Shadow operation target cloud layer identification method, system and device and medium

Abstract

The invention relates to a weather modification work target cloud layer identification method, system and device and a medium, and belongs to the technical field of weather cloud layer identification, and the identification method comprises the steps: obtaining original cloud layer image data of a target cloud layer region, carrying out the preprocessing, and generating a standardized feature image; performing multi-level feature extraction on the standardized feature image based on a feature extraction network to generate a feature map; traversing the feature map based on a region suggestion network, and generating an initial detection result; performing feature alignment processing on the bounding box of the candidate region based on a region-of-interest pooling algorithm, and outputting feature vectors of the candidate region to a classification regression network to obtain a detection result set; performing full convolution segmentation operation on each candidate region in the detection result set to generate a pixel-level segmentation mask; the cloud cover proportion, the cloud layer category, the texture morphological characteristics and the vertical height parameters of the target cloud layer area are calculated respectively, and a cloud layer characteristic parameter set is constructed and obtained; the cloud feature recognition accuracy can be improved.

Description

Technical Field

[0001] This application relates to the technical field of meteorological cloud identification, and in particular, to a method, system, device, and medium for identifying target clouds for weather modification operations. Background Art

[0002] Weather modification (abbreviated as "WM") operation is a technical means to improve or regulate local climate conditions by artificially intervening in natural processes. Its main purpose is to promote precipitation, reduce hail, or improve weather conditions through artificial means, so as to achieve the goals of regulating climate, promoting agricultural irrigation, reducing forest fires, etc. Typical WM operations include rain enhancement, artificial hail suppression, etc., and these operations are of great significance for improving agricultural production and ensuring the climate adaptability in human life and economic activities.

[0003] Currently, WM operations usually need to be carried out at an appropriate time and must be targeted at specific clouds and their meteorological conditions. For example, only when the clouds have sufficient moisture, thickness, and updrafts, can operations such as artificial rain enhancement achieve the expected effects. Therefore, accurately evaluating various characteristics of clouds, especially the cloud type, shape, cloud height, etc., is a prerequisite for effective operations.

[0004] However, due to the variability and complexity of clouds, the existing methods have weak adaptability under different weather conditions. Especially when dealing with cloud images at different scales and complex backgrounds, problems such as inaccurate feature extraction and incorrect cloud classification often occur, resulting in inaccurate decision support and reducing the efficiency of weather modification operations. Summary of the Invention

[0005] In order to improve the accuracy of cloud feature recognition, this application provides a method, system, device, and medium for identifying target clouds for weather modification operations.

[0006] In the first aspect, this application provides a method for identifying target clouds for weather modification operations, adopting the following technical solution: A method for identifying target clouds for weather modification operations, the identification method includes: Obtain the original cloud image data of the target cloud area; Preprocess the original cloud image data to generate a standardized feature image; Pre-construct a deep convolutional neural network architecture, and perform multi-level feature extraction on the standardized feature image based on the feature extraction network to generate a feature map; Traverse the feature map based on the region proposal network to generate an initial detection result including candidate region bounding boxes; Perform feature alignment processing on the candidate region bounding boxes based on the region of interest pooling algorithm, and output the feature vectors of the candidate regions; Input the feature vectors of the candidate regions into a classification regression network, perform multi-class classification prediction and bounding box position regression calculation, and output a detection result set including cloud class labels and target cloud bounding boxes; Perform full convolutional segmentation operations on each candidate region in the detection result set to generate pixel-level segmentation masks; According to the pixel-level segmentation masks and cloud class labels, calculate the cloud amount ratio, cloud class, texture morphology features, and vertical height parameters of the target cloud region respectively, and construct a cloud feature parameter set.

[0007] By adopting the above technical solutions, the high-resolution visible light imaging system is combined with a deep convolutional neural network. On the basis of realizing multi-scale feature extraction of cloud images, through the collaborative optimization of the Region Proposal Network (RPN) and the Fully Convolutional Network (FCN), a multi-dimensional parameter system including pixel-level segmentation masks, cloud class coding matrices, and morphological feature vectors is constructed, effectively solving the technical problems such as low efficiency existing in traditional observation methods when facing complex backgrounds. It not only realizes the automatic and accurate analysis of cloud amount distribution, cloud class, and texture features, but also can provide multi-dimensional decision-making support for weather modification operations, including the judgment of catalytic timing and the delimitation of operation scope, through the real-time generated cloud feature parameter set, significantly improving the scientificity and timeliness of weather modification operations.

[0008] Optionally, the step of traversing the feature map based on the Region Proposal Network to generate an initial detection result including candidate region bounding boxes includes: Apply a convolutional kernel on the feature map for sliding window scanning to generate an intermediate feature map; Perform dual-branch convolution operations on the intermediate feature map to generate an anchor box classification confidence tensor and an anchor box position offset tensor for each spatial position respectively; Based on a preset reference anchor box template, generate corresponding anchor box groups at each spatial position of the feature map; According to the anchor box classification confidence tensor and the anchor box position offset tensor, calculate the initial bounding box coordinates of the candidate regions based on the anchor box groups; According to the initial bounding box coordinates of the candidate regions, calculate the intersection over union (IoU) between the initial bounding box of the candidate region and the true cloud annotation box, and filter and retain the anchor boxes in the anchor box groups that meet the preset IoU threshold; Perform spatial correction on the filtered anchor box groups to obtain effective anchor box groups; Perform iterative processing on the effective anchor box groups based on the non-maximum suppression algorithm, and output the optimized candidate region bounding boxes as the initial detection results.

[0009] By adopting the above technical solution, combining the Region Proposal Network (RPN) and the object detection framework, the precise detection of cloud targets can be efficiently achieved. Through steps such as multi-level feature extraction, anchor box regression, and non-maximum suppression, the target clouds in the image can be effectively identified, thereby providing real-time and precise data support for weather modification operations.

[0010] Optionally, the steps of performing feature alignment processing on the candidate region bounding box based on the Region of Interest Pooling (ROI Pooling) algorithm and outputting the feature vector of the candidate region include: Based on the original cloud image data and the scaling ratio parameter of the feature map, map the candidate region to the corresponding spatial position on the feature map; Evenly divide the mapped candidate region into K rows and K columns of grid sub-regions in the feature map space; where K is a preset positive integer; Perform a max pooling operation on each grid sub-region, extract the maximum pooling result of the features of each grid sub-region, and splice them in spatial order to obtain the feature vector of the candidate region.

[0011] By adopting the above technical solution, the steps of performing feature alignment processing on the candidate region based on the Region of Interest Pooling (ROI Pooling) algorithm can effectively convert candidate regions of different sizes into a unified-size feature representation. This step solution can not only handle the processing problems of candidate regions of different sizes but also extract the most representative features through max pooling, thereby improving the performance and efficiency of the object detection task.

[0012] Optionally, the steps of pre-constructing the deep convolutional neural network architecture include: Obtain the original image sample set of meteorological clouds and perform standardization processing, and divide it into a training set and a validation set; Construct a three-level network architecture including a feature extraction network, a region proposal network, and a classification and regression network; Input the training set into the three-level network architecture, generate the corresponding shared feature map based on the feature extraction network, and perform sliding window convolution operations based on the region proposal network to generate a multi-scale anchor box map; By calculating the intersection over union of the anchor box and the labeled bounding box, filter positive and negative samples and construct a binary classification label; Based on the binary classification label, optimize the classification branch of the region proposal network through the binary cross-entropy loss function, and optimize the regression branch of the region proposal network through the smooth L1 loss function; Map the candidate regions generated by the region proposal network to the shared feature map, perform region of interest pooling operations, and generate fixed-size feature vectors; Input the fixed-size feature vector into the fully connected layer of the classification and regression network for dimensionality reduction processing, and output a dimensionality reduction feature matrix; Perform multi-task prediction based on the dimensionality-reduced feature matrix, and calculate the class probability distribution and bounding box offset; Obtain the labeled target cloud class labels and bounding box coordinates in the training set, calculate the classification loss according to the class probability distribution and the target cloud class labels, and calculate the regression loss according to the bounding box offset and the bounding box coordinates; Jointly optimize the classification loss and the regression loss through the backpropagation algorithm, and update the weight parameters of the classification regression network; Freeze the weight parameters of the feature extraction network, jointly train the region proposal network and the classification regression network to generate a primary training model; Unfreeze the weight parameters of the feature extraction network, perform global parameter fine-tuning based on the primary training model, and update the weight parameters of the feature extraction network, the region proposal network, and the classification regression network; Validate the primary training model based on the validation set until the total loss function meets the preset conditions or the model iteration times reach the preset times, and output the completed deep convolutional neural network architecture.

