A large field of view ground-based infrared cloud image generation system and cloud cover identification method

Abstract

The application relates to the technical field of image recognition, and particularly discloses a large-view-field ground-based infrared cloud image generation system and a cloud cover recognition method, which is composed of an infrared imaging unit, a temperature and humidity sensing unit, an operation control unit and a waterproof shell; the temperature and humidity sensing unit comprises two groups of temperature and humidity sensors, and the two groups of temperature and humidity sensors are respectively installed on the outer side of the waterproof shell and are located in the same plane as the infrared lens; the system generates a non-sky mask and a temperature mask by stretching and enhancing the image grayscale and eliminating the influence of the sun pixel, uniformly assigns a low value to the brightness temperature null value, splices the enhanced image and the brightness temperature into a lightweight DeepLabV3+ segmentation model, outputs the sky, cloud, sun and invalid area results, and obtains the cloud cover by counting the cloud pixel proportion in the effective area. The scheme does not need to be scanned and spliced, and is suitable for all-day real-time cloud monitoring.

Description

A wide-view field-based infrared cloud image generation system and cloud cover identification method Technical Field

[0001] This invention relates to the field of image recognition technology, and more specifically, to a large-view field-based infrared cloud image generation system and a cloud cover recognition method. Background Technology

[0002] As a key carrier of energy balance and water cycle in the atmospheric system, accurate observation of cloud macroscopic parameters (such as cloud cover, cloud height, and cloud type) and microscopic parameters (such as cloud droplet spectrum and water content) plays a vital role in understanding weather processes, improving the accuracy of weather forecasts, and studying climate change.

[0003] Currently, ground-based cloud observation mainly relies on two technical approaches: one is traditional manual observation combined with visible light equipment observation, namely visual estimation by meteorological observers and all-sky visible light camera photography. This method has low observation costs but significant limitations. For example, it is constrained by lighting conditions, operating only during the day and unable to conduct effective observations at night or under strong or weak light conditions. Manual cloud cover estimation is easily affected by subjective factors, resulting in inconsistent and difficult-to-quantify results. The other approach is active remote sensing equipment observation, such as ceilometers and weather radar. Ceilometers can accurately measure cloud base height, but they only detect single points horizontally and cannot obtain large-scale two-dimensional cloud distribution information; weather radar, while having a wide detection range, is expensive, complex to deploy, and has limited detection capabilities for low and thin clouds.

[0004] To address these shortcomings, ground-based infrared remote sensing technology for cloud observation has been developed. Infrared radiation sensors image clouds by detecting the thermal radiation (long-wave infrared, 8-14µm) emitted by the cloud base and the atmosphere below the cloud. Their advantages include being unaffected by sunlight, enabling all-day (day and night) observation, and being less affected by weather conditions such as fog and dust storms. However, due to limitations in infrared sensing technology, ground-based cloud infrared radiation remote sensing equipment has a limited field of view. For example, the CIR-7 (CloudsInfraredRadiometer) developed by Gillotay et al. at the Belgian Space Atmospheric Research Institute has an effective field of view of only 12° for a single probe. It requires mounting seven probes on a semi-circular ring and scanning and stitching the sky to obtain a full-sky infrared cloud image. The WSIRCMS (WholeSkyInfraredCloudMeasuringSystem) developed by a domestic university of science and technology improves the field of view for a single observation, but it is still only 45°×60°.

[0005] To address the limited field of view of infrared radiation sensors, infrared cloud observation equipment often employs a pan-tilt-zoom (PTZ) system for multi-angle imaging, followed by image stitching algorithms to create a large field-of-view image. While scanning and stitching can capture cloud distribution data across the entire sky, data acquisition is done in a time-division, multi-exposure manner. This process is time-consuming and fails to capture the instantaneous state of cloud conditions. When clouds move or change shape rapidly, the stitched images suffer from misalignment and ghosting, severely impacting the real-time performance and accuracy of cloud cover identification.

