Inversion Method Based on Radar Reflectivity Cascade Model of Himawari Satellite Multi-channel Brightness Temperature

Abstract

This invention provides a method for inverting radar reflectivity using a cascaded model based on the brightness temperature of the Himawari satellite. The method includes the following steps: constructing a cascaded model based on a UNet structure; selecting Himawari satellite data and radar reflectivity mosaic data at corresponding times; training the constructed cascaded model; and inverting radar reflectivity data using the Himawari satellite data. The beneficial effects of this invention are: it achieves high-precision inversion of radar reflectivity under different weather conditions, greatly improving the resolution and stability of meteorological monitoring. It can be widely applied in short-term forecasting, severe convection monitoring, and quantitative precipitation estimation, providing strong technical support for meteorological early warning and disaster prevention.

Description

Technical Field

[0001] This invention belongs to the field of atmospheric sounding technology, and in particular relates to a radar reflectivity cascade model inversion method based on the brightness temperature of the multi-channel Himawari satellite. Background Technology

[0002] In the field of meteorological remote sensing, the technology of retrieving radar reflectivity from satellite brightness temperature data still faces significant challenges. Traditional methods mainly rely on physical radiative transfer models or simple statistical regression, which are insufficient to effectively characterize the complex nonlinear mapping relationship between multi-channel brightness temperature and radar reflectivity. Physical models have high computational complexity, with single-scene processing taking more than 30 minutes, and are extremely sensitive to atmospheric profile input data; statistical methods, due to the failure of linear assumptions, show significantly increased retrieval errors under extreme weather scenarios. More seriously, 71% of the global ocean surface and areas such as the Qinghai-Tibet Plateau lack ground-based radar coverage, forming persistent monitoring blind spots, while existing satellite retrieval products have insufficient spatial resolution (>10km) and poor timeliness, failing to meet the needs of short-term forecasting of severe convective weather.

[0003] The core problem at the data level is the excessively high proportion of zero-value samples, leading to severe overfitting of conventional models during training and generating a large number of spurious echoes. This reduces product reliability, with a false negative rate of 35% in precipitation areas and a systematic underestimation in areas of heavy precipitation. Current technology fails to fully utilize the collaborative information from the brightness and temperature data of the Himawari satellite's 13 channels; spatiotemporal matching uses fixed time windows and simple grid interpolation, causing discrepancies between cloud top and cloud interior positions. Oversampling strategies in sample processing disrupt the original data distribution.

[0004] These current issues result in significant deficiencies in the spatiotemporal continuity and physical consistency of the inversion results, making it difficult to accurately capture the evolutionary characteristics of severe convective weather systems. Particularly during sudden severe weather events such as thunderstorms, strong winds, and hail, existing models exhibit high response delays and blurred echo structures, leading to insufficient timeliness and precision in early warnings. Furthermore, the models do not sufficiently extract information from multi-channel brightness temperature data and fail to effectively integrate complementary features between infrared channels, further limiting the accuracy and stability of radar reflectivity inversion. Summary of the Invention

[0005] In view of this, the present invention aims to propose a radar reflectivity cascade model inversion method based on the brightness temperature of the multi-channel Himawari satellite, in order to solve the technical problems of weak reflectivity intensity, incomplete blind zone coverage, complex calculation of existing models or large errors in extreme weather, false echoes caused by overfitting of zero-value samples, insufficient utilization of multi-channel data, and delay in response to strong convection.

[0006] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0007] A radar reflectivity cascade model inversion method based on the multi-channel brightness temperature of the Himawari satellite includes the following steps:

[0008] S1. Construct a cascaded model;

[0009] S2. Based on satellite data and radar reflectivity mosaic, satellite data-radar reflectivity labels are created through data filtering and spatiotemporal matching.

[0010] S3. Train the cascaded model using satellite data and radar reflectivity labels;

[0011] S4. Using a cascaded model, invert the radar reflectivity;

[0012] In step S1, the cascaded model includes a zero-value classifier, a non-zero regressor, a cascaded decision function, and a joint loss function;

[0013] In step S4, the radar reflectivity is inverted using a cascaded model, including:

[0014] S41. Input satellite data;

[0015] S42. The zero-value classifier extracts features without radar echoes from satellite data and outputs the zero-value probability in the interval [0, 1].