[0013] By adopting the above technical solutions, through multi-level feature extraction, accurate candidate region generation and screening, region of interest pooling alignment, and multi-task learning, the classification accuracy and bounding box localization accuracy of the model are significantly improved. By gradually optimizing the classification and regression losses, the model can efficiently process cloud targets of different sizes and shapes, and finally realizes an end-to-end deep learning model with high accuracy, robustness, and generalization ability in the cloud detection task.

[0014] Optionally, the step of performing full convolutional segmentation operations on each candidate region in the detection result set to generate a pixel-level segmentation mask includes: Input the feature vector of the candidate region into the segmentation sub-network of the fully convolutional neural network, and perform upsampling processing on the feature vector of the candidate region through at least three deconvolution layers to generate an initial segmentation map; Perform per-pixel classification operations on the initial segmentation map to generate a probability distribution map containing the probability values of each pixel belonging to the target cloud; Perform binarization processing on the probability distribution map based on a preset probability threshold, and output a pixel-level segmentation mask containing cloud pixel marking information.

[0015] By adopting the above technical solutions, the pixel-level segmentation of the target cloud layer is accurately achieved. Through the upsampling processing of the deconvolution layer, the spatial details of the feature map of the candidate region are restored, providing high-resolution input for the subsequent pixel-by-pixel classification. The probability distribution map generated by the pixel-by-pixel classification operation can accurately predict the probability of each pixel belonging to the cloud layer, and the threshold-based binarization processing effectively converts the probability value into a clear cloud segmentation mask, ensuring the accurate identification and marking of the cloud area. The whole process not only improves the accuracy of cloud segmentation, but also makes full use of the advantages of deep learning technology to provide an efficient and accurate solution when processing complex meteorological images.

[0016] Optionally, the step of respectively calculating the cloud amount ratio, cloud category, texture morphological characteristics and vertical height parameters of the target cloud area according to the pixel-level segmentation mask and the cloud category label, and constructing a cloud feature parameter set includes: Performing binarization processing based on the pixel-level segmentation mask to generate a valid cloud area pixel labeling matrix; The cloud amount ratio parameter is calculated by counting the ratio of the number of pixels in the effective cloud area to the total number of pixels according to the effective cloud area pixel labeling matrix; According to the cloud category label, the preset cloud category coding table is matched, and the candidate area is categorized and coded to obtain the cloud category parameters; Extracting grayscale image data from the cloud layer area corresponding to the pixel-level segmentation mask, calculating contrast, correlation and energy values based on the grayscale co-occurrence matrix, and generating texture morphological feature parameters using a local binary pattern algorithm; The texture morphological feature parameters are input into a pre-trained height regression model, and a nonlinear mapping is performed on the texture morphological feature and the height correlation relationship through a multi-layer perception network to output a vertical height parameter; The cloud amount ratio parameter, cloud category parameter, texture morphology characteristic parameter and vertical height parameter are integrated to construct a cloud characteristic parameter set.

[0017] By adopting the above technical solutions, multi-dimensional features are integrated to construct a cloud feature parameter set, which provides a scientific basis for decision-making for weather modification (human shadow) operations, making the decision more accurate and efficient. The construction of this data set not only improves the accuracy of cloud feature extraction, but also provides important data support for meteorological research and disaster warning.

[0018] In the second aspect, the present application provides a human shadow operation target cloud recognition system, which adopts the following technical solution: A human shadow operation target cloud layer recognition system, the recognition system comprising: An acquisition module is used to acquire original cloud image data of a target cloud area; A preprocessing module for preprocessing the original cloud image data to generate a standardized feature image; A model construction module for pre - constructing a deep convolutional neural network architecture; A feature extraction module for performing multi - level feature extraction on the standardized feature image based on a feature extraction network to generate a feature map; A bounding box generation module for traversing the feature map based on a region proposal network to generate an initial detection result including candidate region bounding boxes; A candidate region determination module for performing feature alignment processing on the candidate region bounding boxes based on the region of interest pooling algorithm and outputting the feature vectors of the candidate regions; A classification and regression module for inputting the feature vectors of the candidate regions into a classification and regression network, performing multi - class classification prediction and bounding box position regression calculation, and outputting a set of detection results including cloud class labels and target cloud bounding boxes; A convolutional segmentation module for performing full - convolutional segmentation operations on each candidate region in the set of detection results to generate pixel - level segmentation masks; A cloud parameter set construction module for calculating the cloud amount ratio, cloud class, texture morphology features, and vertical height parameters of the target cloud region according to the pixel - level segmentation masks and cloud class labels, and constructing a cloud feature parameter set.

[0019] In a third aspect, the present application provides a computer device, adopting the following technical solution: A computer device includes a memory, a processor, and a computer program stored on the memory. The processor executes the computer program to implement the steps of the method as described in the first aspect.

[0020] In a fourth aspect, the present application provides a computer - readable storage medium, adopting the following technical solution: A computer - readable storage medium stores a computer program that can be loaded and executed by a processor to perform any of the methods in the first aspect.

[0021] In summary, the present application includes at least one of the following beneficial technical effects: It not only improves the accuracy of cloud recognition, but also significantly enhances the scientific nature and effectiveness of weather modification operations through multi - dimensional analysis, real - time feedback, and data - driven decision - making support, providing an innovative solution for applications in related fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is the first flow diagram of the method for identifying the target cloud in weather modification operations in one embodiment of the present application.

[0023] Figure 2It is the second process schematic diagram of the cloud identification method for cloud seeding operation target in one embodiment of the present application.

[0024] Figure 3 It is the third process schematic diagram of the cloud identification method for cloud seeding operation target in one embodiment of the present application.

[0025] Figure 4 It is the fourth process schematic diagram of the cloud identification method for cloud seeding operation target in one embodiment of the present application.

[0026] Figure 5 It is the fifth process schematic diagram of the cloud identification method for cloud seeding operation target in one embodiment of the present application.

[0027] Figure 6 It is the sixth process schematic diagram of the cloud identification method for cloud seeding operation target in one embodiment of the present application. Detailed implementation manners

[0028] In order to make the objectives, technical solutions and advantages of the present application clearer, the following further describes the present application in detail with reference to the appended Figures 1-6 drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0029] An embodiment of the present application discloses a cloud identification method for cloud seeding operation target.

[0030] Referring to Figure 1 , a cloud identification method for cloud seeding operation target, the identification method includes: Step S101, obtaining the original cloud image data of the target cloud area; Among them, the cloud image data includes the optical characteristics in the visible light band. The visible light band image is the most direct optical manifestation of the cloud, and can capture the shape, position, and meteorological conditions of the cloud.

[0031] Specifically, in order to accurately identify the cloud seeding operation target cloud, a high-resolution visible light camera is used to obtain high-quality cloud images to ensure that the captured cloud details are clear enough. At the same time, the camera has a sensor with good dynamic range and low noise performance to cope with use under different lighting conditions.

[0032] In the actual application process, the high-resolution visible light camera is installed at a high point in the cloud seeding operation point area to ensure that its field of view covers the target cloud area. The angle and focal length of the camera can be controlled and adjusted to enable it to clearly capture the fine features of the cloud; at the same time, appropriate parameters such as exposure time, ISO sensitivity, and white balance are set to optimize the image quality and ensure that the camera can work stably under various weather conditions, avoiding image quality degradation caused by environmental changes.

[0033] Step S102, preprocess the original cloud image data to generate a standardized feature image; Among them, the preprocessing steps include but are not limited to eliminating image noise, enhancing contrast, adjusting the image size, normalizing pixel values, and so on.

[0034] Specifically, in addition to using conventional noise processing and image enhancement, etc., the data preprocessing also performs resizing and normalization processing on the cloud layer of the human shadow target. First, use noise elimination techniques such as Gaussian blur for spatial domain denoising to reduce image noise; then enhance the contrast of the clouds in the image by means of histogram equalization. In order to meet the input size requirements of the neural network, the image will be adjusted to a specific size, and at the same time, the pixel values will be normalized so that they are distributed in a smaller range, which helps to improve the stability and convergence speed of model training.

[0035] It can be understood that during the process of changing the image size, the transformation relationship (such as scaling ratio, cropping position, etc.) between the original image and the adjusted image will also be recorded, and this information is saved in the im_info variable for use in subsequent steps.

[0036] Step S103, pre-construct a deep convolutional neural network architecture, and perform multi-level feature extraction on the standardized feature image based on the feature extraction network to generate a feature map; Among them, the standardized image will be fed into a pre-trained deep convolutional neural network, which has been learned through a large-scale image dataset. Through the multi-level feature extraction network structure of the deep convolutional neural network (Faster R-CNN), meaningful feature maps can be effectively extracted from the cloud image, and these feature maps will serve as the basis for the region proposal network to generate candidate region bounding boxes.