[0006] Therefore, there is a need in this field for a system and method that can rapidly generate and identify cloud cover in a wide field-based infrared cloud image at all times. Summary of the Invention

[0007] To overcome the aforementioned deficiencies of the prior art, this invention provides a large field-of-view ground-based infrared cloud image generation system and cloud cover identification method. The system acquires a large field-of-view infrared cloud image in a single imaging process, introduces brightness temperature data to assist in the identification of cloud areas and temperature and humidity data of the internal and external working environment, and combines a deep learning model to achieve accurate identification of cloud point pixels and cloud cover generation, thereby solving the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a large-field-of-view ground-based infrared cloud image generation system, comprising an infrared imaging unit, a computation and control unit, and a waterproof housing; further comprising a temperature and humidity sensing unit; the infrared imaging unit employs a large-field-of-view infrared lens and is coupled with an uncooled infrared focal plane detector to obtain a large-field-of-view sky infrared image with a narrow-side field of view of not less than 160° and a wide-side field of view of 180° in a single observation imaging in the thermal infrared band of 8–14 μm, and simultaneously obtains the original sky brightness temperature data within the same field of view range as the large-field-of-view sky infrared image; the temperature and humidity sensing unit comprises two sets of temperature and humidity sensors, which are respectively installed on the outside of the waterproof housing and are on the same plane as the infrared lens; the computation and control unit is ... large-field-of-view ground-based infrared cloud image generation system; the infrared imaging unit employs a large-field-of-view infrared lens and is coupled with an uncooled infrared focal plane detector to obtain a large-field-of-view sky infrared image with a narrow-side field of view of not less than 160° and a wide-side field of view of 180°, and simultaneously obtains the original sky brightness temperature data within the same field of view range as the large-field-of The computation control unit performs data preprocessing on the grayscale image and the original sky brightness temperature data. The data preprocessing includes at least: performing grayscale stretching enhancement on the grayscale image and removing the grayscale values ​​of pixels corresponding to the sun's position when determining the maximum grayscale value for grayscale stretching; generating a mask for shielding non-sky areas; uniformly assigning low values ​​to null value areas in the original sky brightness temperature data and generating a temperature mask for invalid brightness temperature areas; the computation control unit concatenates the enhanced grayscale image with the preprocessed original sky brightness temperature data, inputs it into a supervised cloud image segmentation model built based on DeepLabV3+ to output the segmentation results of the sky, clouds, sun, and invalid areas, and generates cloud cover within the areas defined by the non-sky area mask and the temperature mask.

[0009] As a further aspect of the present invention, two sets of temperature and humidity sensors are used to acquire temperature and humidity data of the external working environment where the system is located and temperature and humidity data inside the system. The temperature and humidity data inside the waterproof housing are used to monitor the status of the internal working environment of the system, and the external temperature and humidity data are used to assist in the optimization of the cloud cover recognition algorithm.

[0010] As a further aspect of the present invention, the system's working modes include a continuous working mode and a custom working cycle mode; in the continuous working mode, the system immediately begins the next working cycle after completing image acquisition and cloud cover identification; in the custom working cycle mode, the observation period can be set and the observation period is longer than the time required to complete one observation cycle in the continuous working mode.

[0011] As a further embodiment of the present invention, the computing control unit adopts an edge computing module as a data processing platform. The edge computing module is a JetsonNano module, which completes the control of image acquisition by the infrared imaging unit, data transmission, and subsequent image and data processing.

[0012] As a further aspect of the present invention, the supervised cloud image segmentation model, while retaining the dilated spatial pyramid pooling module and decoder of DeepLabV3+, reduces the depth of the backbone network to reduce the amount of computation, and retains dilated convolution to obtain image-level global features.

[0013] As a further aspect of the present invention, brightness temperature data is a proprietary term in the art. Brightness temperature data is obtained by an infrared camera. The infrared camera of this solution can simultaneously acquire images and brightness temperature data. That is, the data collected by the infrared imaging unit includes a large field-of-view sky infrared image and sky temperature data within the same field of view.

[0014] As a further aspect of the present invention, a method for cloud cover identification from a large-view field-based infrared cloud image is also provided, comprising the following steps:

[0015] Step 1: Acquire a wide field-of-view sky infrared image and raw sky brightness temperature data within the same field of view as the wide field-of-view sky infrared image;

[0016] Step 2: Preprocess the wide field-of-view sky infrared image and the original sky brightness temperature data. The preprocessing includes: performing grayscale stretching enhancement on the grayscale image and removing the grayscale values ​​of pixels corresponding to the sun's position when determining the maximum grayscale value for grayscale stretching; masking the non-sky areas of the wide field-of-view sky infrared image; uniformly assigning low values ​​to the null value areas of the original sky brightness temperature data and generating a temperature mask for the invalid brightness temperature areas.

