Multi-scale high-precision rainfall inversion method based on PrecipNet model and multi-source visible light data

Abstract

The invention relates to a multi-scale high-precision rainfall inversion method based on a PrecipNet model and multi-source visible light data, and relates to the technical field of rainfall inversion based on a multi-source visible light cloud picture. The method comprises the following steps: preprocessing data; making a multi-scale rainfall inversion data set; an improved PrecipNet algorithm model is constructed; carrying out rainfall inversion; performing precision analysis; comparing and analyzing the model; ablation experiment; and case analysis. The invention discloses a multi-scale high-precision rainfall inversion method based on a PrecipNet model and multi-source visible light data, and aims to overcome the limitation of traditional numerical forecasting in refined rainfall inversion. Wide-range visible light data of a sunflower 8 # satellite and high-resolution visible light data of a Jilin 1 # satellite are integrated; and an attention mechanism, an LSTM framework, a multi-scale feature fusion module and a progressive training strategy are introduced, so that the precision and spatial resolution of rainfall inversion are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of rainfall inversion based on multi-source visible light cloud images, and particularly relates to a multi-scale high-precision precipitation inversion method based on the PrecipNet model and multi-source visible light data. Background Art

[0002] Precipitation, as an important part of the earth's water cycle, has a profound impact on agricultural production, ecosystem balance, and the economic development of human society. Accurate precipitation information is crucial for fields such as disaster warning, climate discovery, and hydrological simulation. However, traditional precipitation observation methods such as ground rain gauges and meteorological radars have significant limitations. First, ground rain gauges can only provide point data and cannot comprehensively reflect the precipitation distribution over a large area. Especially in areas with complex terrain or sparse population, the installation density is limited, making it difficult to achieve high-precision regional precipitation monitoring. Second, although meteorological radars have a relatively wide coverage range, their signals are easily attenuated under complex terrain conditions, resulting in a decrease in echo quality in some areas and affecting the accuracy of precipitation estimation. In addition, the installation and maintenance costs of radar equipment are high, and regular calibration is required to ensure the reliability of data, which further limits its widespread application.

[0003] With the rapid development of remote sensing technology, satellite remote sensing data has been widely used in precipitation observation. Satellite remote sensing has the advantages of wide coverage, high spatio-temporal resolution, and low data acquisition cost, and can effectively make up for the deficiencies of traditional ground observation methods. For example, IMERG (Integrated Multi-satellitE Retrievals for Global Precipitation Measurement) uses data from multiple satellites to generate global precipitation products, providing global precipitation estimates with high spatio-temporal resolution; CMORPH (Climate Prediction Center Morphing Technique) provides high-precision global precipitation estimates by fusing data from different satellite sensors, especially suitable for areas lacking ground observations. In addition, the Himawari-8 satellite and the FY-4A satellite are operating in Japan and China respectively. The high-quality cloud images and atmospheric parameter information they provide are widely used in weather monitoring and forecasting in the East Asian region, especially in the real-time tracking of typhoon and heavy rain events. These cases fully demonstrate the extensive application of satellite remote sensing in global precipitation monitoring and its significant advantages, providing strong technical support for dynamically tracking precipitation processes and timely updating forecast results.

[0004] In recent years, significant progress has been made in geophysical inventions with deep learning neural network technology, especially in the parameterization of model physics, ENSO prediction, and precipitation observation. In particular, deep learning models such as convolutional neural networks (CNNs), long short-term memory networks (LSTMs), and attention mechanisms have been widely used in the fields of precipitation estimation and nowcasting, significantly improving the accuracy and applicability of precipitation inversion. For example, Negin Hayatbini et al. invented satellite precipitation estimation using conditional generative adversarial networks (CGANs) in 2019, and the results showed overall improvements of CGAN compared to the PERSIANN-CCS product; the global precipitation nowcasting model proposed by Rahimi et al. based on the UNet and convolutional long short-term memory (ConvLSTM) architectures performed well on the IMERG dataset, especially showing significant advantages in capturing rapidly changing weather systems. In addition, Kühnlein et al. achieved precipitation inversion in the European region using Meteosat satellite visible light data combined with a random forest model, and subsequent inventions showed that replacing the random forest with a convolutional neural network (CNN) could further improve the model performance, especially in fine-grained precipitation estimation under complex terrain. Li Jun et al. and Zhang Peng et al. respectively demonstrated the potential of the ResNet architecture and multi-scale feature fusion module in improving the accuracy of precipitation estimation using Fengyun-2 and Himawari-8 satellite data combined with deep learning methods. In summary, deep learning technology not only overcomes the limitations of traditional methods in dealing with high-dimensional data by automatically extracting complex non-linear features, but also greatly improves the accuracy and real-time performance of precipitation estimation, providing strong support for more accurate and comprehensive precipitation monitoring.

[0005] Although there have been numerous inventions in precipitation monitoring based on remote sensing satellites, most of the existing methods rely on data from a single satellite and cannot simultaneously meet the requirements of large-scale coverage and high-precision monitoring of local areas. By combining multiple data sources, the deficiencies of a single data source can be compensated for, enabling large-scale and high-precision precipitation monitoring. For example, the Himawari-8 geostationary satellite provides extensive cloud mass information, and its high temporal resolution enables real-time monitoring of weather changes; while the Jilin-1 high-resolution optical satellite, with its excellent spatial resolution (up to sub-meter level), allows the model to better capture the microscopic characteristics of cloud masses. Multi-source data fusion not only provides more detailed information in terms of spatial resolution but also more frequent observations in terms of temporal resolution, thereby enhancing the ability to track rapidly changing weather systems. At the same time, the application of visible light data in precipitation monitoring is relatively lacking. Visible light images can clearly capture the texture and structural characteristics of clouds, which is crucial for identifying different types of cloud masses and their potential precipitation activities. The resolution of visible light data is relatively high compared to infrared data, and it can capture more detailed information, such as cloud top reflectivity, texture characteristics, etc., with significant advantages. For example, in complex terrain or urban environments, high-resolution visible light data can help more accurately locate the occurrence location and intensity of local precipitation events. Therefore, in daytime precipitation observations, the use of visible light data can achieve more refined precipitation monitoring than infrared data.