[0037] In one embodiment of the present application, the feature extraction network adopts ResNet50. As a deep residual network, it can extract the low-level to high-level features of the image layer by layer through its multi-layer convolutional and pooling layers. ResNet50 adopts residual modules to solve the problem of gradient disappearance in the training of deep networks, ensuring the smooth propagation of gradients and enabling the network to train deeper models. After the image passes through these convolutional operations, the network represents the extracted feature information in a multi-level manner, and finally generates a feature map containing rich semantic information and spatial information. The dimension of this feature map is usually a compressed low-dimensional representation, with strong abstraction ability and retaining the spatial structure features of the input image.

[0038] Step S104, traverse the feature map based on the region proposal network to generate an initial detection result including candidate region bounding boxes; Among them, the Region Proposal Network (RPN) is one of the core components in the deep convolutional neural network architecture (Faster R-CNN). Its goal is to generate candidate regions in the image that may contain the target object (clouds in the embodiments of this application). The bounding boxes corresponding to these candidate regions are called "candidate region bounding boxes".

[0039] In this step, the working principle of the Region Proposal Network (RPN) is to traverse the feature map in small blocks through a sliding window mechanism and generate multiple anchor boxes with different sizes and aspect ratios at each position. Each anchor box corresponds to a local area in the image, and by calculating the score of each area, it is determined whether it contains clouds (targets). The RPN will calculate the score of each anchor box and generate a set of candidate regions. Through this method, the RPN can efficiently identify the regions that may contain clouds and provide preliminary region localization for subsequent classification and regression.

[0040] It can be understood that by sliding on the feature map and generating anchor boxes, the RPN enables the system to quickly and efficiently determine potential cloud regions. This process provides accurate anchor boxes for subsequent object detection and classification, while significantly improving the detection speed. The RPN can adapt to clouds of various sizes and shapes, ensuring that the generated anchor boxes can accurately cover the positions where the clouds are located, and is a key link in the entire system framework.

[0041] Step S105: Perform feature alignment processing on the candidate region bounding boxes based on the Region of Interest (RoI) Pooling algorithm, and output the feature vectors of the candidate regions; Among them, the Region of Interest (RoI) Pooling solves the problem of how to map candidate region bounding boxes of different sizes to a feature map of a fixed size. In traditional convolutional neural networks, the size of the input image is usually fixed, but the sizes of candidate region bounding boxes vary. Directly feeding candidate region bounding boxes of different sizes into the fully connected layer for classification and regression will result in information loss. RoI Pooling maps each candidate region bounding box to the feature map and performs a pooling operation on it, unifying all candidate region bounding boxes into an output of a fixed size, thereby ensuring that regardless of the change in the size of the candidate region bounding box, the dimension of the output feature is always the same. These unified feature vectors will then be fed into the classifier and regressor for further processing.

[0042] It can be understood that the introduction of RoI Pooling enables the system to efficiently process information from different candidate regions and convert it into fixed-length feature vectors. This feature alignment operation not only improves the performance of the classifier but also enables the network to operate stably in multi-scale situations, thereby enhancing the efficiency and accuracy of object detection. Through RoIPooling, the network avoids problems caused by inconsistent input sizes and ensures the quality of feature extraction.

[0043] Step S106: Input the feature vectors of the candidate regions into the classification and regression network, perform multi-class classification prediction and bounding box position regression calculation, and output a set of detection results including cloud class labels and target cloud bounding boxes. Among them, multi-class classification prediction and bounding box regression are performed in the classification and regression network. First, for the classification task, by classifying the feature vectors of the candidate regions, it is determined which cloud class (such as cumulus cloud, stratus cloud, etc.) the region belongs to. The classifier will output the class label to which each candidate region bounding box belongs and assign a probability value to each class. Second, for the regression task, the position of each candidate region bounding box is fine-tuned. By calculating the difference between the actual bounding box and the predicted box, its position is optimized, and this process is achieved through bounding box regression, so as to obtain a more accurate positioning of the target cloud bounding box.

[0044] It can be understood that outputting the class label and position accuracy of the candidate region bounding boxes simultaneously effectively optimizes the classification information and positioning information of each candidate region bounding box. The classification and regression network improves the accuracy of object detection through the joint classification and regression tasks, ensuring the reliability and accuracy of the detection results. Through fine regression adjustment, the system can accurately fit the target bounding box and improve the spatial accuracy of the detection results.

[0045] Step S107: Perform full convolutional segmentation operation on each candidate region in the set of detection results to generate a pixel-level segmentation mask. Among them, the segmentation mask contains binary marking information indicating whether each pixel belongs to a cloud or not. In the embodiment of the present application, the full convolutional segmentation operation (FCN) is an efficient pixel-level image segmentation method. Different from traditional object detection methods, image segmentation not only focuses on the bounding box or class label of an object but requires each pixel to be classified into a specific class. The core purpose of this step is to perform a convolution operation on each candidate region bounding box based on the feature information of the candidate region, thereby generating a fine-grained pixel-level segmentation mask. In this way, each pixel in the image can be accurately classified as belonging to a cloud, background, or other classes, and further provide more detailed data for subsequent cloud feature extraction (such as cloud amount calculation, texture analysis, etc.).

[0046] Specifically, the FCN first performs convolutional feature extraction on the image and restores the spatial resolution of the image through a deconvolution layer or upsampling operation to ensure that the size of the final output image is the same as that of the input image. This segmentation mask represents the probability value of each pixel belonging to the cloud layer. To obtain an accurate segmentation result, the FCN classifies each pixel in the image through pixel-by-pixel prediction based on the previously trained model, thereby generating a segmentation map containing the target cloud area. These segmentation maps not only provide the accurate boundaries of the clouds but also reflect the complex texture and morphological characteristics of the clouds.

[0047] Step S108: According to the pixel-level segmentation mask and the cloud layer category label, calculate the cloud amount ratio, cloud layer category, texture morphological characteristics, and vertical height parameters of the target cloud area respectively, and construct a cloud layer feature parameter set.

[0048] Among them, the cloud layer feature parameter set includes a cloud amount distribution matrix, a cloud layer category coding matrix, a texture feature vector, and a height scalar value, which are used as the basic data source for the auxiliary decision-making of cloud seeding operations.

[0049] Specifically, the calculation of the cloud amount ratio is based on the pixel segmentation of the target cloud area to calculate the proportion of the cloud layer in the image; the cloud layer category is obtained through the cloud layer category label, the texture morphological characteristics are extracted through the texture analysis of the cloud layer image, and the vertical height of the cloud layer is estimated by combining texture information and optical characteristics. These parameters constitute a detailed cloud layer feature parameter set and are provided to the operators as a decision-making basis.

[0050] It can be understood that various cloud layer feature parameters are integrated in matrix form to form a cloud layer feature parameter matrix to provide comprehensive cloud layer feature data, including not only the quantitative data of the cloud amount but also key features such as cloud layer category, texture, and cloud height. These data are of great significance for artificial weather modification (cloud seeding) operations, can provide scientific support for decision-making, and help operators better judge whether to perform operations such as cloud catalysis.

[0051] In the above embodiment, by combining the high-resolution visible light imaging system with the deep convolutional neural network, on the basis of realizing multi-scale feature extraction of cloud layer images, through the collaborative optimization of the Region Proposal Network (RPN) and the Fully Convolutional Network (FCN), a multi-dimensional parameter system including a pixel-level segmentation mask, a cloud class coding matrix, and a morphological feature vector is constructed, effectively solving the technical problems such as low efficiency existing in traditional observation methods when facing complex backgrounds. It not only realizes the automatic and accurate analysis of cloud amount distribution, cloud layer category, and texture characteristics but also can provide multi-dimensional decision-making support for artificial weather modification operations, including the judgment of catalysis timing and the determination of operation scope, through the real-time generated cloud layer feature parameter set, significantly improving the scientificity and timeliness of cloud seeding operations.

[0052] Reference Figure 2 , as an implementation of step S104, the step of generating an initial detection result including candidate region bounding boxes based on traversing the feature map by the region proposal network includes: Step S201, applying a convolutional kernel on the feature map for sliding window scanning to generate an intermediate feature map; Among them, the convolutional kernel scans on the feature map through a sliding window operation. The convolution operation is a core step in the convolutional neural network. It processes the input image through the convolutional kernel to extract local features. In the RPN (Region Proposal Network), the convolutional kernel generates an intermediate feature map by sliding and scanning the feature map. This intermediate feature map contains visual information at each spatial position and has a spatial structure consistent with the size of the input image. Specifically, the size of the convolutional kernel is usually 3×3, and the convolution operation is performed on the entire feature map in a sliding manner. Each convolution generates a local feature map, which reflects the feature information of a specific region, and finally forms an intermediate feature map containing spatial information and semantic information.