[0017] Step 3: Construct a supervised learning cloud image segmentation model based on DeepLabV3+ suitable for large field-of-view infrared cloud images. The input data of the cloud image segmentation model includes preprocessed infrared image data, brightness temperature data within the same field of view, and temperature and humidity data of the internal and external working environment.

[0018] Step 4: Create training and test datasets. Within the valid region, label pixels as four categories: sky, clouds, sun, and invalid region. Divide the total data into training and test sets in a 4:1 ratio.

[0019] Step 5: After concatenating the infrared image data and brightness temperature data, input the data into the network for training, and obtain the segmentation model by adjusting the learning rate, convolution kernel size, number of iterations, activation function type and batch data size;

[0020] Step 6: Validate the trained segmentation model on the test set and adjust the training parameters based on the segmentation accuracy to select the optimal model;

[0021] Step 7: Based on the segmentation results of the optimal model and combined with non-sky area masks and temperature masks to limit the effective area, cloud cover is generated.

[0022] As a further aspect of the present invention, the large field-of-view sky infrared image in step one is a 16-bit grayscale image.

[0023] As a further aspect of the present invention, the absence of null values ​​in the original sky brightness temperature data includes null value anomalies caused by the brightness temperature being too low in cloudless areas, exceeding the effective observation temperature range of the infrared camera, and the null value areas are uniformly assigned preset low values ​​so that the brightness temperature data can be used for subsequent stitching input and masking.

[0024] As a further aspect of the present invention, the non-sky region mask is a pre-configured effective region mask, which masks non-sky region pixels during data preprocessing, model training, model inference and cloud generation.

[0025] As a further aspect of the present invention, the training dataset is generated by using the imaging characteristics of large field-of-view infrared cloud images for auxiliary annotation. The imaging characteristics are that the gray values ​​of the infrared cloud images of cloudless skies increase in a quadratic function form as the observed zenith angle corresponding to the pixel increases.

[0026] As a further aspect of the present invention, the optimal model outputs the segmentation results of the sky, clouds, sun and invalid regions, wherein the sun pixels can be removed from the maximum grayscale value statistics during the enhancement process to avoid affecting the grayscale stretching effect.

[0027] Compared with existing technologies, the technical effects and advantages of a large-view field-based infrared cloud image generation system are as follows: For the first time in this field, this system introduces temperature and humidity data from the internal and external working environments of the imaging system into a large-view field-based infrared cloud image generation system, which is used to assist the cloud cover recognition algorithm. In the computation and control unit, a technical solution combining a deep learning model is proposed to achieve accurate identification of cloud point pixels and cloud cover generation. The system adopts an integrated waterproof shell and dual internal and external temperature and humidity monitoring, supporting continuous mode and custom periodic mode scheduling; each period saves the entire process data, including the original / enhanced image, brightness temperature, mask, segmentation image, and cloud cover, facilitating long-term outdoor operation status diagnosis and model iteration; cloud cover is defined according to the proportion of cloud pixels in the effective area, with a clear and traceable output aperture, facilitating comparison and verification.

[0028] Compared with existing technologies, the technical effects and advantages of a large field-of-view field-based infrared cloud image generation system are as follows: This invention proposes a robust preprocessing method for large field-of-view infrared cloud images: Solar regions are located based on high grayscale connected components, and their pixels are removed from the maximum grayscale statistics, followed by grayscale stretching enhancement; simultaneously, pre-set non-sky region masks and brightness temperature masks are generated, null values ​​are uniformly assigned low values, and invalid labels are retained, ensuring that training, inference, and cloud cover statistics are all completed within a consistent effective region, resulting in more stable and reproducible results. This invention stitches enhanced infrared images and processed brightness temperature data together by channel to form a dual-modal input, constructs a lightweight DeepLabV3+ supervised segmentation model, and outputs four categories of results: sky / cloud / sun / invalid; compared with a single threshold method, it is more robust to nighttime, thin cloud, low cloud, and brightness temperature null value scenarios, and can be deployed in real time on edge devices such as JetsonNano. In addition, a single exposure can obtain a 16-bit sky infrared image with a wide field of view of 180° wide side and ≥160° narrow side in the 8-14μm thermal infrared band, and simultaneously output the original brightness temperature matrix of the same field of view, realizing data acquisition of "the same moment and the same field of view", avoiding time-sharing acquisition, ghosting, misalignment and cloud movement blur caused by scanning and stitching, and significantly improving the real-time performance of instantaneous cloud condition monitoring and cloud cover identification. Attached Figure Description

[0029] Figure 1 is a schematic diagram of the structure of a large field-view ground-based infrared cloud image generation system according to the present invention.