[0016] S43. Based on the zero-value probability and dynamic threshold τ, the cascaded decision function determines whether to directly output zero reflectivity or to activate the non-zero regressor.

[0017] S44. If the non-zero regressor is activated, the reflectivity value of the echo region is calculated and output.

[0018] S45. Based on the gradient balancing mechanism, the backpropagation weights are dynamically adjusted using the joint loss function to avoid classifier degradation and regressor overfitting.

[0019] Furthermore, the zero-value classifier employs a lightweight UNet classification network based on residual coding, including a four-level downsampling structure and global feature fusion.

[0020] Furthermore, the non-zero regressor employs an improved lightweight Res-UNet architecture, including a four-level encoder, a bottleneck layer, and a four-level decoder; the non-zero regressor introduces depthwise separable convolution techniques.

[0021] Furthermore, in the cascaded decision function, a dynamic threshold τ is set. If the classifier output probability > τ, zero reflectivity is directly output; otherwise, a non-zero regressor is activated.

[0022] Furthermore, in the joint loss function, the classifier uses BCE Loss and the regressor uses MSE Loss, and the two are weighted and summed for backpropagation; a gradient balancing mechanism is introduced, which reduces the weight of the classifier when the proportion of zero samples is greater than a set value, and increases the weight of the regressor when the error of non-zero samples is greater than a set value.

[0023] Furthermore, in step S2, based on satellite data and radar reflectivity mosaic, satellite data-radar reflectivity labels are created through data filtering and spatiotemporal matching, including:

[0024] S21. Select satellite data with obvious cloud areas in the Sunflower Satellite Channel;

[0025] S22, Set the spatiotemporal alignment precision;

[0026] S23. Based on the spatiotemporal alignment accuracy, select ground radar mosaics synchronized with satellite data;

[0027] S24. Spatial segmentation: Remove samples with a null value rate >70% from the radar mosaic, and set the null value region to zero.

[0028] S25. Based on the screening results of steps S21 to S24, the screened sample set is divided into training set, validation set and test set according to a set ratio, and satellite data-radar reflectivity labels are created.

[0029] Furthermore, in step S3, a cascaded model is trained using satellite data and radar reflectivity labels, including:

[0030] Based on satellite data and radar reflectivity labels, a cascaded model is trained, and the network structure and loss weight parameters are continuously adjusted until the loss curve is stable and close to 0, and the radar echo recognition accuracy is ≥90%.

[0031] Compared with existing technologies, the radar reflectivity cascade model inversion method based on multi-channel brightness temperature of the Sunflower satellite described in this invention has the following advantages:

[0032] (1) By introducing a cascaded model structure, the spatiotemporal continuity and physical consistency of radar reflectivity inversion are effectively improved.

[0033] (2) The collaborative work of the zero-value classifier and the non-zero regressor avoids the misjudgment of weak signals in traditional methods and improves the timeliness and accuracy of the response to severe convective weather systems.

[0034] (3) Multi-channel brightness temperature data are fully fused through lightweight U-Net, which enhances the ability to extract complementary features between infrared channels, thereby significantly improving the inversion accuracy and stability.

[0035] (4) The data screening and labeling process is more rigorous, eliminating samples with high null values, ensuring the quality of training data, and further optimizing model performance.

[0036] (5) The design of the joint loss function takes into account both classification and regression tasks. It is optimized in a unified manner by weighted summation, so that the model can maintain overall stability while improving its sensitivity and adaptability to changes in radar reflectivity. Attached Figure Description

[0037] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0038] Figure 1 This is a schematic diagram of the overall architecture of the cascaded model described in an embodiment of the present invention;

[0039] Figure 2 This is a schematic diagram of satellite data-radar reflectivity tags according to an embodiment of the present invention;

[0040] Figure 3 This is a schematic diagram of the prediction results of the cascaded model described in this embodiment of the invention on July 23, 2025. Detailed Implementation

[0041] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0042] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0043] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0044] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0045] like Figures 1 to 3 As shown, a radar reflectivity cascade model inversion method based on the multi-channel brightness temperature of the Himawari satellite includes the following steps:

[0046] Step 1: Construct a cascaded model;

[0047] Step 2: Select satellite data and radar reflectivity mosaic for the corresponding time and create corresponding labels;

[0048] Step 3: Based on the constructed cascaded model, perform radar reflectivity inversion calculation.