[0053] Step S202, performing a two-branch convolution operation on the intermediate feature map to generate an anchor box classification confidence tensor and an anchor box position offset tensor for each spatial position respectively; Specifically, the two-branch convolution operation processes the intermediate feature map through two independent convolutional branches. The value in the anchor box classification confidence tensor indicates whether the target is included in the region, and the anchor box position offset tensor provides details on how the anchor box should be offset.

[0054] In the embodiment of the present application, the first branch generates an anchor box classification confidence tensor through a 1×1 convolutional kernel, which is used to represent the probability that each anchor box contains a target; the second branch also generates an anchor box position offset tensor through a 1×1 convolution, which is used to adjust the center coordinates and its width and height values of the anchor box. The outputs of these two tensors determine the classification information and position adjustment information of each candidate region.

[0055] It can be understood that this two-branch convolution operation can effectively provide necessary classification confidence and position regression information for each anchor box, and provide accurate anchor box positions and class predictions for subsequent object detection. Through this process, the system can accurately adjust the position of each anchor box and predict whether it contains a target, thereby providing a preliminary result for object detection.

[0056] Step S203, generating a corresponding anchor box group at each spatial position of the feature map based on a preset reference anchor box template; Among them, based on a preset multiple anchor box templates, the system will generate corresponding anchor box groups at each spatial position of the feature map. Each anchor box template corresponds to different sizes and aspect ratios to adapt to different sizes and shapes of targets that may exist in the image. The design of the anchor box templates takes into account the typical physical size range of the target cloud layer and ensures that the diversity of cloud layer targets can be covered. By the method of generating anchor box groups through a sliding window, the system can cover all potential target areas in the image and provide candidate region bounding boxes for subsequent classification and regression tasks.

[0057] It should be noted that through the generation strategy based on anchor box templates, the system can generate anchor boxes of various sizes for targets of different sizes and shapes. This multi-scale anchor box generation method enhances the robustness of the model, enabling it to adapt to various images and target categories. The generation of anchor box groups ensures that the candidate regions can cover all target areas in the image, thus providing an accurate starting point for subsequent object detection and position regression.

[0058] Step S204, based on the anchor box classification confidence tensor and the anchor box position offset tensor, calculate the initial bounding box coordinates of the candidate regions based on the anchor box group; Among them, by calculating the anchor box classification confidence tensor and the anchor box position offset tensor, the system can calculate the initial bounding box coordinates of each candidate region based on the anchor box group. The classification confidence tensor provides the probability that each anchor box contains a target, while the position offset tensor provides the offset of the anchor box relative to the true target bounding box. Combining these two pieces of information helps the system generate the initial bounding box of the candidate region from the anchor box group.

[0059] Step S205, according to the initial bounding box coordinates of the candidate regions, calculate the intersection over union (IoU) between the initial bounding box of the candidate region and the true cloud annotation box, and filter and retain the anchor boxes in the anchor box group that meet the preset IoU threshold; Specifically, the system screens the candidate regions by calculating the intersection over union (IoU) between the initial bounding box of the candidate region and the true cloud annotation box. The IoU is an index to measure the degree of overlap. By setting a predetermined threshold, only the anchor boxes with an IoU higher than this threshold will be considered as valid positive samples, while the anchor boxes with an IoU lower than another threshold will be regarded as negative samples. Through this screening mechanism, the system can remove those redundant boxes that cannot represent the true target.

[0060] Step S206, perform spatial correction on the filtered anchor box group to obtain an effective anchor box group; Among them, by performing spatial correction on the filtered anchor box group, it is ensured that the coordinates of these anchor boxes match the true target boxes more closely. By fine-tuning the anchor box group, the system can further optimize the position and size of the initial bounding box of the candidate region to make it more conform to the boundary of the actual target.

[0061] Step S207: Iteratively process the set of valid anchor boxes based on the non-maximum suppression algorithm, and output the optimized candidate region bounding boxes as the initial detection results.

[0062] Among them, non-maximum suppression (NMS) is a classic algorithm in object detection, which is used to remove redundant parts in the set of valid anchor boxes. In this step, NMS calculates the intersection over union (IoU) of the anchor boxes, removes those boxes that overlap too much with other anchor boxes, and finally retains those most representative anchor boxes. NMS selects the anchor box with the highest classification confidence and deletes other boxes with too high overlap with it, and outputs the final detection results.

[0063] It should be noted that the set of anchor boxes is a collection composed of multiple anchor boxes, which is used to generate possible target regions; the candidate regions are the target regions that are generated in the region proposal network and have been screened and regression-adjusted; and the candidate region bounding boxes are the finally optimized bounding boxes that represent the accurate positions of the target objects.

[0064] In the above implementation manner, by combining the region proposal network (RPN) and the object detection framework, the accurate detection of cloud targets can be efficiently realized. Through steps such as multi-level feature extraction, anchor box regression, and non-maximum suppression, the target clouds in the image can be effectively identified, and then real-time and accurate data support can be provided for artificial weather modification operations.

[0065] Refer to Figure 3 , as an implementation manner of step S105, the steps of performing feature alignment processing on the candidate region bounding boxes based on the region of interest pooling algorithm and outputting the feature vectors of the candidate regions include: Step S301: Map the candidate regions to the corresponding spatial positions on the feature map based on the scaling ratio parameter between the original cloud image data and the feature map; Among them, in order to ensure that the candidate regions can be uniformly processed on the feature map, it is necessary to accurately map these candidate regions to the feature map according to the scaling ratio of the image. This process adjusts the coordinates of the candidate boxes (such as the x and y coordinates of the upper left corner and the width w and height h) according to the scaling ratio of the feature map, and the scaling ratio is usually calculated based on the relationship between the original image and the feature map.

[0066] Step S302: Uniformly divide the mapped candidate regions into grid sub-regions with K rows and K columns in the feature map space; where K is a preset positive integer; Specifically, each mapped candidate region is divided into several small grid sub-regions, which can enable candidate regions of different sizes to have the same processing method. In this step, each candidate region is divided into a K×K grid (such as 14x14), where K is a preset positive integer. This operation is to ensure the uniformity of the pooling operation. The features within each grid will be processed and extracted separately, and finally a feature representation of a fixed size is formed.

[0067] Step S303: Perform a max pooling operation on each grid sub-region, extract the max pooling result of the features of each grid sub-region, and splice them in the spatial order to obtain the feature vector of the candidate region.

[0068] Among them, performing a max pooling operation within each grid sub-region means selecting the maximum value within the region as the representative feature of the grid. The max pooling operation can retain the most significant features within the region and avoid the interference of smaller values. Then, these max pooling results are spliced in the spatial order (i.e., in the order from left to right and from top to bottom), and finally the feature vector of the entire candidate region is obtained.

[0069] In the above implementation, the steps of performing feature alignment processing on candidate regions based on the Region of Interest (ROI) Pooling algorithm can effectively convert candidate regions of different sizes into feature representations of a unified size. This step solution can not only handle the processing problems of candidate regions of different sizes, but also extract the most representative features through max pooling, thereby improving the performance and efficiency of the object detection task.

[0070] Refer to Figure 4 , as an implementation of step S103, the steps of pre-constructing a deep convolutional neural network architecture include: Step S401: Obtain the original image sample set of meteorological clouds and perform normalization processing, and divide it into a training set and a validation set; Among them, the sample set includes multiple original meteorological cloud images and corresponding annotation information, and the annotation information includes the target cloud category label and the bounding box coordinates; Specifically, a large number of cloud images are obtained from meteorological satellites or other acquisition channels. These images should cover various cloud categories and weather conditions to ensure the diversity and representativeness of the data. For each image, in addition to the image itself, the category label of the target cloud and its corresponding bounding box coordinates (annotation information) need to be provided. The bounding box coordinates define the position and size of the cloud in the image and are indispensable for training the object detection model.

[0071] Furthermore, to ensure that each image can enter the network for processing and the computational efficiency during the processing is not affected by input size differences. First, the image sizes are unified. The shorter side of the image is adjusted to a fixed size (such as 600 pixels), and then the longer side is adjusted accordingly according to the aspect ratio of the image, so that the longer side of the image does not exceed a specified maximum value (such as 1000 pixels). Then, the pixel values of the image are normalized. Each pixel value of the image is subtracted by the mean vector of the ImageNet dataset and divided by the standard deviation vector, so as to map the pixel values of all images to a standard range (usually between 0 and 1).

[0072] Next, after the image normalization is completed, 70% of the data is used as the training set, and the remaining 30% is used for the validation set. The training set is used to train the model, and the validation set is used to monitor and optimize the performance during the model training process to avoid overfitting and ensure the generalization ability of the model on unknown data.

[0073] Step S402, construct a three-level network architecture including a feature extraction network, a region proposal network, and a classification and regression network; In some embodiments, ResNet50 can be used as the feature extraction network. ResNet50 is a convolutional neural network based on the residual network (ResNet). It alleviates the vanishing gradient problem in deep networks by introducing skip connections. ResNet50 consists of five residual modules (conv1 to conv5), and each module contains multiple residual blocks. These residual blocks add the input and output through skip connections, thereby effectively preventing the disappearance of gradients during backpropagation, enabling the network to be trained deeper and capable of learning more complex features.