[0030] Figure 2 is a flowchart illustrating the cloud cover identification method for a large-view field-based infrared cloud image according to the present invention.

[0031] Figure 3 is a schematic diagram of the infrared cloud image preprocessing and sample results of the present invention. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Example 1

[0034] As shown in Figure 1, this embodiment employs an integrated outdoor structure deployed on top of a fixed bracket at an open observation point, with the lens facing the zenith. The system consists of an infrared imaging unit, a temperature and humidity sensing unit, a computing and control unit, and a waterproof housing. All components are integrated within the waterproof housing to accommodate long-term outdoor operation.

[0035] The infrared imaging unit uses a wide field-of-view infrared lens and an uncooled infrared focal plane detector to obtain a wide field-of-view sky infrared image with a narrow side field of view of no less than 160° and a wide side field of view of 180° in a single observation in the thermal infrared band of 8–14 μm. The wide field-of-view sky infrared image is a 16-bit grayscale image. Simultaneously with this image, the original sky brightness temperature data within the same field of view range is obtained and used for subsequent segmentation and inference together with the infrared image.

[0036] The temperature and humidity sensing unit comprises two sets of temperature and humidity sensors. One set is mounted on the outside of the waterproof housing, coplanar with the infrared lens, and the other inside the waterproof housing. The outer temperature and humidity sensor records the external working environment and serves as auxiliary information for optimizing the cloud cover recognition algorithm. The inner temperature and humidity sensor monitors the internal working environment of the system. Specifically, the outer temperature and humidity sensor outputs the ambient temperature and relative humidity. The computational control unit switches between preprocessing and post-processing parameter sets based on the preset relative humidity range: increasing grayscale stretching intensity and employing more sensitive discrimination parameters for thin cloud regions under high humidity conditions, and increasing noise suppression intensity to reduce false cloud spots under low humidity conditions. The inner temperature and humidity sensor assists in assessing the risk of lens condensation. When a condensation risk is determined, the reliability of the current frame result is marked, and it can be optionally excluded from cloud cover time-series updates.

[0037] The computing control unit uses an edge computing module as a data processing platform. In this embodiment, the Jetson Nano module is selected to control the infrared imaging unit's image acquisition, data transmission, and subsequent image and data processing, and output segmentation results and cloud cover results.

[0038] The system operates periodically, with two modes: continuous operation and a customizable cycle. In continuous operation, the computational control unit completes one cycle of "acquisition—preprocessing—inference—cloud cover generation—storage / reporting" and immediately begins the next cycle. In customizable cycle mode, the observation cycle is parameterized by the computational control unit and is longer than the duration required to complete one cycle in continuous operation to ensure stable scheduling. Within each observation cycle, the infrared imaging unit obtains one 16-bit grayscale image and one frame of raw sky brightness temperature data in the same field of view with a single exposure. Both are stored with the same timestamp as the smallest data unit for subsequent processing. The brightness temperature data and the infrared image correspond spatially by pixel or pixel block. When the brightness temperature output resolution is lower than the infrared image resolution, the computational control unit aligns the brightness temperature matrix to the infrared pixel grid according to a predetermined mapping relationship for subsequent stitching input. This alignment does not change the mask boundaries of invalid brightness temperature regions, as invalid regions are clearly marked by temperature masks and maintain consistent caliber in training, inference, and cloud cover statistics.

[0039] Figure 2 shows a flowchart of the cloud cover identification method, used to process the grayscale image and the original sky brightness temperature data. An optional flowchart is as follows: Data preprocessing includes infrared cloud image enhancement, brightness temperature data outlier processing, infrared cloud image masking, and invalid brightness temperature region masking. After data preprocessing, training and testing datasets are created, followed by the construction of a cloud image segmentation network. After the cloud image segmentation network is constructed, the cloud image segmentation model is trained, and finally, cloud image segmentation testing is performed.