[0006] In addition, the existing precipitation retrieval algorithm models still have deficiencies in feature extraction and fail to fully utilize multi-temporal features. Many models only rely on static image features and ignore the temporal evolution characteristics of precipitation events. When dealing with complex weather systems, some models also lack the ability to focus on key areas, resulting in a relatively high false alarm rate. To overcome these problems, the inventors introduced advanced technologies such as long short-term memory networks (LSTM) and attention mechanisms. LSTM was proposed by Hochreiter and Schmidhuber in 1997. By introducing memory cells and gating mechanisms, it can effectively capture long-term dependencies in time series and significantly improve the modeling ability of the temporal evolution characteristics of precipitation events. The attention mechanism was first proposed by Bahdanau et al. in 2014, allowing the model to dynamically focus on different parts of the input sequence, enhancing the ability to identify key areas and reducing the false alarm rate. These technologies not only make up for the limitations of traditional methods in dealing with high-dimensional non-linear data but also greatly improve the accuracy and real-time performance of precipitation estimation, providing strong support for more accurate and comprehensive precipitation monitoring. Summary of the Invention

[0007] The present invention aims to solve the technical problems that most existing precipitation retrieval methods rely on data from a single satellite, cannot simultaneously meet the requirements of large-scale coverage and high-precision monitoring of local areas, and have insufficient feature extraction, and provides a multi-scale high-precision precipitation retrieval method based on the PrecipNet model and multi-source visible light data.

[0008] To solve the above technical problems, the technical solution of the present invention is as follows:

[0009] A multi-scale high-precision precipitation retrieval method based on the PrecipNet model and multi-source visible light data, comprising the following steps:

[0010] Step 1: Data preprocessing;

[0011] Use the visible light band of the Himawari-8 geostationary satellite data as the basic data set for precipitation satellite data retrieval; use the high-resolution visible light data of the Jilin-1 high-resolution optical satellite for refined precipitation retrieval; fuse the ERA5 precipitation data with the ground radar station observation data to construct a high-resolution and high-precision precipitation data set;

[0012] Step 2: Produce a multi-scale precipitation retrieval data set;

[0013] The multi-scale precipitation retrieval data set includes: a large-scale precipitation retrieval data set combining Himawari-8 data with ERA5 precipitation data, and a small-area refined precipitation retrieval data set combining Jilin-1 data with high-resolution and high-precision precipitation data;

[0014] Step 3: Construct an improved PrecipNet algorithm model;

[0015] An attention mechanism module is provided between the encoder and the decoder of the improved PrecipNet algorithm model, enabling the PrecipNet algorithm model to focus on the cloud cluster area and improve the accuracy of precipitation retrieval;

[0016] The improved PrecipNet algorithm model uses an LSTM structure to process the correlation between multi-temporal data, extracts important features between multi-temporal cloud images, enables the improved PrecipNet algorithm model to capture the temporal evolution characteristics of precipitation, and improves the model's prediction ability for continuous precipitation events;

[0017] A multi-scale feature fusion module is provided in the improved PrecipNet algorithm model, enabling the improved PrecipNet algorithm model to simultaneously learn the macroscopic and microscopic features of cloud clusters and improve the accuracy of precipitation retrieval;

[0018] Step 4: Precipitation retrieval;

[0019] Perform multi-scale precipitation retrieval from large areas to small areas;

[0020] Step 5: Precision analysis;

[0021] Introduce multiple meteorological verification indicators for precision analysis;

[0022] Step 6: Model comparison and analysis;

[0023] Compare the improved PrecipNet algorithm model with mainstream algorithm models to verify the performance advantages of the improved PrecipNet algorithm model in the precipitation inversion task;

[0024] Step 7: Ablation experiment;

[0025] Compare the model performance under different module combinations to verify the contributions of each module in the improved PrecipNet algorithm model;

[0026] Step 8: Case analysis;

[0027] Select strong precipitation processes for case analysis to verify the performance in actual precipitation events.

[0028] In the above technical solution, in step 1,

[0029] Use the high-resolution visible light data of Jilin-1 high-resolution optical satellite for refined precipitation inversion. Specifically: First, perform radiometric calibration to convert the original data into surface reflectance and eliminate the sensor response differences; then perform geometric correction to eliminate the geometric distortion of the image and ensure the spatial accuracy of the image; finally, perform orthorectification to eliminate the image deformation caused by terrain undulation and improve the geometric consistency of the image;

[0030] Fuse the ERA5 precipitation data with the ground radar station observation data to construct a high-resolution and high-precision precipitation dataset. Specifically: First, fuse the ERA5 precipitation data with the ground radar station observation data and use the ground radar station observation data to correct the systematic error of the ERA5 precipitation data; then use the inverse distance weighted interpolation method to interpolate the fused precipitation data onto the grid of the invention area to generate spatially continuous precipitation distribution data; finally, perform quality control on the precipitation data through various outlier detection algorithms to remove outliers and missing values and ensure the reliability of the data.

[0031] In the above technical solution, in step 2,

[0032] The large-scale precipitation inversion dataset combining Himawari-8 data and ERA5 precipitation data is used for large-scale precipitation inversion;

[0033] The small-area refined precipitation inversion dataset combining Jilin-1 data and high-resolution and high-precision precipitation data is used for small-area refined precipitation inversion.

[0034] In the above technical solution, in step 4, precipitation inversion includes two stages:

[0035] In the first stage, the overall characteristics of cloud clusters and the preliminary precipitation distribution are obtained using Himawari-8 data, providing important characteristic inputs for the second stage;

[0036] In the second stage, by combining Jilin-1 data, the spatial resolution and accuracy of precipitation inversion are further improved.

[0037] In the above technical solution, in step 5, the verification indicators for accuracy analysis include: hit rate, false alarm rate, critical success index, and root mean square error.