[0049] Specifically as follows:

[0050] A1. Zero-value classifier: Employs a lightweight UNet classification network based on residual coding, containing a four-level downsampling structure and a global feature fusion layer, outputting zero-value probabilities via a sigmoid function. The expression is as follows:

[0051] ;

[0052] In the formula, These are the probability predictions of the regressor. This is the weight matrix for the first layer. Here is the weight matrix for the second layer, and ReLU is the activation function, expressed as ReLU(z) = max(0, z). For feature extraction function, This is the Sigmoid function.

[0053] A2. Non-zero regressor: It adopts an improved lightweight Res-UNet structure, which includes a four-level encoder-bottleneck layer-four-level decoder and outputs non-zero reflectivity.

[0054] A3. Cascaded Decision: If the classifier output probability > threshold τ, then the final output is set to zero; otherwise, the regressor output is used as the radar reflectivity inversion value. The expression is as follows:

[0055] ;

[0056] In the formula, For the regressor output, To achieve dynamic allocation of computing resources, the threshold is specially adjusted.

[0057] A4. Joint Loss: The classifier uses BCE Loss, and the regressor uses MSE Loss. The weighted sum of these two losses is then used for backpropagation. The expression for the joint loss function is as follows:

[0058] ;

[0059] In the formula, and These are adjustable weighting coefficients. For binary labels (all reflectance is zero = 1, non-zero = 0). For joint losses, For the BCE loss term of the classifier, The MS loss term for the regressor. Given the input feature matrix, This represents the true value of radar reflectivity. This represents the probability prediction value of the regressor.

[0060] In step 2, select Himawari satellite data with obvious cloud areas and create tags. Details are as follows:

[0061] B1. Select Himawari satellite data with obvious cloud areas and synchronously matched ground observation mosaics with obvious radar reflectivity distribution;

[0062] B2. Use the radar reflectivity mosaic of the corresponding time as the label;

[0063] B3. Perform spatial segmentation on the corresponding data, filter samples with a null value rate of less than 70% in the radar reflectivity mosaic (samples with a null value rate of more than 70% are removed), and set the null value region to zero; divide the filtered sample set into training set, validation set and test set.

[0064] Example 1:

[0065] Step 1: Constructing the Cascade Model

[0066] C1, Zero-value classifier: Based on a lightweight UNet, it outputs the "probability of no reflectivity";

[0067] The zero-value classifier, as the core innovative component of the cascaded model, is specifically designed to solve the challenge of accurately identifying radar-echo-free areas in meteorological satellite remote sensing. This module employs a four-level residual downsampling architecture, using progressively increasing channel numbers (32→64→128→256) residual convolutional blocks and 2×2 max pooling operations to achieve deep feature extraction from the 13 channels of the Himawari satellite data. In the feature processing stage, adaptive global average pooling is used to efficiently compress feature maps of arbitrary sizes into 1×1 vectors, completely eliminating spatial redundancy while preserving key channel-dimensional information. Finally, a probability prediction head containing fully connected layers, Dropout regularization, and a Sigmoid activation function outputs the zero-value probability in the [0,1] interval, forming a precise basis for judging radar-echo-free areas.

[0068] C2, Non-zero reflector: Based on the improved lightweight RES-UNet, it outputs non-zero reflectivity;

[0069] The non-zero regressor, as the core inversion component of the cascaded model, is specifically designed to solve the challenge of high-precision inversion of radar reflectivity in radar echo regions. This module adopts a lightweight RES-UNet architecture. In the encoder section, a four-level residual downsampling module (basic channel count 128) extracts multi-scale brightness and temperature features step-by-step. Each level includes a residual convolution block and a 2×2 max pooling operation, achieving deep feature fusion of 13 channels of data from the Himawari satellite. In the decoder section, transposed convolution upsampling and skip connection techniques are used to accurately fuse low-resolution abstract features with high-resolution spatial information, ultimately outputting a high-precision reflectivity image through a 1×1 convolution. A depthwise separable convolution technique is introduced, decomposing the standard convolution into channel-group convolution and 1×1 pointwise convolution, reducing the number of parameters by 70% while maintaining feature extraction capabilities. A residual bottleneck structure is added at the skip connections, reducing the dimensionality to 1 / 4 of the original channel count through 1×1 convolution before feature processing, further compressing the model parameters by 30%.