[0040] Preprocessing is organized in parallel with image-side and brightness-temperature-side processing within the same data unit: on the image side, grayscale stretching and enhancement are performed, and non-sky region masking is applied; on the brightness-temperature side, null values ​​are uniformly assigned low values, and temperature masks for invalid brightness-temperature regions are generated. The enhanced map, non-sky region mask, brightness-temperature processing results, and temperature mask output from the preprocessing will serve as common inputs for subsequent bimodal stitching, model training / inference, and cloud cover generation. This ensures that training and inference use the same preprocessing definition, and cloud cover statistics use the same valid region limitation.

[0041] The enhancement of the 16-bit grayscale image uses grayscale stretching. Let the original grayscale value of the i-th pixel in the original 16-bit grayscale image be... After enhancement, the grayscale of this pixel is , This represents the minimum grayscale value of the original infrared image pixels in that frame. Given the maximum grayscale value used for grayscale stretching in this frame, the grayscale stretching is performed using the following formula:

[0042] ;

[0043] Where i represents the i-th pixel in the infrared image (or the pixel index of the two-dimensional coordinates (x, y)); This represents the grayscale value of the pixel in the original 16-bit infrared grayscale image; This indicates the grayscale value of the pixel in the enhanced infrared image; This represents the minimum grayscale value of all pixels in the original infrared image of that frame; This indicates the maximum grayscale value used for grayscale stretching in this frame, and is used in the calculation. The grayscale value of the pixel corresponding to the sun position is removed and then the maximum value is taken to ensure that the upper bound of the enhanced mapping is not affected by the strong point of the sun. This is a 16-bit mapping scale factor used to map the normalization result to a 16-bit grayscale range.

[0044] This embodiment determines Perform solar position pixel culling: When there are no clouds or the cloud layer is thin, the grayscale of pixels in the solar imaging area is significantly higher than that of pixels in other areas. If the maximum grayscale of the entire image is directly taken as the solar position pixel culling, the solar position pixel culling will not be effective. This will cause most pixel mappings to concentrate in the lower grayscale range; therefore, when calculating the maximum grayscale value used for grayscale stretching, first identify the sun's position region, then remove the grayscale values ​​of pixels in that region before calculating. The reproducible realization of the sun's position region employs a "high grayscale connected component localization" rule: The top 0.1% of pixels in grayscale from the original grayscale image are selected as a candidate set. Connected component aggregation is performed on this candidate set, and the connected component with the largest area is selected as the sun's candidate region. If the mean grayscale of this candidate region is greater than three times the mean grayscale of the entire image, it is confirmed as the sun's position region and is subject to elimination. After elimination, the region used for enhancement is obtained. Then, grayscale stretching is performed on the entire image. Figure 3(a) shows the original infrared grayscale image before enhancement, and Figure 3(b) shows the enhanced infrared cloud image after grayscale stretching.

[0045] Simultaneously with enhancement, a mask for shielding non-sky regions is generated. In this embodiment, the non-sky region mask uses a pre-configured effective region mask as the unified effective field-of-view boundary throughout the entire process. The generation of the effective region mask is completed during the calibration phase after equipment installation: a frame of infrared image is acquired under clear sky conditions, a stable sky field-of-view boundary is extracted, and a binary mask M is generated. IR In this model, the effective sky region is set to 1, and the non-sky region is set to 0. M is then used in data preprocessing, model training, model inference, and cloud cover generation. IR Pixels in non-sky areas are masked to ensure that the training annotation range, inference statistics range, and final cloud cover output range are consistent.