[0038] In the above technical solution, step 6 includes:

[0039] First, precipitation inversion is performed on Himawari-8 data using different algorithm models;

[0040] Then, high-precision precipitation inversion for small areas is carried out by combining Jilin-1 data;

[0041] Finally, the accuracies on the small-area validation sets are compared.

[0042] In the above technical solution, step 7 includes:

[0043] Testing the basic PrecipNet model, the PrecipNet model with an attention mechanism, the PrecipNet model with an LSTM framework, the PrecipNet model with a multi-scale feature fusion module, and improving the performance of the PrecipNet model.

[0044] In the above technical solution, step 8 specifically includes:

[0045] First, data preprocessing is performed on Himawari-8 data and Jilin-1 data, and the preprocessing includes radiometric calibration, geometric correction, and orthorectification;

[0046] Then, model prediction is carried out, and the preprocessed data is input into the improved PrecipNet algorithm model to generate precipitation inversion results;

[0047] Finally, result verification is carried out, and the prediction results of the improved PrecipNet algorithm model are compared with the observed data of ground radar stations to evaluate the performance of the improved PrecipNet algorithm model.

[0048] The present invention has the following beneficial effects:

[0049] The multi-scale high-precision precipitation inversion method based on the PrecipNet model and multi-source visible light data of the present invention aims to overcome the limitations of traditional numerical forecasting in refined precipitation inversion. By integrating the large-scale visible light data of the Sunflower-8 satellite and the high-resolution visible light data of the Jilin-1 satellite, and introducing the attention mechanism, LSTM framework, multi-scale feature fusion module and progressive training strategy, the accuracy and spatial resolution of precipitation inversion are significantly improved.

[0050] Experimental results show that the method of the present invention can not only effectively combine the spatial coverage and resolution advantages of the two data sources to improve the ability to capture macroscopic distribution and microscopic details, but also the improved PrecipNet model surpasses mainstream algorithms such as DeepLabV3+, U-Net, and Pix2PixGAN in inversion accuracy (RMSE is 0.513, MAE is 0.407) and precipitation event detection capability (POD is 0.85, FAR is 0.45).

[0051] The multi-scale high-precision precipitation inversion method based on the PrecipNet model and multi-source visible light data of the present invention further enhances the generalization ability and refined inversion ability of the model through a two-stage progressive training strategy. Case analysis shows that the improved PrecipNet algorithm model performs well in practical applications, especially in the identification and inversion of small-scale precipitation events, providing strong technical support for disaster warning, water resources management and climate innovation.

[0052] The multi-scale high-precision precipitation inversion method based on the PrecipNet model and multi-source visible light data of the present invention provides a new deep learning multi-source data fusion method, demonstrates a new paradigm for modeling complex meteorological events, and verifies the potential of high-resolution data in small-area precipitation inversion.

[0053] The multi-scale high-precision precipitation inversion method based on the PrecipNet model and multi-source visible light data of the present invention improves the accuracy of disaster warning and the data accuracy of climate change invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0055] Figure 1 This is a flow chart of the multi-scale high-precision precipitation inversion method based on the PrecipNet model and multi-source visible light data.

[0056] Figure 2 This is a schematic diagram of the sunflower-8 combined with the ERA5 precipitation dataset, where (a) is the sunflower cloud image data; (b) is the precipitation data, and (c) is the cloud image features.

[0057] Figure 3 Schematic diagram of precipitation dataset combining Jilin-1 data with high resolution and high precision, where (a)-(c) are large-scale cloud images, precipitation data, and cloud features respectively, (d) is the Jilin-1 cloud image, (e)-(f) correspond to regional precipitation and cloud features respectively, and (g) is the high-precision precipitation inversion result.

[0058] Figure 4 PrecipNet model structure diagram, where (a) is the PrecipNet model structure diagram, (b) is the schematic diagram of the GAM attention mechanism, and (c) is the schematic diagram of the LSTM structure.

[0059] Figure 5 Large-scale precipitation inversion result map of Himawari-8 data, where (a) and (d) are both Himawari-8 cloud images, (b) and (e) are both true precipitation distribution maps, and (c) and (f) are both predicted precipitation distribution maps.

[0060] Figure 6 Fine-scale precipitation inversion result map of Jilin-1 data for a small area, where (a) and (d) are both Jilin-1 cloud images, (b) and (e) are both true precipitation distribution maps, and (c) and (f) are both predicted precipitation distribution maps.

[0061] Figure 7 Precipitation inversion result map at 12:00 on July 15, 2022, where (a) is the Himawari-8 cloud image, (b) is the large-scale true precipitation, (c) is the large-scale precipitation inversion result of Himawari-8, (d) is the Jilin-1 cloud image, (e) is the small-area true precipitation, and (f) is the small-area precipitation inversion result of Jilin-1. Detailed implementation manners

[0062] The inventive concept of the present invention is as follows:

[0063] The multi-scale high-precision precipitation inversion method based on the PrecipNet model and multi-source visible light data of the present invention uses deep learning technology and combines multi-source visible light remote sensing data to achieve multi-scale, high-precision, and fine-scale precipitation inversion, which not only provides important technical support for disaster warning, water resource management, and climate research, but also provides a new example for the application of deep learning in the meteorological field.

[0064] The multi-scale high-precision precipitation inversion method based on the PrecipNet model and multi-source visible light data of the present invention aims at the problem that most existing precipitation inversion methods rely on the data of a single satellite and cannot simultaneously meet the requirements of large-scale coverage and high-precision monitoring of local areas. For the first time, the visible light data of the Himawari-8 and Jilin-1 satellites are combined to make full use of the complementary advantages of the two types of data in terms of spatial coverage and resolution. The Himawari-8 data provides large-scale cloud mass information, while the Jilin-1 data provides high-resolution internal cloud mass structure information, providing a data basis for refined precipitation inversion.