[0070] C3. Cascaded Decision: If the classifier output probability > threshold τ, then the final output is set to zero; otherwise, the U-Net result is adopted.

[0071] The cascaded decision-making mechanism, as the core of intelligent control, enables dynamic collaborative work between the zero-value classifier and the non-zero regressor. This mechanism intelligently determines whether to activate the reflectance inversion process by analyzing the probability signal (range [0,1]) output by the zero-value classifier in real time, combined with a seasonal adaptive threshold strategy. When the probability value is higher than the dynamic threshold, zero reflectance output is directly generated; when the probability value is lower than the threshold, a lightweight residual U-Net is activated for deep computation, ultimately fusing the two types of output to generate a complete reflectance product.

[0072] C4. Joint Loss: The classifier uses BCE Loss, the regressor uses MSE Loss, and the two are weighted and summed for backpropagation.

[0073] A joint loss function is constructed using a weighted strategy of α:β=1:1:

[0074] ;

[0075] A gradient balancing mechanism is introduced to dynamically adjust the backpropagation weights, thereby avoiding classifier degradation (automatically reducing the α weight when the proportion of zero samples is >60%) and regressor overfitting (increasing the β weight when the error of non-zero samples is >1dBZ), and ensuring the collaborative optimization of the two networks.

[0076] Step 2: Sample creation based on sunflower data and radar reflectivity mosaic.

[0077] D1. To ensure that the data sample has good temporal and spatial generalization and can be applied to different seasons and regions, the Sunflower Satellite observation data of the lower Yangtze River region from July 2024 to July 2025 was selected, which includes 13 channels excluding visible light. The data was downloaded and processed.

[0078] D2. Select radar reflectivity mosaic data from the lower reaches of the Yangtze River from July 2024 to July 2025. After quality control, remove terrain-obstructed areas to ensure that the spatiotemporal alignment accuracy of the data is ≤5 minutes / 1km.

[0079] D3. The selected Sunflower satellite data and radar reflectivity mosaic were spatiotemporally matched and cropped to 256×256 to form a sample set. To improve the training efficiency of the model, the samples were screened, and only samples with a radar reflectivity effective value ratio of more than 30% were retained. Finally, a total of 4037 training samples were obtained (2896 training sets, 571 validation sets, and 570 test sets).

[0080] Step 3, Model Training and Prediction

[0081] Using the dataset constructed in step two, the cascaded model is trained. By adjusting parameters, changing the loss function and network structure, the network loss curve is fitted and close to 0. At the same time, it can accurately identify more than 90% of effective radar reflectivity, thus meeting the requirements of efficient and accurate reflectivity inversion.

[0082] like Figure 3 As shown, this is a schematic diagram of the prediction results of the cascaded model on July 23, 2025. The first column is the actual radar reflectivity, and the second column is the corresponding inversion result.

[0083] A cascaded deep learning model is trained using a multi-channel Himawari satellite and radar reflectivity mosaic dataset. By continuously optimizing the network structure and loss function parameters, it achieves efficient and accurate radar reflectivity retrieval. Real-time dynamic monitoring of severe convective weather is conducted to obtain high spatiotemporal resolution reflectivity distribution and precipitation system evolution trends, providing a basis for decision-making in severe weather warnings. Combined with digital elevation models and land cover data, it can identify terrain-enhanced precipitation areas and urban flooding risk zones, thereby enabling quantitative assessment and damage analysis of meteorological disaster impacts. This invention achieves high-precision radar reflectivity retrieval under different weather conditions through the synergistic application of a zero-value classifier and a lightweight residual U-Net, significantly improving the resolution and stability of meteorological monitoring.

[0084] Advantages and benefits of this invention:

[0085] (1) By introducing a cascaded model structure, the spatiotemporal continuity and physical consistency of radar reflectivity inversion are effectively improved.

[0086] (2) The collaborative work of the zero-value classifier and the non-zero regressor avoids the misjudgment of weak signals in traditional methods and improves the timeliness and accuracy of the response to severe convective weather systems.

[0087] (3) Multi-channel brightness temperature data are fully fused through lightweight U-Net, which enhances the ability to extract complementary features between infrared channels, thereby significantly improving the inversion accuracy and stability.

[0088] (4) The data screening and labeling process is more rigorous, eliminating samples with high null values, ensuring the quality of training data, and further optimizing model performance.