[0046] Brightness temperature processing includes uniformly assigning low values ​​to null values ​​and generating temperature masks for invalid brightness temperature regions. The original sky brightness temperature data is denoted as matrix T. For each frame of brightness temperature data, null and invalid values ​​are first located, and a temperature mask M is generated. T : Valid brightness temperature positions are set to 1, and null / invalid positions are set to 0. Null values ​​are uniformly assigned low values ​​to obtain the processed brightness temperature matrix T′, allowing the brightness temperature data to participate in subsequent stitching input and making the values ​​verifiable. This embodiment uses a reproducible low-value setting method: The minimum brightness temperature of the valid pixel set in this frame is set to... The default low value is:

[0047] ;

[0048] in, This represents the minimum brightness temperature value of the "effective brightness temperature pixel set" in the original sky brightness temperature data of this frame; This represents the preset low value used to replace the position of the brightness temperature null / invalid value; constant 5 is the low value offset (unit is consistent with the brightness temperature data), used to make... The value is less than the minimum effective brightness temperature of the frame, thus numerically distinguishing it from the effective brightness temperature, and is used in conjunction with a temperature mask for subsequent stitching input and effective area definition. This embodiment will use all M... T The value assigned to the position of =0 is This setting is independent of the camera range constant, making it easy to implement encoding across different devices; it also retains M... T Used for subsequent effective area limitation and statistical elimination. The null values ​​in the original sky brightness temperature data include anomalies caused by the brightness temperature being too low in cloudless areas, exceeding the effective observation temperature range of the infrared camera.

[0049] After the above preprocessing, the computation control unit concatenates the enhanced 16-bit grayscale image with the brightness temperature data after null value processing to form a dual-modal input. The concatenation is performed in the channel dimension: the enhanced infrared image is denoted as I′, and the processed brightness temperature matrix is ​​denoted as T′, forming the input tensor X = concat(I′, T′). Non-sky region mask M IR With temperature mask M TAlong with the input, a mask constraint is used for effective region definition and training / inference. Based on this concatenated input, a supervised cloud image segmentation model is constructed. This model is built based on DeepLabV3+ and outputs four categories of segmentation results: sky, clouds, sun, and invalid regions. In terms of network structure, the Spatial Pyramid Pooling (ASPP) module and decoder of DeepLabV3+ are retained; at the same time, the backbone network depth is reduced to decrease computational load, while dilated convolutions are retained to obtain image-level global features. The backbone reduction can be achieved by "reducing duplicate residual blocks and maintaining an output stride of 1 / 16," allowing the ASPP to receive features at this scale and output a segmentation probability map, while the decoder is used to recover boundary details. When deployed at the edge, this structure can use inference acceleration frameworks such as TensorRT to meet the inference latency requirements of periodic operation.

[0050] To obtain a deployable optimal model, this embodiment creates training and testing datasets and performs four-class labeling. During the creation of the training dataset, pixels within the valid regions are labeled as sky, clouds, sun, and invalid regions; invalid regions include M... IR =0 region and M T =0. The invalid class in the network output is used to characterize the exception M. IR =0 and M T Abnormal pixels other than 0 (such as lens condensation, obstruction, local overexposure, etc.) are represented by M in cloud cover statistics. V =M IR ·M T The statistical domain is defined, and invalid pixels are not included in the cloud category. The total data is divided into training and testing sets in a 4:1 ratio. During the annotation process, the characteristics of large field-of-view infrared cloud imagery are used for auxiliary annotation: the grayscale value of the infrared cloud imagery of a cloudless sky increases in a quadratic function form as the observed zenith angle of the pixel increases; in this embodiment, this background trend is fitted on the clear sky sample to generate a reference background, which is used to help determine the boundaries of the sky category and the thin cloud category. The annotation results still maintain the four-category system.

[0051] During the training phase, the input network is spliced ​​together with input X. Loss is calculated within the effective region during training, and the segmentation model is obtained by adjusting the learning rate, convolutional kernel size, number of iterations, activation function type, and batch data size. This embodiment provides a set of reproducible training configuration examples, as shown in Table 1, as a basis for training log recording and verification.

[0052] Table 1. Record of Model Training Parameters and Configurations during the Training and Testing Phases

[0053]

[0054] During the testing phase, the trained segmentation model is validated on the test set, and the training parameters are adjusted based on the segmentation accuracy to select the optimal model. The optimal model outputs the segmentation results for the sky, clouds, sun, and invalid regions. The sun pixels are used in the enhancement process to remove the maximum grayscale value to avoid affecting the grayscale stretching effect. After the optimization is completed, the optimal model is fixed and deployed to the JetsonNano platform for online inference.