[0065] The multi-scale high-precision precipitation inversion method based on the PrecipNet model and multi-source visible light data of the present invention aims at the problem that existing precipitation inversion models still have deficiencies in feature extraction and fail to make full use of multi-temporal features. Many models only rely on static image features and ignore the time evolution characteristics of precipitation events. The present invention proposes an improved PrecipNet deep learning model, introducing an attention mechanism and an LSTM framework to improve the precipitation inversion accuracy of the model. The attention mechanism enables the model to focus on the key areas of the cloud mass, while the LSTM framework captures the time evolution characteristics of precipitation. And a multi-scale feature fusion module and a progressive training strategy are adopted. The multi-scale feature fusion module can extract the macroscopic and microscopic features of the cloud mass at the same time, while the progressive training strategy gradually improves the refined inversion ability of the model through two-stage training.

[0066] The present invention will be described in detail below with reference to the accompanying drawings.

[0067] The present invention mainly includes steps such as data preprocessing, dataset production, rainfall inversion model training, comparison experiments of different models, ablation experiments, and analysis of specific cases. The technical process is as Figure 1 shown.

[0068] The multi-scale high-precision precipitation inversion method based on the PrecipNet model and multi-source visible light data proposed by the present invention has a flowchart as Figure 1 shown.

[0069] Step 1: Data preprocessing;

[0070] (1) Himawari data

[0071] This invention uses the visible light band data of the Himawari-8 geostationary satellite (hereinafter referred to as Himawari-8) as the basic data set for precipitation satellite data inversion. Himawari-8 has high-precision observation capabilities. Its visible light cloud images show the brightness, texture, and shape of clouds, and can distinguish different types of clouds, such as cirrus clouds, stratus clouds, cumulus clouds, etc. Cloud clusters of different types and structures exhibit different texture characteristics. In addition, the visible light cloud images can clearly show the shielding range and movement path of cloud clusters, help infer the changes in precipitation areas, and track the dynamic changes of cloud systems through continuous observations, evaluate the possibility and development trend of precipitation, which is of great significance for weather forecasting and climate monitoring.

[0072] (2) Jilin-1 data

[0073] The Jilin-1 satellite constellation is the first commercially remote sensing satellite constellation independently developed in China and is operated by Changguang Satellite Technology Co., Ltd. The Jilin-1 high-resolution optical satellite (Gaofen-03 satellite) used in this invention is an important part of the Jilin-1 constellation, with a spatial resolution of 0.5 meters, capable of clearly depicting the fine structures of the earth's surface and cloud clusters. This high-resolution feature gives it a unique advantage in fine-scale precipitation inversion in small areas.

[0074] In this invention, the Jilin-1 high-resolution visible light data is mainly used for fine-scale precipitation inversion. The Jilin-1 images in the invention area are first radiometrically calibrated to convert the original data into surface reflectance and eliminate the sensor response differences. Then geometric correction is carried out to eliminate the geometric distortion of the images and ensure the spatial accuracy of the images. Finally, orthorectification is carried out to eliminate the image deformation caused by terrain undulation and improve the geometric consistency of the images.

[0075] (3) Precipitation data

[0076] Due to the different resolutions of visible light images, this invention uses two types of precipitation data, namely ERA5 precipitation data and high-resolution and high-precision precipitation data synthesized based on ERA5 precipitation data. ERA5 precipitation data is the fifth-generation global atmospheric reanalysis data set released by the European Centre for Medium-Range Weather Forecasts (ECMWF), with relatively high spatio-temporal resolutions. Its horizontal resolution is 31 kilometers and it provides a time resolution of every hour. This resolution enables ERA5 precipitation data to capture many important weather phenomena and climate change trends, providing rich detailed information for the invention.

[0077] For high-resolution precipitation data, the present invention fuses ERA5 rainfall data with ground radar site observation data to construct a precipitation dataset with high resolution and high precision. First, the ERA5 rainfall data is fused with the ground radar site observation data, and the systematic error of the ERA5 precipitation data is corrected using the ground radar site observation data. Then, the inverse distance weighted interpolation method (IDW) is used to interpolate the fused precipitation data onto the grid of the invention area to generate spatially continuous precipitation distribution data. Finally, quality control is performed on the precipitation data through a variety of outlier detection algorithms to remove outliers and missing values, ensuring the reliability of the data. The finally generated precipitation dataset features high spatio-temporal resolution, high precision, and wide application. The time resolution is 1 hour, and the spatial resolution is 10 m, which can meet the needs of fine-scale precipitation inversion in small areas.

[0078] (4) Experimental area

[0079] The invention area is selected as the central and eastern regions of China and the Pacific region (60°N - 60°S, 80°W - 160°W), covering the eastern coastal areas of China and some waters in the western Pacific. The geographical scope of the invention area is vast, including provinces and cities such as Jiangsu, Zhejiang, Anhui, and Shanghai along the eastern coast of China, as well as some waters in the Pacific Ocean. The land area of this region accounts for about two-thirds, and most areas belong to the southeast monsoon climate zone, with significant climate characteristics and complex spatio-temporal precipitation distribution.

[0080] Step 2: Produce a multi-scale precipitation inversion dataset;

[0081] The dataset of the present invention consists of two parts, which are designed for precipitation inversion requirements at different spatial scales, forming a multi-scale precipitation dataset that can fully reflect the spatio-temporal distribution characteristics of precipitation. The specific composition of the multi-scale precipitation inversion dataset is as follows:

[0082] (1) Himawari-8 data combined with ERA5 precipitation data

[0083] This dataset is mainly used for large-scale precipitation inversion, combining the high spatio-temporal resolution visible light data of the Himawari-8 satellite and the ERA5 reanalysis precipitation data. The input data is the hourly Himawari-8 visible light data, including three visible light bands of 0.46 μm (blue light), 0.51 μm (green light), and 0.64 μm (red light), with a time resolution of 10 minutes and a spatial resolution interpolated to 5 km, forming a total of 18-band data. The output data is: hourly ERA5 precipitation data and cloud image features. The ERA5 reanalysis precipitation data is interpolated to a 5 km spatial resolution as the target data for large-scale precipitation inversion. The characteristic data of the cloud image comes from the second-level product inverted by Himawari-8, mainly reflecting the cloud optical thickness and effective particle radius. The schematic diagram of the large-scale precipitation inversion dataset is as Figure 2 shown.