[0089] (5) The design of the joint loss function takes into account both classification and regression tasks. It is optimized in a unified manner by weighted summation, so that the model can maintain overall stability while improving its sensitivity and adaptability to changes in radar reflectivity.

[0090] 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 method for retrieving radar reflectivity from a satellite multi-channel brightness temperature cascade model based on sun-synchronous polar orbiting microwave radiometer data, characterized in that: The method comprises the following steps: S1, constructing a cascade model; S2, based on satellite data and radar reflectivity mosaics, satellite data-radar reflectivity labels are made through data screening and space-time matching; S3, using satellite data-radar reflectivity labels, the cascade model is trained; S4, using the cascade model, radar reflectivity is retrieved; In step S1, the cascade model comprises a zero-value classifier, a non-zero regressor, a cascade decision function and a joint loss function; In step S4, using the cascade model, radar reflectivity is retrieved, comprising: S41, inputting satellite data; S42, the zero-value classifier extracts features of no radar echo in the satellite data and outputs zero-value probability in the interval [0, 1]; S43, based on the zero-value probability and a dynamic threshold τ, the cascade decision function determines whether to directly output zero reflectivity or to start the non-zero regressor; S44, if the non-zero regressor is started, the reflectivity value of the echo area is calculated and outputted; S45, based on the gradient balance mechanism, the joint loss function dynamically adjusts the back propagation weight to avoid classifier degradation and regressor overfitting.

2. The method according to claim 1, wherein the method is a method for retrieving the radar reflectivity from the multi-channel brightness temperature of the sun-synchronous satellite based on the sun-synchronous satellite. The zero-value classifier adopts a lightweight UNet classification network based on residual coding, and the lightweight UNet classification network comprises a four-level downsampling structure and global feature fusion.

3. The method according to claim 1, wherein the method is a method for retrieving the radar reflectivity factor from the multi-channel brightness temperature of the sun-synchronous satellite based on the sun-synchronous satellite. The non-zero regressor adopts a lightweight Res-UNet structure, and the lightweight Res-UNet structure comprises a four-level encoder, a bottleneck layer and a four-level decoder; the non-zero regressor introduces a depth separable convolution technology.

4. The method according to claim 1, wherein the method is a method for retrieving the radar reflectivity factor from the multi-channel brightness temperature of the sun-synchronous satellite based on the sun-synchronous satellite. In the cascade decision function, a dynamic threshold τ is set, if the classifier output probability > τ, zero reflectivity is directly outputted, otherwise the non-zero regressor is started.

5. The method according to claim 1, wherein the method is a method for retrieving the radar reflectivity factor from the multi-channel brightness temperature of the sun-synchronous satellite based on the sun-synchronous satellite. In the joint loss function, the classifier uses BCE Loss and the regressor uses MSE Loss, and the two are weighted and summed for back propagation; the gradient balance mechanism is introduced, when the proportion of zero-value samples > a set value, the weight of the classifier is reduced, and when the error of non-zero samples > a set value, the weight of the regressor is increased.

6. The method according to claim 1, wherein the method is a method for retrieving the radar reflectivity from the multi-channel brightness temperature of the sun-synchronous satellite based on the sun-synchronous satellite. In step S2, based on satellite data and radar reflectivity mosaics, satellite data-radar reflectivity labels are made through data screening and space-time matching, comprising: S21, selecting satellite data with obvious cloud area in the Kaihua satellite channel; S22, setting the space-time alignment accuracy; S23, based on the space-time alignment accuracy, selecting ground radar mosaics synchronized with the satellite data; S24, spatial segmentation, removing samples with a null value rate > 70% in the radar mosaic, and setting the null value area as zero; S25, based on the screening results of steps S21 to S24, the screened sample set is divided into a training set, a validation set and a test set according to a set proportion, and satellite data-radar reflectivity labels are made.

7. The method according to claim 6, wherein the method is a method for retrieving the radar reflectivity factor from the multi-channel brightness temperature of the sun-synchronous satellite based on the sun-synchronous satellite. In step S3, using satellite data-radar reflectivity labels, the cascade model is trained, comprising: Based on satellite data-radar reflectivity labels, the cascade model is trained, and the network structure and loss weight parameters are continuously adjusted until the loss curve is smooth and tends to 0, and the radar echo recognition accuracy is ≥ 90%.

Images