[0074] Moreover, the output of ResNet50 is a multi-level feature map, which contains different levels of semantic information of the input image. In object detection tasks, usually a certain middle-level feature map (such as conv4_x) is selected as the shared feature map because this layer of feature map usually retains both spatial details (such as object edges, textures, etc.) and higher-level semantic information (such as object categories, shapes, etc.), so it is suitable for both generating candidate regions for object detection tasks and providing rich features for subsequent classification and regression tasks.

[0075] For the Region Proposal Network (RPN), the RPN receives the shared feature map as input, slides a small convolutional window over the feature map, and generates multiple anchor boxes with different sizes and aspect ratios. These anchor boxes are used to cover different target regions in the image. The output of the RPN includes two branches: a classification branch (to determine whether the anchor box is foreground or background) and a regression branch (to predict the position offset of the anchor box). Through these candidate regions, the RPN can provide high-quality candidate regions for the subsequent classification and regression network.

[0076] The classification and regression network further processes the candidate regions generated by the RPN, and performs fine-grained classification and bounding box regression on each candidate region. The classification and regression network usually contains several fully connected layers and a multi-task prediction head. The multi-task prediction head outputs the class probability and the corrected offset of the bounding box simultaneously, so as to achieve accurate object localization and classification.

[0077] It can be understood that through this three-level architecture design, the model can simultaneously generate candidate regions and classify and locate subsequent objects, so as to achieve end-to-end object detection tasks.

[0078] Step S403: Input the training set into the three-level network architecture, generate the corresponding shared feature map based on the feature extraction network, and perform a sliding window convolution operation based on the Region Proposal Network to generate a multi-scale anchor box map; Specifically, the training set images are input into the deep network architecture. First, a shared feature map is generated through the feature extraction network (such as ResNet50). The feature extraction network extracts different levels of features from the original image through multiple convolutional operations, forming a high-dimensional feature map. These feature maps contain information such as edges, textures, and object shapes in the image for subsequent networks to use.

[0079] Furthermore, the Region Proposal Network (RPN) receives the shared feature map as input and uses a sliding window convolutional layer to slide over the feature map to generate anchor boxes. At each feature map position, the RPN generates multiple anchor boxes with different scales and aspect ratios to cover possible target regions in the image. A common anchor box generation strategy is to generate 9 anchor boxes for each position, corresponding to 3 scales and 3 aspect ratios respectively. The anchor boxes can cover targets of different sizes and shapes, thereby improving the accuracy of object detection.

[0080] Step S404: By calculating the intersection over union of the anchor box and the annotated bounding box, filter positive and negative samples and construct binary classification labels; Specifically, the intersection over union (IoU) between each anchor box and the labeled bounding box is calculated. IoU is an indicator for measuring the overlapping degree of two boxes, and its value ranges from 0 to 1. The larger the value, the more the overlap. When IoU ≥ 0.7, the anchor box is regarded as a positive sample; when IoU ≤ 0.3, the anchor box is regarded as a negative sample; and the anchor boxes in between are usually regarded as ignored samples or boundary case samples.

[0081] Step S405: Based on the binary classification labels, optimize the classification branch of the region proposal network through the binary cross-entropy loss function, and optimize the regression branch of the region proposal network through the smooth L1 loss function; Among them, the classification loss uses the binary cross-entropy loss function, which is used to optimize the foreground and background classification of the anchor boxes. The cross-entropy loss is a commonly used loss function in classification problems and can effectively evaluate the difference between the predicted class distribution and the true label.

[0082] The regression loss uses the smooth L1 loss function, which is mainly used to optimize the coordinate regression of the anchor boxes. The smooth L1 loss is a function that combines the L1 and L2 losses. It can handle small errors and effectively avoid the influence of large errors, and is more stable than the pure L2 loss in object detection tasks. This loss function helps to accurately predict the position of the target bounding box, thereby improving the localization accuracy of the model.

[0083] Step S406: Map the candidate regions generated by the region proposal network to the shared feature map, and perform the region of interest (RoI) pooling operation to generate a fixed-size feature vector; Among them, the generated candidate regions are further processed through the RoI Pooling (region of interest pooling) operation. The role of RoI Pooling is to align candidate regions of different sizes to a fixed-size feature map. This is achieved by dividing each candidate region into several small grids and performing a pooling operation (usually max pooling) on each grid to extract the most significant features and reduce the dimension, so that the size of the finally output feature vector is unified. RoI Pooling ensures that regardless of the size of the input region, the size of the output feature vector is fixed, usually set to a 7×7 grid.

[0084] Step S407: Input the fixed-size feature vector into the fully connected layer of the classification regression network for dimensionality reduction processing, and output a dimensionality-reduced feature matrix; Among them, the fixed-size feature vector obtained through RoI Pooling is input into the fully connected layer for dimensionality reduction processing. The role of the fully connected layer is to convert the high-dimensional feature vector into a smaller and more refined representation. This step usually reduces the 7×7×1024 feature vector to a smaller dimension (such as 4096 dimensions) to provide a simplified input for the subsequent multi-task prediction head.

[0085] Step S408: Perform multi-task prediction based on the dimensionality-reduced feature matrix, and calculate the class probability distribution and bounding box offsets. Specifically, on the dimensionality-reduced feature vectors, the network performs multi-task prediction, outputting the class probability distribution (e.g., cloud class) for each candidate region and the offsets (dx, dy, dw, dh) of the bounding boxes. The class probability distribution is output through the Softmax function, while the offsets of the bounding boxes are predicted through a regression network.

[0086] Step S409: Obtain the labeled target cloud class labels and bounding box coordinates in the training set, calculate the classification loss based on the class probability distribution and the target cloud class labels, and calculate the regression loss based on the bounding box offsets and the bounding box coordinates. Specifically, the network is optimized by calculating the cross-entropy loss between the class probability distribution predicted by the model and the true class labels, and the regression loss between the predicted bounding box and the true bounding box. The calculation of these two loss terms helps the network continuously correct the prediction results during training.

[0087] Step S410: Jointly optimize the classification loss and the regression loss through the backpropagation algorithm, and update the weight parameters of the classification and regression network. Specifically, through the backpropagation algorithm, the weights of the classification and regression network are updated according to the calculated losses. The backpropagation algorithm can optimize the network parameters through gradient descent, enabling the model to gradually learn better feature representations during training.

[0088] Step S411: Freeze the weight parameters of the feature extraction network, jointly train the region proposal network and the classification and regression network, and generate a primary training model. Specifically, in the initial stage of model training, freeze the weights of the feature extraction network such as ResNet50, and focus on training the RPN and the classification and regression network to quickly optimize the candidate region generation and classification and regression tasks.

[0089] Step S412: Unfreeze the weight parameters of the feature extraction network, perform global parameter fine-tuning based on the primary training model, and update the weight parameters of the feature extraction network, the region proposal network, and the classification and regression network. Specifically, after the initial training is completed, unfreeze the weights of the feature extraction network and perform global fine-tuning. This step helps the network as a whole learn more suitable feature representations for object detection.

[0090] Step S413: Validate the primary training model based on the validation set until the total loss function meets the preset conditions or the model iteration times reach the preset number, and output the constructed deep convolutional neural network architecture.

[0091] Specifically, the trained model is verified through a validation set. Once the total loss function satisfies the stopping condition or reaches the maximum number of iterations, the training is terminated, and the final model is output. Among them, the total loss function is a weighted sum of the regression loss and the classification loss. The specific formula is: ; In the above formula, α and β are weight coefficients, reg_loss is the classification loss function, cls_loss is the regression loss function, and the total loss function total_loss is used to balance the importance of the regression loss and the classification loss. In this way, both the regression and classification tasks can be optimized simultaneously, thus achieving better performance in the object detection task.

[0092] In the above embodiment, through multi-level feature extraction, accurate generation and screening of candidate regions, region of interest pooling alignment, and multi-task learning, the classification accuracy and bounding box localization accuracy of the model are significantly improved. By gradually optimizing the classification and regression losses, the model can efficiently process cloud targets of different sizes and shapes, and finally realizes an end-to-end deep learning model with high accuracy, robustness, and generalization ability in the cloud detection task.