[0084] (2) The Jilin-1 data is combined with high-resolution and high-precision precipitation data

[0085] This dataset is mainly used for fine-scale precipitation inversion in small areas, combining high-resolution visible light data from the Jilin-1 satellite and ground radar station observation data. The input data is the hourly high-resolution visible light data of Jilin-1 and the cloud cluster characteristics and large-scale precipitation information retrieved from Himawari data in the corresponding area. The spatial resolution is unified to 10 meters through upscaling and downscaling. The output data is high-resolution precipitation data obtained by fusing ground radar station observation data and ERA5 precipitation data. A schematic diagram of the fine-scale precipitation inversion dataset is as Figure 3 shown.

[0086] (3) Dataset division and augmentation

[0087] The dataset consists of two parts, namely the large-scale precipitation inversion dataset of Himawari-8 data combined with ERA5 precipitation data, and the small-area fine-scale precipitation inversion dataset of Jilin-1 data combined with high-resolution and high-precision precipitation data. The experimental data is from 2020 to 2024. Due to the different sample numbers and characteristics of the two parts of the dataset, different data division and augmentation strategies are adopted in the present invention.

[0088] For the Himawari-8 data combined with ERA5 precipitation data, this dataset contains more than 3,000 samples, and the data volume is relatively sufficient, so there is no need for data augmentation. The dataset is divided into a training set, a test set, and a validation set according to a ratio of 3:1:1. Among them, there are 1,800 samples in the training set, 600 samples in the test set, and 600 samples in the validation set (accounting for 20% of the total data). The time resolution of Himawari-8 data is 10 minutes, and 6 time instances of data are generated every hour. Therefore, the time coverage of the dataset is wide, and it can fully reflect the spatio-temporal variation characteristics of precipitation. This dataset is mainly used for the training and validation of large-scale precipitation inversion models.

[0089] For the Jilin-1 data combined with high-resolution and high-precision precipitation data, this dataset contains 500 samples, and the data volume is relatively small. Therefore, data augmentation methods need to be used to expand it to improve the generalization ability of the model. To expand the dataset, the present invention adopts methods of random rotation, random flipping, and random cropping. After data augmentation, there are 1,500 samples in the training set, 200 samples in the test set, and 200 samples in the validation set.

[0090] Step 3: Construct an improved PrecipNet algorithm model;

[0091] PrecipNet is a deep learning model specifically designed for precipitation retrieval. Based on the U-Net architecture and with multiple innovative optimizations, this model enhances its ability to identify and predict different types of precipitation events by introducing specific mechanisms and technologies.

[0092] The structural diagram of the improved PrecipNet algorithm model of the present invention is as shown in Figure 4 the following figure. It mainly includes the following innovations:

[0093] (1) Introduction of the GAM attention mechanism. The attention mechanism enables the model to focus more on important regions in the image, thereby improving the model's feature extraction ability and prediction accuracy. In the present invention, an attention mechanism module is added between the encoder and decoder of the improved PrecipNet algorithm model, enabling the model to focus more on the cloud cluster region, thus improving the accuracy of precipitation retrieval. The attention mechanism focuses on basic features and suppresses unnecessary features, thereby enhancing the model's feature extraction ability. Among them, the channel attention sub-module maintains three-dimensional information through 3D arrangement and enhances the dependence between cross-dimensional channels and space through dual MLP (Multi-Layer Perceptron), while the spatial attention sub-module focuses on spatial information through two convolutions. The GAM attention mechanism integrates channel and spatial attention together, taking into account the interaction between channels and space, retaining more feature information, and improving the performance of the model. In the present invention, the GAM attention mechanism is used to enable the model to focus on precipitation cloud clusters and reduce the interference of other redundant features.

[0094] (2) Introduction of the LSTM framework: LSTM is a recurrent neural network that can effectively process time series data. In the present invention, the LSTM framework is added to the improved PrecipNet algorithm model, enabling the model to capture the temporal evolution characteristics of precipitation, thereby improving the model's prediction ability for continuous precipitation events. With the help of the LSTM structure, the characteristics between multiple time stages can be analyzed. By transmitting the current input value and the hidden layer state of the previous step, LSTM can adjust the hidden layer state and storage unit in real time, and then grasp the time order of sequence data. In the present invention, the LSTM structure is mainly used to process the correlation between multi-temporal data and extract important features between multi-temporal cloud images, which serves as an important support for the precipitation retrieval model.

[0095] (3) Use of a multi-scale feature fusion module. The multi-scale feature fusion module can fuse feature information of different scales, thereby improving the model's adaptability to precipitation events of different scales. In the present invention, a multi-scale feature fusion module is added to the improved PrecipNet algorithm model, enabling the model to simultaneously learn the macroscopic and microscopic features of cloud clusters, thus improving the accuracy of precipitation retrieval.

[0096] Step 4: Precipitation retrieval;

[0097] The experimental process of the present invention realizes multi-scale precipitation inversion from large areas to small regions through the collaborative inversion of two stages. In the first stage, the Himawari-8 data is used to obtain the overall characteristics of cloud clusters and the preliminary precipitation distribution, providing important characteristic inputs for the second stage. In the second stage, combined with the Jilin-1 data, the spatial resolution and accuracy of precipitation inversion are further improved. Through the application of multi-source data fusion and deep learning technology, the experimental process can effectively support high-precision precipitation inversion in large areas and small regions, providing important technical support for inventions and applications in related fields.

[0098] (1) Large-area precipitation inversion based on Himawari-8 data

[0099] First, input the visible light cloud map data of the Himawari-8 satellite for 1 hour, with a total of 18 bands. Then use the improved PrecipNet algorithm model to extract the cloud cluster characteristics in the Himawari-8 data, including cloud top height, cloud optical thickness, effective radius of cloud particles, etc. Through multi-band fusion and multi-scale feature extraction, the overall characteristics of cloud clusters are obtained. Based on the extracted cloud cluster characteristics, use the improved PrecipNet algorithm model to perform large-area precipitation inversion and generate the precipitation distribution map for that hour. The spatial resolution of the inversion result is 5 km, and the time resolution is 1 hour. Figure 5 It is part of the large-area precipitation inversion results.