[0055] During the online operation phase, the process of "acquisition—preprocessing—stitching—inference—cloud cover generation" is repeated for each observation cycle. The inference output is denoted as S for the four types of segmentation results, where the set of cloud-type pixels is denoted as S0. cloud Cloud formation occurs within the area defined by the non-sky area mask and the temperature mask. The effective area mask is defined as:

[0056] ;

[0057] in, This is a non-sky region mask (effective region mask), used to block pixels in non-sky regions. The mask value is 0 or 1. A value of 1 indicates that the pixel belongs to the effective sky field of view, and a value of 0 indicates that the pixel belongs to the non-sky region. This is a temperature mask (invalid brightness temperature area mask). A value of 1 indicates that the brightness temperature of the corresponding pixel is valid, and a value of 0 indicates that the brightness temperature of the corresponding pixel is empty or invalid. For the final effective region mask, it represents the set of pixels that simultaneously satisfy "effective sky field of view" and "effective brightness temperature"; where ⋅ represents element-wise multiplication.

[0058] Let P(⋅) be the pixel counting operator, then cloud cover is defined as:

[0059] ;

[0060] Where C represents the cloud cover output value; S represents the pixel-level classification result image output by the segmentation model. This represents the set of pixels (or an equivalent binary image of the cloud category) that are classified as "cloud" in the classification result image. This represents the effective area mask, used to limit the statistical range; solar pixels belong to the solar category and are not... Invalid regions pass They were excluded from the statistics.

[0061] In this embodiment, the system's computation control unit performs bitwise operations on the segmentation result image and the mask to obtain the cloud pixel count within the effective area, and outputs the cloud amount C as the cloud amount result for that period. Taking the segmentation result as input, the non-sky area mask is first applied to shield the non-sky area, and then the temperature mask is applied to shield the brightness and temperature invalid areas. After obtaining the effective area, the proportion of cloud-type pixels is counted and the cloud amount is output.

[0062] To ensure the traceability and verifiability of the results, this embodiment saves the timestamp, outer temperature and humidity, inner temperature and humidity, original 16-bit infrared image, enhanced infrared image, original brightness temperature matrix, processed brightness temperature matrix, non-sky region mask, temperature mask, four-class segmentation result image, and cloud cover C for each cycle. Figure 3(c) shows an example of the infrared mask and annotation format. The infrared mask in the figure is a type of non-sky region mask, used to verify the consistency between the four-class segmentation output and the effective region boundary in the online inference. At the same time, the saved brightness temperature mask can be used to verify whether the nighttime brightness temperature null value locations have been correctly assigned low values ​​and excluded from the effective region statistics. This is achieved through the continuous implementation of the above system structure, preprocessing method, dual-modal splicing input, DeepLabV3+ structure reduction strategy, training and testing division, adjustable training parameter set, optimal model selection, and effective region cloud cover statistics.

[0063] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0064] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A large-view field-based infrared cloud image generation system, comprising an infrared imaging unit, a computation and control unit, and a waterproof housing, characterized in that, It also includes a temperature and humidity sensing unit; the infrared imaging unit uses a wide field-of-view infrared lens and an uncooled infrared focal plane detector to obtain a wide field-of-view sky infrared image with a narrow side field of view of not less than 160° and a wide side field of view of 180° in a single observation in the thermal infrared band of 8-14μm, and simultaneously obtains the original sky brightness temperature data within the same field of view as the wide field-of-view sky infrared image; the wide field-of-view sky infrared image is a grayscale image; the temperature and humidity sensing unit includes two sets of temperature and humidity sensors, which are respectively installed on the inner and outer sides of the waterproof housing and are on the same plane as the infrared lens. The computation control unit performs data preprocessing on the large field-of-view sky infrared image and the original sky brightness temperature data. The data preprocessing includes at least: performing grayscale stretching enhancement on the grayscale image and removing the grayscale value of the pixel corresponding to the sun position when determining the maximum grayscale value for grayscale stretching; generating a mask for shielding non-sky areas; uniformly assigning low values ​​to null value areas in the original sky brightness temperature data and generating a temperature mask for invalid brightness temperature areas; the computation control unit stitches the enhanced grayscale image with the preprocessed original sky brightness temperature data, inputs it into a supervised cloud image segmentation model built based on DeepLabV3+ to output the segmentation results of the sky, clouds, sun and invalid areas, and generates cloud cover within the area defined by the non-sky area mask and the temperature mask.

