intervals according to the peak of the brightness histogram of the cropped image: ， Among them, represents the brightness level index, represents the number of pixels with brightness in the cropped image, represents rounding down, represents rounding up, represents the maximum possible brightness value of the image, represents the segmentation threshold for the low brightness interval, represents the segmentation threshold for the high brightness interval; Divide the brightness range into three intervals , , , and apply the following piecewise linear mapping to the original brightness value of each pixel : ， Among them, represents the original brightness value of the pixel in the cropped image, represents the corrected brightness value, represents dividing the brightness range into three equal segments.

5. The crop disaster monitoring, early warning and assessment platform in the main grain-producing area according to claim 1, wherein, The geographical location encoding unit includes: a latitude and longitude mapping layer for converting GPS coordinates into an embedded vector of a fixed dimension; a channel suppression layer for generating a feature channel mask matrix according to the embedded vector to suppress crop disease features irrelevant to the current geographical location; The feature channel mask matrix is generated by a sigmoid function, and its element values range between [0,1], and are used to attenuate the feature responses related to non-local crops in the channel dimension.

6. The crop disaster monitoring, early warning and assessment platform in the main grain-producing area according to claim 5, characterized in that The disaster recognition model construction sub-module further includes: a few-shot learning unit integrating a meta-learning framework based on a prototype network, configured to construct a category prototype vector library using historical disaster images; an incremental training interface for triggering dynamic expansion of the prototype vector library and fine-tuning of model parameters in response to newly issued disaster type annotation data.

7. The crop disaster monitoring, early warning and assessment platform in the main grain-producing area according to claim 6, characterized in that, The prototype vector library is communicatively connected to a model hot update module, and the hot update module is configured to: receive a newly issued disaster feature description file pushed by the agricultural management department; parse the key morphological parameters in the feature description file to generate corresponding prototype vector patches; inject the prototype vector patches into the disaster recognition model construction sub-module through a differential update protocol.

8. The crop disaster monitoring, early warning and assessment platform in the main grain producing areas as claimed in claim 7, characterized in that, The disaster recognition model construction sub-module is also connected to a crop feature confusion suppression module, and the crop feature confusion suppression module includes: Multi-crop feature decoupling layer, inserting a geographical location-related channel weight filter before the fully connected layer of the feature extraction network; Cross-crop false alarm correction unit, generating a spatial constraint rule based on the main producing area crop planting distribution database, and post-processing and optimizing the disaster probability distribution output by the model.

9. The crop disaster monitoring, early warning and assessment platform in the main grain producing areas according to claim 1, characterized in that, The multi-source data fusion sub-module includes: Multi-time scale gating network, configured to extract temporal features from meteorological forecast data and real-time sensor data respectively, and dynamically adjust the fusion ratio of the two types of data through a learnable weight matrix; Feature alignment unit, using a sliding window mechanism to match the spatial correspondence between data with different resolutions; The dimension of the learnable weight matrix is equal to the time step of the meteorological feature vector, and is jointly trained and updated with the disaster recognition model through the backpropagation algorithm.

10. A crop disaster monitoring, early warning and assessment platform in the main grain-producing areas as described in claim 9, characterized in that, In the multi-time scale gating network, the steps of temporal feature extraction and dynamic fusion include: For the meteorological forecast data sequence and the sensor data sequence , two symmetric causal one-dimensional convolutional branches are respectively constructed. ， Among them, represents the meteorological forecast input vector at the moment, represents the meteorological forecast data sequence, represents the ground sensor data sequence, represents the time offset of the convolution kernel, represents the convolution kernel length, represents the weight of the th convolution kernel in the first layer of the meteorological branch, represents the dimension of the input feature vector at each moment, represents the output feature dimension of the convolution layer, represents the bias, represents the output feature of the first layer, represents the rectified linear unit function; ， Among them, represents the weights of the th convolutional kernel in the second layer of the meteorological branch, represents the bias, represents the output features of the second layer of the meteorological branch; The sensor branch mirrors the above structure to generate and with the corresponding parameters being respectively: 、 、 、 ， Subsequently, perform temporal average pooling on the deep outputs of the two branches: ， , Among them, represents the total sequence length, are respectively the global temporal feature vectors of the meteorological and sensor branches; Concatenate the features of the two branches and input them into the fusion layer to generate dynamic fusion weights through a learnable matrix: ， Among them, represents feature splicing, represents the fusion weight matrix, represents the fusion bias, represents the normalization function, satisfies and ; Finally, output the fused features: ， Among them, represents the multi-source time-series features after dynamic weighting.

Images