[0100] (2) Fine-scale precipitation inversion in small areas based on Jilin-1 data

[0101] First, input the high-resolution visible light data of Jilin-1 with a spatial resolution of 10 meters, as well as the cloud cluster characteristics and precipitation distribution map obtained from the inversion in the first stage cropped according to the coverage area of the Jilin-1 satellite. Then fuse the cloud cluster characteristics obtained from the Himawari-8 data inversion with the high-resolution data of Jilin-1 to generate a multi-scale and multi-resolution feature map. Use the attention mechanism and multi-scale feature fusion module to enhance the model's ability to extract key regions and fine features. Finally, based on the fused features, use the improved PrecipNet algorithm model to perform fine-scale precipitation inversion in small areas and generate a high-precision precipitation distribution map. The spatial resolution of the inversion result is 10 meters, and the time resolution is 1 hour. Figure 6 It is part of the high-precision precipitation inversion results.

[0102] Step 5: Accuracy analysis;

[0103] The present invention introduces multiple verification indicators commonly used in meteorology, mainly including the probability of detection (POD), false alarm rate (FAR), critical success index (CSI), and root mean square error (RMSE). The calculation methods of the adopted indicators are shown in Table 1.

[0104] Table 1 Specific Meanings of Precipitation Indicators

[0105]

[0106]

[0107] Among them, a means that there is rain in the model inversion and rain in the observation; b means that the model inversion is clear sky and there is rain in the observation; c is that there is rain in the model inversion and clear sky in the observation; d is that the model inversion is clear sky and clear sky in the observation.

[0108] Among them, POD represents the possibility that the model correctly identifies precipitation pixel points when precipitation is actually observed. FAR shows the error rate of the model in identifying precipitation among the simulated precipitation pixel points. CSI is the proportion of precipitation pixel points among the correctly classified pixel points when removing the correctly classified non-precipitation points. In addition, the root mean square error RMSE is used in the present invention to measure the deviation between the inversion value and the true value.

[0109] In order to evaluate the precipitation inversion accuracy of the improved PrecipNet algorithm model, the present invention compared and analyzed the model inversion results of the validation set with the label data, and listed the accuracy of the large-scale inversion model and the accuracy of the small-area refined precipitation inversion combined with the data of Jilin-1 satellite respectively.

[0110] Through the two-stage inversion results, the advantages of Jilin-1 satellite data in small-area refined precipitation inversion can be verified. The Himawari-8 satellite data has high temporal resolution and large-area coverage ability, and can achieve high-dynamic and large-area precipitation inversion. However, its spatial resolution is relatively low (5 km), making it difficult to meet the refined inversion requirements of small-scale precipitation events. The Jilin-1 satellite data provides a spatial resolution of meters, which can clearly depict the internal structure of cloud clusters and surface details. Through comparison, the significant contribution of Jilin-1 satellite data in improving spatial resolution and inversion accuracy can be intuitively reflected. The experimental results are shown in Table 2.

[0111] Table 2 Accuracy Evaluation

[0112] Model Spatial resolution RMSE MAE POD FAR Large-scale inversion 5km 0.741 0.579 0.76 0.55 Refined inversion 10m 0.513 0.407 0.85 0.45

[0113] The experimental results show that in terms of spatial resolution: the spatial resolution of the large-scale inversion using Himawari-8 satellite data is 5 km, and the resolution is 10 m after the refined inversion combined with Jilin-1 satellite data, significantly improving the ability to depict small-scale precipitation events. In terms of inversion accuracy: RMSE decreased from 0.741 to 0.513; MAE decreased from 0.579 to 0.407; POD increased from 0.76 to 0.85, an increase of 5%; FAR decreased from 0.55 to 0.45. The introduction of Jilin-1 satellite data significantly improves the spatial resolution and inversion accuracy of the model, especially in small-area refined precipitation inversion.

[0114] Step 6: Model comparison and analysis;

[0115] To verify the superiority of the improved PrecipNet algorithm model in algorithm design, the present invention compares it with several mainstream algorithm models (such as DeepLabV3+, U-Net, Pix2PixGAN, etc.). First, different algorithm models are used to perform precipitation inversion on Himawari-8 data, then combined with Jilin-1 data for small-area high-precision precipitation inversion, and finally the accuracy on the small-area validation set is compared. DeepLabV3+ is a semantic segmentation model based on dilated convolution, U-Net is a classic encoder-decoder structure model, and Pix2PixGAN is an image conversion model based on generative adversarial network. These models perform excellently in image segmentation and prediction tasks, but have limitations in processing multi-source data and multi-scale feature fusion. Through comparison, the performance advantages of the improved PrecipNet algorithm model in the precipitation inversion task can be highlighted. The experimental results are shown in Table 3.

[0116] Table 3 Accuracy comparison

[0117] Model RMSE MAE POD FAR DeepLabV3+ 0.680 0.650 0.72 0.59 U-Net 0.760 0.735 0.69 0.63 Pix2PixGAN 0.590 0.530 0.76 0.52 PrecipNet 0.513 0.407 0.85 0.45

[0118] The experimental results show that both the RMSE and MAE of the improved PrecipNet algorithm model are lower than those of other models, indicating higher inversion accuracy. And the POD is the highest (0.85) and the FAR is the lowest (0.45), indicating higher hit rate and lower false alarm rate in precipitation event detection. The improved PrecipNet algorithm model significantly improves the model performance by introducing the attention mechanism, LSTM framework and multi-scale feature fusion module, which is better than other mainstream algorithm models.

[0119] Step 7: Ablation experiment;

[0120] To verify the contributions of each module in the improved PrecipNet algorithm model, the present invention designs an ablation experiment to compare the model performance under different module combinations. Specifically, the performance of the basic PrecipNet model, the PrecipNet model with attention mechanism added, the PrecipNet model with LSTM framework added, the PrecipNet model with multi-scale feature fusion module added, and the improved PrecipNet algorithm model are respectively tested. Through the ablation experiment, the specific contributions of each module (such as attention mechanism, LSTM framework, multi-scale feature fusion module) to the improvement of model performance can be quantified, so as to verify the rationality and effectiveness of the model design. The experimental results are shown in Table 4.