[0093] Referring to Figure 5 , as an embodiment of step S107, the steps of performing a full convolutional segmentation operation on each candidate region in the detection result set to generate a pixel-level segmentation mask include: Step S501, input the feature vector of the candidate region into the segmentation sub-network of the fully convolutional neural network, and perform upsampling processing on the feature vector of the candidate region through at least three deconvolution layers to generate an initial segmentation map; Among them, after the candidate region feature vector undergoes previous processing (such as the extraction of candidate boxes and feature maps generated by RPN), it is input into the segmentation sub-network of the fully convolutional neural network (FCN). The deconvolution layer, also known as the upsampling layer, restores the spatial resolution in the image by gradually performing spatial upsampling on the feature map. It uses a convolutional kernel and a stride to perform upsampling on the feature map, gradually restoring the high-resolution details of the input image. Here, the spatial resolution of the feature map is gradually increased through three deconvolution layers, and finally an initial segmentation map is generated. This segmentation map has a lower resolution but contains preliminary segmentation information of the cloud region. The principle of deconvolution is based on the inverse process of the convolutional layer, and the image size can be restored through the padding process while retaining edge and semantic information. The key in this stage is to convert the low-resolution candidate region feature map into an initial segmentation map with more spatial details through the deconvolution layer.

[0094] Exemplarily, the segmentation sub-network of the fully convolutional neural network includes: a first transposed convolutional layer that upsamples the input feature map using a 4×4 convolutional kernel and a padding method with a stride of 2; a second transposed convolutional layer that refines the features of the output of the first transposed convolutional layer using a 3×3 convolutional kernel and a padding method with a stride of 1; and a third transposed convolutional layer that restores the feature map to the resolution of the original input image using a 2×2 convolutional kernel and a padding method with a stride of 2.

[0095] Step S502: Perform a per-pixel classification operation on the initial segmentation map to generate a probability distribution map containing the probability values of each pixel belonging to the target cloud layer. Among them, after the initial segmentation map is processed by the transposed convolutional layer, a higher-resolution feature map is obtained. Next, each pixel needs to be classified. At this stage, the network predicts the position of each pixel one by one through a classification branch to generate a class probability distribution map. Specifically, the network will correspond the position of each pixel to the target cloud layer category, predict whether the pixel belongs to the cloud layer (foreground) or non-cloud layer (background), and calculate the probability of each pixel position belonging to each category.

[0096] In the embodiment of the present application, first, the initial segmentation map is fed into a convolutional layer (usually a 3×3 convolution) to adjust the number of channels of the image to the number of cloud layer categories plus a background category (for example, if there are only two types of cloud layers, then the number of channels will be 3); then the Softmax function is applied to each pixel to calculate the probability of the pixel belonging to each category. The Softmax function is used to convert the score of each pixel into a probability value, so that each pixel point will have a class probability distribution.

[0097] Exemplarily, assume that there are three cloud layer categories in the image. After convolution and Softmax operations, each pixel point in the image will obtain a three-dimensional probability distribution. For example, the probability that a certain pixel belongs to category 1 (cumulus cloud) is 0.7, the probability that it belongs to category 2 (cirrus cloud) is 0.2, and the probability that it belongs to the background category is 0.1.

[0098] Step S503: Perform binarization processing on the probability distribution map based on a preset probability threshold, and output a pixel-level segmentation mask containing cloud pixel marking information.

[0099] Among them, after per-pixel classification and obtaining the class probability distribution of each pixel, the next step is to convert these probabilities into specific segmentation results. This is usually done through binarization processing with a threshold (such as 0.5), that is, comparing the probability value of each pixel with the threshold. If the probability value of the pixel point is greater than or equal to the set threshold, it is considered that the pixel belongs to the target cloud layer (i.e., marked as 1), otherwise it is considered that the pixel does not belong to the cloud layer (i.e., marked as 0).

[0100] It should be noted that in some practical applications, the threshold may need to be dynamically adjusted. Especially when the transparency of the cloud image is low or there are thin clouds, the threshold may need to be reduced (such as 0.3) to adapt to different cloud conditions. The result of binarization will obtain a binary segmentation mask, which identifies which pixels belong to the cloud region.

[0101] In the above embodiments, the pixel-level segmentation of the target cloud layer is accurately achieved. Through the upsampling process of the deconvolution layer, the spatial details of the candidate region feature map are restored, providing a high-resolution input for subsequent per-pixel classification. The probability distribution map generated by the per-pixel classification operation can accurately predict the probability that each pixel belongs to the cloud layer, and the binarization process based on the threshold effectively converts the probability values into a clear cloud layer segmentation mask, ensuring the accurate identification and marking of the cloud layer region. The whole process not only improves the accuracy of cloud layer segmentation, but also makes full use of the advantages of deep learning technology to provide an efficient and accurate solution when dealing with complex meteorological images.

[0102] Referring to Figure 6 , as an embodiment of step S108, according to the pixel-level segmentation mask and the cloud layer category label, the steps of calculating the cloud amount ratio, cloud layer category, texture morphology characteristics, and vertical height parameters of the target cloud layer region respectively, and constructing a cloud layer feature parameter set include: Step S601, perform binarization processing based on the pixel-level segmentation mask to generate a valid cloud region pixel marking matrix; Among them, after the pixel-level segmentation mask of the target cloud layer is generated, the next step is to clearly distinguish the cloud layer region from the non-cloud layer region. Binarization processing is to convert the pixel values in the image into two possible values by setting a threshold.

[0103] In this step, through binarization processing, the class probability of each pixel in the segmentation mask is converted into 0 or 1, where 1 indicates that the pixel belongs to the cloud layer and 0 indicates that the pixel does not belong to the cloud layer. Usually, a preset probability threshold (such as 0.5) is used. If the class probability of a certain pixel is greater than or equal to the threshold, it means that the pixel belongs to the cloud layer; otherwise, it is considered a non-cloud layer pixel. The purpose of this binarization processing is to clearly distinguish the cloud layer region from the non-cloud layer region in the image and generate a pixel marking matrix containing the valid cloud layer region. This matrix serves as the basic data for cloud amount calculation and cloud layer category analysis in subsequent steps.

[0104] Step S602, calculate the ratio of the number of valid cloud region pixels to the total number of pixels according to the valid cloud region pixel marking matrix, and calculate the cloud amount ratio parameter; Among them, the cloud cover ratio is an important parameter in meteorology to measure the degree of cloud cover. By counting the number of pixels marked as 1 in the effective cloud area pixel marking matrix (i.e., the number of pixels in the cloud area N_cloud), and the total number of pixels in the entire image N_total, the cloud cover ratio can be calculated.

[0105] Specifically, the calculation formula for the cloud cover ratio is to divide the number of pixels in the cloud area by the total number of pixels in the image, obtaining a ratio representing the cloud cover degree. This ratio can provide a quantitative basis for weather forecasting and cloud analysis. Especially when conducting weather modification operations, the cloud cover degree directly affects the selection and seeding strategy of cloud catalysts. The value of the cloud cover ratio is usually expressed as a percentage. For example, a cloud cover ratio of 60% means that 60% of the area in the image is covered by clouds.

[0106] Step S603, match the preset cloud category coding table according to the cloud category label, and perform category coding on the candidate area to obtain the cloud category parameter; Among them, after the cloud area segmentation and cloud amount calculation are completed, the next step is to classify the clouds. Specifically, each cloud area is marked with a different cloud category. Cloud categories include fair-weather cumulus, towering cumulus, stratus cloud, cumulonimbus cloud, etc. Each cloud category has different meteorological characteristics. By matching the cloud category label with the preset cloud category coding table, the category information of each cloud area can be converted into a standardized coding form.

[0107] For example, assume that in the preset coding table, fair-weather cumulus is coded as 01, towering cumulus is coded as 02, etc. Then, during the classification process, the model will look up the coding table according to the category label of the cloud area and assign a corresponding code to the cloud area. This coding method makes the data of cloud categories more standardized in subsequent processing, facilitating further analysis and decision-making. The finally generated cloud category parameter can be used in decision-making systems such as weather modification to help identify different categories of clouds and provide decision support for related operations.

[0108] In some embodiments, the classification of the target cloud includes seven main cloud categories, each cloud having unique morphological characteristics and meteorological meanings, as follows: Fair-weather cumulus: Its shape is flat, with a slightly arched top, a horizontal width greater than the vertical thickness, a flat bottom, and a convex top. It usually forms when the air convection is weak, the weather is stable, and it is common in sunny days. Its distribution is isolated, and there may be local shadows, but it does not bring precipitation.

[0109] Cumulonimbus: It has a thick vertical development, with a thickness of up to 4 - 5 kilometers. It is tall in shape, with an overlapping dome - shaped top, and its appearance is similar to cauliflower. Cumulonimbus forms under strong air convection conditions, often accompanied by extreme weather such as thunderstorms. Its surface has obvious shadows, the sun - irradiated part is brighter, and the shaded part is darker.

[0110] Fractocumulus: Its edges are broken and the outline is incomplete, with an irregular shape, and it is mostly composed of tiny water droplets. Fractocumulus usually forms before the development of fair - weather cumulus or after cumulus clouds are dispersed by the wind, indicating stable weather and usually not bringing significant weather changes.