2. The large-view field-based infrared cloud image generation system according to claim 1, characterized in that, The wide field-of-view sky infrared image is a 16-bit grayscale image; the temperature and humidity sensor inside the waterproof housing monitors the internal working environment of the system, and the temperature and humidity sensor outside the waterproof housing acquires the external working environment of the system.

3. The large-view field-based infrared cloud image generation system according to claim 1, characterized in that, The system has two working modes: continuous working mode and custom working cycle mode. In continuous working mode, the system starts the next working cycle immediately after completing image acquisition and cloud cover identification. In custom working cycle mode, the observation period can be set and is longer than the time required to complete one observation cycle in continuous working mode.

4. The large-view field-based infrared cloud image generation system according to claim 1, characterized in that, The computing control unit uses an edge computing module as a data processing platform. The edge computing module is a Jetson Nano module, which completes the control of image acquisition by the infrared imaging unit, data transmission, and subsequent image and data processing.

5. The large-view field-based infrared cloud image generation system according to claim 1, characterized in that, The supervised cloud image segmentation model retains the dilated spatial pyramid pooling module and decoder of DeepLabV3+, reduces the depth of the backbone network to reduce computation, and retains dilated convolution to obtain image-level global features.

6. A method for cloud cover identification in a large-view field-based infrared cloud image, characterized in that, A wide-field-view ground-based infrared cloud image generation system according to any one of claims 1-5 includes the following steps: Step 1: Acquire a wide-field-view sky infrared image and original sky brightness temperature data within the same field of view as the wide-field-view sky infrared image; Step 2: Preprocess the wide-field-view sky infrared image and the original sky brightness temperature data, including: performing grayscale stretching enhancement on the grayscale image and removing the grayscale values ​​of pixels corresponding to the sun's position when determining the maximum grayscale value for grayscale stretching; masking non-sky regions in the wide-field-view sky infrared image; uniformly assigning low values ​​to null regions in the original sky brightness temperature data and generating a temperature mask for invalid brightness temperature regions; Step 3: Constructing a supervised learning cloud image segmentation model suitable for wide-field-view infrared cloud images based on DeepLabV3+, wherein the cloud image segmentation model... The input data includes preprocessed infrared image data, brightness and temperature data within the same field of view, and temperature and humidity data of the internal and external working environment; Step 4: Create training and test datasets, labeling pixels in the effective area as four categories: sky, clouds, sun, and invalid areas, and dividing the total data into training and test sets in a 4:1 ratio; Step 5: Input the infrared image data and brightness and temperature data into the network for training, and obtain the segmentation model by adjusting the learning rate, convolution kernel size, number of iterations, activation function type, and batch data size; Step 6: Validate the trained segmentation model on the test set and adjust the training parameters according to the segmentation accuracy to select the optimal model; Step 7: Generate cloud cover based on the segmentation results of the optimal model and by combining non-sky area masks and temperature masks to limit the effective area.

7. The cloud cover identification method for a large-view field-based infrared cloud image according to claim 6, characterized in that, The absence of null values ​​in the original sky brightness temperature data includes null value anomalies caused by the brightness temperature being too low in cloudless areas, exceeding the effective observation temperature range of the infrared camera. Furthermore, the null value areas are uniformly assigned preset low values ​​so that the brightness temperature data can be used for subsequent stitching input and masking.

8. The cloud cover identification method for a large-view field-based infrared cloud image according to claim 6, characterized in that, The non-sky region mask is a pre-configured effective region mask, which masks pixels in the non-sky region during data preprocessing, model training, model inference, and cloud generation.

9. The cloud cover identification method for a large-view field-based infrared cloud image according to claim 6, characterized in that, The training dataset was created using the imaging characteristics of large field-of-view infrared cloud images for auxiliary annotation. The imaging characteristics are that the gray values ​​of infrared cloud images in cloudless skies increase in a quadratic function form as the observed zenith angle corresponding to the pixel increases.

10. A method for cloud cover identification in a large-view field-based infrared cloud image according to claim 6, characterized in that, The optimal model outputs the segmentation results of the sky, clouds, sun and invalid regions. In the enhancement process, the sun pixels can be removed from the statistics of the maximum gray value to avoid affecting the gray-scale stretching effect.

Images