[0121] Table 4 Ablation experiment

[0122] Model configuration RMSE MAE POD FAR Basic U-NET model 0.760 0.735 0.69 0.63 + Attention mechanism 0.710 0.682 0.72 0.58 + LSTM framework 0.680 0.630 0.73 0.54 + Feature fusion module 0.600 0.515 0.79 0.50 PrecipNet model 0.513 0.407 0.85 0.45

[0123] The experimental results show that after introducing the attention mechanism, the RMSE decreases from 0.760 to 0.710, the MAE decreases from 0.735 to 0.682, the POD increases from 0.69 to 0.72, and the FAR decreases from 0.63 to 0.58. This indicates that the attention mechanism can effectively improve the model's ability to extract key features. After introducing the LSTM framework, the RMSE further decreases from 0.710 to 0.680, the MAE decreases from 0.682 to 0.630, the POD increases from 0.72 to 0.73, and the FAR decreases from 0.58 to 0.54, indicating that the LSTM framework can capture the temporal evolution characteristics of precipitation and improve the model's dynamic modeling ability. After introducing the multi-scale feature fusion module, the RMSE decreases from 0.680 to 0.600, the MAE decreases from 0.630 to 0.515, the POD increases from 0.73 to 0.79, and the FAR decreases from 0.54 to 0.50, indicating that the multi-scale feature fusion module can improve the model's adaptability to precipitation events of different scales. The performance of the complete improved PrecipNet algorithm model is optimal, and the RMSE, MAE, POD, and FAR all reach the best values, which are 0.513, 0.407, 0.85, and 0.45 respectively, indicating that the synergistic effect of each module significantly improves the overall performance of the model.

[0124] Step 8: Case analysis;

[0125] To further verify the performance of the precipitation retrieval method proposed in the present invention in actual precipitation events, the present invention selects a typical heavy precipitation process for case analysis. This precipitation event occurred on September 14, 2022, covering the central and eastern regions of China, including Xiamen, Jiangsu, Zhejiang, Anhui, Shanghai and other places (26.76°-37°N, 114°-124.24°E). This precipitation process was triggered by Typhoon Muifa, with large precipitation intensity, hourly precipitation exceeding 50 mm in some local areas, uneven spatial distribution, and obvious zonal distribution characteristics, which is an ideal case for verifying the model performance.

[0126] Apply the improved PrecipNet algorithm model to this precipitation event to generate high-precision precipitation retrieval results. First, perform data preprocessing on the Himawari-8 data and Jilin-1 data, including radiometric calibration, geometric correction, and orthorectification. Then, perform model prediction, input the preprocessed data into the improved PrecipNet algorithm model to generate precipitation retrieval results. Finally, perform result verification, compare the model prediction results with the ground radar station observation data to evaluate the performance of the model.

[0127] Figure 7Shows the precipitation inversion results at 12:00 on July 15, 2022, including Himawari-8 data, Jilin-1 data, large-scale inversion results, small-area refined inversion results, and precipitation observation data at two scales. Himawari-8 data can provide large-scale cloud cluster information and can reflect the overall distribution characteristics of precipitation, but the spatial resolution is low and it is difficult to capture small-scale precipitation events. Jilin-1 data provides high-resolution internal cloud structure information and can clearly depict the fine characteristics of precipitation, but the time resolution is low and it is difficult to reflect the dynamic changes of precipitation. The improved PrecipNet algorithm model results combine the advantages of Himawari-8 data and Jilin-1 data, can not only reflect the overall distribution characteristics of precipitation, but also capture small-scale precipitation events, and both the spatial resolution and inversion accuracy are significantly improved. By comparison, it can be seen that the prediction results of the improved PrecipNet algorithm model have a high consistency with the real precipitation data, especially in the identification and inversion of small-scale precipitation events. Through case analysis, the present invention verifies the performance of the algorithm in actual precipitation events. The model can accurately reflect the spatial distribution and intensity changes of precipitation, especially in the identification and inversion of small-scale precipitation events. The case analysis results show that the improved PrecipNet algorithm model has high inversion accuracy and reliability, and can provide important technical support for disaster warning, water resource management and climate research.

[0128] The multi-scale high-precision precipitation inversion method based on the PrecipNet model and multi-source visible light data of the present invention aims to overcome the limitations of traditional numerical forecasting in refined precipitation inversion. By integrating the large-scale visible light data of the Himawari-8 satellite and the high-resolution visible light data of the Jilin-1 satellite, and introducing an attention mechanism, an LSTM framework, a multi-scale feature fusion module and a progressive training strategy, the accuracy and spatial resolution of precipitation inversion are significantly improved.

[0129] The experimental results show that the method of the present invention can not only effectively combine the spatial coverage and resolution advantages of the two data sources and improve the ability to capture macroscopic distribution and microscopic details, but also the improved PrecipNet model exceeds mainstream algorithms such as DeepLabV3+, U-Net, and Pix2PixGAN in terms of inversion accuracy (RMSE is 0.513, MAE is 0.407) and precipitation event detection ability (POD is 0.85, FAR is 0.45).

[0130] The multi-scale high-precision precipitation inversion method based on the PrecipNet model and multi-source visible light data of the present invention uses a two-stage progressive training strategy to further enhance the generalization ability and refined inversion ability of the model. Case analysis shows that the improved PrecipNet algorithm model performs excellently in practical applications, especially good at identifying and inverting small-scale precipitation events, providing strong technical support for disaster warning, water resource management and climate research.

[0131] The multi-scale high-precision precipitation inversion method based on the PrecipNet model and multi-source visible light data of the present invention provides a new deep learning multi-source data fusion method, demonstrates a new paradigm for complex meteorological event modeling, and verifies the potential of high-resolution data in small-area precipitation inversion.