[0111] Stratus: The stratus cloud body is uniform and layered, with a color of gray or gray - white. Its thickness is usually between 400 - 500 meters, and the height of the cloud base from the ground is generally no more than 2000 meters. Stratus usually does not bring precipitation, shows a relatively flat shape, and has stable weather, commonly seen on sunny days.

[0112] Fractostratus: Its shape is incomplete and the edges are broken. It is usually composed of small water droplets and has a variable form. Fractostratus generally appears before the formation of fair - weather cumulus or when it is dispersed by the wind, indicating stable weather and usually not bringing precipitation.

[0113] Cumulonimbus incus: Cumulonimbus incus is the cloud body with the most powerful vertical development, presenting a tall and thick shape, and its top is usually in the shape of an anvil or cauliflower. The formation of cumulonimbus incus is closely related to the atmospheric instability. It is often accompanied by strong updrafts and can produce severe weather such as thunderstorms, lightning, and hail, which has an important impact on agriculture, transportation, and daily life.

[0114] Nimbostratus: Nimbostratus is a low and extensive cloud layer that usually covers the entire sky and brings continuous precipitation or snow. Such cloud layers usually form under the influence of frontal systems, bringing relatively uniform and continuous precipitation, which has a positive effect on alleviating drought and replenishing water resources.

[0115] It can be understood that the characteristics of these cloud categories are of great significance for artificial weather modification (such as rain enhancement, hail prevention, etc.), which can help select appropriate catalysts and seeding methods. In addition, the cloud classification method provided by the present invention improves the accuracy of artificial weather modification and provides valuable data support for meteorological research and disaster warning.

[0116] Step S604: Extract grayscale image data from the cloud region corresponding to the pixel - level segmentation mask, calculate the contrast, correlation, and energy values based on the gray - level co - occurrence matrix, and generate texture morphological feature parameters using the local binary pattern algorithm; Among them, after cloud classification, it is next necessary to extract the texture features of the clouds in order to perform more precise analysis of the clouds. Texture features play an important role in the morphological analysis of clouds, and they can reveal the surface details of clouds, such as roughness, contrast, etc. By extracting grayscale image data from the segmented cloud regions, the gray-level co-occurrence matrix (GLCM) can be used to analyze the texture features of the images.

[0117] Specifically, the gray-level co-occurrence matrix is a statistical method for describing image texture. It obtains texture information by calculating the co-occurrence frequency of adjacent pixel gray values in the image. By calculating indicators such as the contrast, correlation, and energy of the GLCM, the texture features of the clouds can be obtained. Contrast describes the degree of gray change in the image, correlation reflects the linear relationship between pixels in the image, and energy represents the consistency of the texture. In addition, the local binary pattern (LBP) algorithm is a powerful texture feature extraction method. It can capture the texture characteristics of local regions of the image and generate a binary pattern for each pixel to describe the local changes in the texture.

[0118] Step S605: Input the texture morphological feature parameters into a pre-trained height regression model, perform a non-linear mapping on the correlation between the texture morphological features and the height through a multi-layer perceptron network, and output the vertical height parameters; Among them, the vertical height of the clouds is a very important parameter in cloud analysis. Clouds at different heights have different texture features. In this step, by inputting the extracted texture feature parameters into a pre-trained height regression model, the vertical height of the clouds can be estimated based on the texture morphological features of the clouds. The regression model uses a multi-layer perceptron network (MLP) to perform non-linear mapping, learn the relationship between the texture features and the height of the clouds, and calculate the height of the clouds through forward propagation.

[0119] During this process, the texture features of the clouds are used as input, and spatial position features (such as the coordinates of the center point of the candidate box) are also input into the model as additional features to provide more information about the spatial position of the clouds. Through these features, the network will output a vertical height parameter representing the actual height of the clouds.

[0120] Step S606: Integrate the cloud amount ratio parameter, cloud layer category parameter, texture morphological feature parameter, and vertical height parameter to construct a cloud feature parameter set.

[0121] Among them, by integrating each feature parameter, a complete cloud feature parameter set can be obtained. These parameters include the cloud amount ratio, cloud layer category, texture features, and the vertical height of the clouds. They jointly form a multi-dimensional data set that can comprehensively describe the characteristics of the clouds.

[0122] It is understandable that the set of cloud feature parameters can assist decision-makers in making more accurate weather intervention decisions based on detailed cloud features. For example, the cloud cover ratio can guide the usage amount of cloud catalysts, the cloud category can help select the appropriate catalyst category, and the cloud height can assist in determining whether the clouds in the area are suitable for weather intervention operations.

[0123] In the above embodiments, the multi-dimensional features are integrated, and the constructed set of cloud feature parameters provides a scientific decision-making basis for weather modification operations, making the decision-making more accurate and efficient. The construction of such a data set not only improves the accuracy of cloud feature extraction but also provides important data support for meteorological research and disaster warning.

[0124] This application combines object detection and image segmentation technologies to solve the problems of cloud identification and analysis in weather modification operations. By integrating the Region Proposal Network (RPN) and the Region-based Convolutional Neural Network (Fast R-CNN), this solution can quickly and accurately extract cloud features, significantly improving the speed and accuracy of cloud object detection. With ResNet50 as the feature extraction network, the system can efficiently capture features of different scales and complexities, providing strong data support for the accurate identification of clouds.

[0125] Another major advantage of this application lies in its multi-dimensional cloud analysis ability. Through the application of Faster R-CNN, the system can not only calculate the cloud cover, identify cloud types, but also describe the cloud shape and estimate the cloud height, providing comprehensive data support for weather modification operations. Compared with traditional methods, this integrated and multi-dimensional analysis framework can provide more accurate and comprehensive cloud information, helping operators make real-time decisions and optimize operation strategies.

[0126] In addition, the high adaptability of the system ensures that it can maintain high performance under different meteorological conditions. Whether it is different types of clouds or complex weather environments, the trained and optimized model can adapt to various changes and improve operation efficiency. Through real-time image acquisition and processing, the system can quickly feedback the latest cloud information, supporting operators to adjust strategies in a timely manner based on the latest meteorological data, reducing resource waste and improving operation effects.

[0127] In summary, this application not only improves the accuracy of cloud identification but also significantly enhances the scientificity and effectiveness of weather modification operations through multi-dimensional analysis, real-time feedback, and data-driven decision support, providing an innovative solution for applications in related fields.

[0128] The embodiments of this application also disclose a system for identifying target clouds in weather modification operations.

[0129] A human shadow operation target cloud layer recognition system, the recognition system comprising: An acquisition module, configured to acquire original cloud layer image data of a target cloud layer area; A preprocessing module, configured to preprocess the original cloud layer image data to generate a standardized feature image; A model construction module, configured to pre-construct a deep convolutional neural network architecture; A feature extraction module, configured to perform multi-level feature extraction on the standardized feature image based on a feature extraction network to generate a feature map; A bounding box generation module, configured to traverse the feature map based on a region proposal network to generate an initial detection result including candidate region bounding boxes; A candidate region determination module, configured to perform feature alignment processing on the candidate region bounding boxes based on a region of interest pooling algorithm and output a feature vector of the candidate regions; A classification and regression module, configured to input the feature vector of the candidate regions into a classification and regression network, perform multi-class classification prediction and bounding box position regression calculation, and output a detection result set including cloud layer class labels and target cloud layer bounding boxes; A convolutional segmentation module, configured to perform full convolutional segmentation operations on each candidate region in the detection result set to generate a pixel-level segmentation mask; A cloud layer parameter set construction module, configured to calculate the cloud amount ratio, cloud layer class, texture morphology features, and vertical height parameters of the target cloud layer area respectively according to the pixel-level segmentation mask and the cloud layer class labels, and construct a cloud layer feature parameter set.

[0130] The human shadow operation target cloud layer recognition system according to an embodiment of the present application can implement any one of the above-mentioned target cloud layer recognition methods, and the specific working processes of each module in the target cloud layer recognition system can refer to the corresponding processes in the above-mentioned method embodiments.

[0131] In several embodiments provided by the present application, it should be understood that the provided methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for example, the division of a certain module is only a logical function division, and there may be other division methods in actual implementation. For example, multiple modules can be combined or integrated into another system, or some features can be ignored or not executed.

[0132] An embodiment of the present application also discloses a computer device.

[0133] The computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the above-mentioned human shadow operation target cloud layer recognition method.

[0134] The embodiments of the present application also disclose a computer-readable storage medium.

[0135] The computer-readable storage medium stores a computer program that can be loaded and executed by a processor and is any one of the methods for identifying a target cloud layer in a human shadow operation as described above.