[0132] The multi-scale high-precision precipitation inversion method based on the PrecipNet model and multi-source visible light data of the present invention improves the accuracy of disaster warning and the data accuracy of climate research.

[0133] Obviously, the above embodiments are only examples for clear illustration and not limitations on the implementation manners. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation manners here. And the obvious changes or modifications derived therefrom are still within the protection scope of the present invention.

Claims

1. A multi-scale high-precision precipitation inversion method based on the PrecipNet model and multi-source visible light data, characterized in that, It includes the following steps: Step 1: Data preprocessing; The visible light band of the Himawari-8 geostationary satellite data is used as the basic data set for precipitation satellite data inversion; The high-resolution visible light data of the Jilin-1 high-resolution optical satellite are used for refined precipitation inversion; The ERA5 precipitation data is fused with the ground radar station observation data to construct a high-resolution and high-precision precipitation data set; Step 2: Produce a multi-scale precipitation inversion data set; The multi-scale precipitation inversion data set includes: a large-scale precipitation inversion data set combining Himawari-8 data with ERA5 precipitation data, and a small-area refined precipitation inversion data set combining Jilin-1 data with high-resolution and high-precision precipitation data; Step 3: Construct an improved PrecipNet algorithm model; An attention mechanism module is set between the encoder and the decoder of the improved PrecipNet algorithm model, enabling the PrecipNet algorithm model to focus on the cloud cluster area and improving the accuracy of precipitation inversion; The improved PrecipNet algorithm model uses an LSTM structure to process the correlation between multi-temporal data, extracts important features between multi-temporal cloud images, enables the improved PrecipNet algorithm model to capture the temporal evolution characteristics of precipitation, and improves the model's prediction ability for continuous precipitation events; A multi-scale feature fusion module is set in the improved PrecipNet algorithm model, enabling the improved PrecipNet algorithm model to simultaneously learn the macroscopic and microscopic features of cloud clusters and improving the accuracy of precipitation inversion; Step 4: Precipitation inversion; Perform multi-scale precipitation inversion from large areas to small areas; Step 5: Accuracy analysis; Introduce multiple meteorological verification indicators for accuracy analysis; Step 6: Model comparison analysis; Compare the improved PrecipNet algorithm model with the mainstream algorithm models to verify the performance advantages of the improved PrecipNet algorithm model in the precipitation inversion task; Step 7: Ablation experiment; Compare the model performance under different module combinations to verify the contributions of each module in the improved PrecipNet algorithm model; Step 8: Case analysis; Select a heavy precipitation process for case analysis to verify the performance in actual precipitation events.

2. The multi-scale high-precision precipitation retrieval method based on the PrecipNet model and multi-source visible light data according to claim 1, wherein, In Step 1, The high-resolution visible light data of the Jilin-1 high-resolution optical satellite are used for refined precipitation inversion. Specifically: First, perform radiometric calibration to convert the original data into surface reflectance and eliminate the sensor response differences; then perform geometric correction to eliminate the geometric distortion of the image and ensure the spatial accuracy of the image; finally, perform orthorectification to eliminate the image deformation caused by terrain undulation and improve the geometric consistency of the image; The ERA5 precipitation data is fused with the ground radar station observation data to construct a high-resolution and high-precision precipitation dataset. Specifically: First, the ERA5 precipitation data is fused with the ground radar station observation data, and the systematic error of the ERA5 precipitation data is corrected using the ground radar station observation data; then, the inverse distance weighted interpolation method is used to interpolate the fused precipitation data onto the grid of the invention area to generate spatially continuous precipitation distribution data; finally, quality control is performed on the precipitation data through various outlier detection algorithms to remove outliers and missing values to ensure the reliability of the data.

3. The multi-scale high-precision precipitation inversion method based on the PrecipNet model and multi-source visible light data according to claim 1, characterized in that, In step 2, The large-scale precipitation inversion dataset combining Himawari-8 data and ERA5 precipitation data is used for large-scale precipitation inversion; The small-area refined precipitation inversion dataset combining Jilin-1 data and high-resolution and high-precision precipitation data is used for small-area refined precipitation inversion.

4. The multi-scale high-precision precipitation inversion method based on the PrecipNet model and multi-source visible light data according to claim 1, characterized in that In step 4, precipitation inversion includes two stages: In the first stage, the overall characteristics of cloud clusters and the preliminary precipitation distribution are obtained using Himawari-8 data, providing important feature inputs for the second stage; In the second stage, combined with Jilin-1 data, the spatial resolution and accuracy of precipitation inversion are further improved.

5. The multi-scale high-precision precipitation inversion method based on the PrecipNet model and multi-source visible light data according to claim 1, wherein, In step 5, the verification indicators for accuracy analysis include: hit rate, false alarm rate, critical success index, and root mean square error.

6. The multi-scale high-precision precipitation inversion method based on the PrecipNet model and multi-source visible light data according to claim 1, characterized in that Step 6 includes: First, different algorithm models are used to perform precipitation inversion on Himawari-8 data; Then, combined with Jilin-1 data, high-precision precipitation inversion for small areas is performed; Finally, the accuracy on the small-area validation set is compared.

7. The multi-scale high-precision precipitation inversion method based on the PrecipNet model and multi-source visible light data according to claim 1, characterized in that, Step 7 includes: Testing the performance of the basic PrecipNet model, the PrecipNet model with an attention mechanism, the PrecipNet model with an LSTM structure, the PrecipNet model with a multi-scale feature fusion module, and the improved PrecipNet algorithm model.

8. The multi-scale high-precision precipitation inversion method based on the PrecipNet model and multi-source visible light data according to claim 1, characterized in that, Step 8 specifically includes: First, data preprocessing is performed on Himawari-8 data and Jilin-1 data. The preprocessing includes radiometric calibration, geometric correction, and orthorectification; Then, model prediction is performed. The preprocessed data is input into the improved PrecipNet algorithm model to generate precipitation inversion results; Finally, result verification is performed. The prediction results of the improved PrecipNet algorithm model are compared with the ground radar station observation data to evaluate the performance of the improved PrecipNet algorithm model.

Images