[0136] Among them, the computer-readable storage medium can be any tangible medium that contains or stores a program, and this program can be used by or in combination with an instruction execution system, device, or component; the program code contained on the computer-readable medium can be transmitted by any appropriate medium, including but not limited to wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0137] It should be noted that in the above embodiments, the descriptions of the various embodiments have their own emphases. For the parts not detailed in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0138] The above are all the preferred embodiments of the present application. Without limiting the protection scope of the present application based on this, any feature disclosed in this specification (including the abstract and drawings), unless specifically described, can be replaced by other equivalent or similar-purpose alternative features. That is, unless specifically described, each feature is only an example of a series of equivalent or similar features.

Claims

1. A method for identifying a target cloud layer for human figure operation, characterized in that, The recognition method includes: Obtaining the original cloud image data of the target cloud area; Preprocessing the original cloud image data to generate a standardized feature image; Pre - constructing a deep convolutional neural network architecture, and performing multi - level feature extraction on the standardized feature image based on the feature extraction network to generate a feature map; Traversing the feature map based on the region proposal network to generate an initial detection result including candidate region bounding boxes; Performing feature alignment processing on the candidate region bounding boxes based on the region of interest pooling algorithm to output the feature vectors of the candidate regions; Inputting the feature vectors of the candidate regions into the classification regression network, performing multi - class classification prediction and bounding box position regression calculation, and outputting a detection result set including cloud class labels and target cloud bounding boxes; Performing a full - convolution segmentation operation on each candidate region in the detection result set to generate a pixel - level segmentation mask; According to the pixel - level segmentation mask and the cloud class labels, calculating the cloud amount ratio, cloud class, texture morphology features, and vertical height parameters of the target cloud area respectively, and constructing a cloud feature parameter set.

2. The method for identifying a target cloud layer for a human figure operation according to claim 1, wherein The step of traversing the feature map based on the region proposal network to generate an initial detection result including candidate region bounding boxes includes: Applying a convolutional kernel on the feature map for sliding window scanning to generate an intermediate feature map; Performing a two - branch convolution operation on the intermediate feature map to generate an anchor box classification confidence tensor and an anchor box position offset tensor for each spatial position respectively; Generating a corresponding anchor box group at each spatial position of the feature map based on a preset reference anchor box template; Calculating the initial bounding box coordinates of the candidate regions based on the anchor box group according to the anchor box classification confidence tensor and the anchor box position offset tensor; Calculating the intersection - over - union of the initial bounding box of the candidate region and the true cloud annotation box according to the initial bounding box coordinates of the candidate region, screening and retaining the anchor boxes in the anchor box group that meet the preset intersection - over - union threshold; Performing spatial correction on the filtered anchor box group to obtain an effective anchor box group; Performing iterative processing on the effective anchor box group based on the non - maximum suppression algorithm, and outputting the optimized candidate region bounding boxes as the initial detection result.

3. The method for identifying the target cloud layer for cloud seeding operation according to claim 2, characterized in that, The step of performing feature alignment processing on the candidate region bounding boxes based on the region of interest pooling algorithm to output the feature vectors of the candidate regions includes: Mapping the candidate regions to the corresponding spatial positions on the feature map based on the scaling ratio parameter between the original cloud image data and the feature map; Uniformly dividing the mapped candidate regions into grid sub - regions with K rows and K columns in the feature map space; where K is a preset positive integer; Performing a max - pooling operation on each grid sub - region, extracting the maximum pooling result of the features of each grid sub - region, and splicing them in spatial order to obtain the feature vectors of the candidate regions.

4. A method for identifying the target cloud layer of a human figure operation according to claim 1, characterized in that The step of pre - constructing a deep convolutional neural network architecture includes: Obtaining a set of original meteorological cloud image samples and performing standardization processing, and dividing them into a training set and a validation set; Constructing a three - level network architecture including a feature extraction network, a region proposal network, and a classification regression network; Input the training set into the three - level network architecture, generate corresponding shared feature maps based on the feature extraction network, and perform sliding window convolution operations based on the region proposal network to generate multi - scale anchor box maps; By calculating the intersection - over - union of the anchor boxes and the labeled bounding boxes, filter positive and negative samples and construct binary classification labels; Based on the binary classification labels, optimize the classification branch of the region proposal network through the binary cross - entropy loss function, and optimize the regression branch of the region proposal network through the smooth L1 loss function; Map the candidate regions generated by the region proposal network to the shared feature maps, perform region of interest pooling operations to generate fixed - size feature vectors; Input the fixed - size feature vectors into the fully - connected layer of the classification and regression network for dimensionality reduction processing, and output a dimensionality - reduced feature matrix; Perform multi - task prediction based on the dimensionality - reduced feature matrix, and calculate the class probability distribution and bounding box offsets; Obtain the labeled target cloud class labels and bounding box coordinates in the training set, calculate the classification loss according to the class probability distribution and the target cloud class labels, and calculate the regression loss according to the bounding box offsets and the bounding box coordinates; Jointly optimize the classification loss and the regression loss through the backpropagation algorithm, and update the weight parameters of the classification and regression network; Freeze the weight parameters of the feature extraction network, jointly train the region proposal network and the classification and regression network to generate a primary training model; Unfreeze the weight parameters of the feature extraction network, perform global parameter fine - tuning based on the primary training model, and update the weight parameters of the feature extraction network, the region proposal network, and the classification and regression network; Validate the primary training model based on the validation set until the total loss function meets the preset conditions or the model iteration times reach the preset times, and output the completed deep convolutional neural network architecture.

5. A method for identifying a target cloud layer for cloud seeding operations according to claim 1, characterized in that, The steps of performing full - convolutional segmentation operations on each candidate region in the detection result set to generate pixel - level segmentation masks include: Input the feature vectors of the candidate regions into the segmentation sub - network of the fully - convolutional neural network, and perform up - sampling processing on the feature vectors of the candidate regions through at least three de - convolutional layers to generate an initial segmentation map; Perform per - pixel classification operations on the initial segmentation map to generate a probability distribution map containing the probability values of each pixel belonging to the target cloud; Perform binary processing on the probability distribution map based on a preset probability threshold, and output a pixel - level segmentation mask containing cloud pixel marking information.

6. A method for identifying a target cloud layer for cloud seeding operations according to any one of claims 1 to 5, characterized in that, The steps of calculating the cloud amount ratio, cloud class, texture morphology features, and vertical height parameters of the target cloud region respectively according to the pixel - level segmentation mask and the cloud class labels, and constructing a cloud feature parameter set include: Perform binary processing based on the pixel - level segmentation mask to generate a valid cloud region pixel marking matrix; According to the valid cloud region pixel marking matrix, calculate the ratio of the number of valid cloud region pixels to the total number of pixels to obtain the cloud amount ratio parameter; Match the preset cloud class coding table according to the cloud class labels, perform class coding on the candidate regions, and obtain the cloud class parameter; Extract grayscale image data from the cloud region corresponding to the pixel-level segmentation mask, calculate the contrast, correlation, and energy values based on the gray-level co-occurrence matrix, and generate texture morphological feature parameters using the local binary pattern algorithm; Input the texture morphological feature parameters into a pre-trained height regression model, perform a non-linear mapping on the correlation between the texture morphology features and height through a multi-layer perceptron network, and output the vertical height parameters; Integrate the cloud amount ratio parameter, cloud layer category parameter, texture morphological feature parameter, and vertical height parameter to construct a cloud layer feature parameter set.

7. A human figure operation target cloud layer recognition system, characterized in that, The recognition system includes: An acquisition module for acquiring the original cloud image data of the target cloud region; A preprocessing module for preprocessing the original cloud image data to generate a standardized feature image; A model construction module for pre-constructing a deep convolutional neural network architecture; A feature extraction module for performing multi-level feature extraction on the standardized feature image based on a feature extraction network to generate a feature map; A bounding box generation module for traversing the feature map based on a region proposal network to generate an initial detection result including candidate region bounding boxes; A candidate region determination module for performing feature alignment processing on the candidate region bounding boxes based on the region of interest pooling algorithm and outputting the feature vectors of the candidate regions; A classification and regression module for inputting the feature vectors of the candidate regions into a classification and regression network, performing multi-class classification prediction and bounding box position regression calculation, and outputting a detection result set including cloud layer category labels and target cloud region bounding boxes; A convolutional segmentation module for performing full convolutional segmentation operations on each candidate region in the detection result set to generate a pixel-level segmentation mask; A cloud layer parameter set construction module for calculating the cloud amount ratio, cloud layer category, texture morphological features, and vertical height parameters of the target cloud region according to the pixel-level segmentation mask and cloud layer category labels, and constructing a cloud layer feature parameter set.

8. A computer device, characterized in that: It includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements a method for identifying target cloud layers in cloud seeding operations according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: A computer program stored that can be loaded and executed by a processor to implement a method for identifying target cloud layers in cloud seeding operations according to any one of claims 1 to 6.

Images