CN113269805B - Rainfall event guided remote sensing rainfall inversion training sample self-adaptive selection method - Google Patents

Abstract

The invention relates to an adaptive selection method of remote sensing precipitation inversion training samples guided by precipitation events. The method includes: obtaining a precipitation binary map, and using a classification algorithm to divide the precipitation area into a connected domain to obtain a plurality of precipitation connected domains; for any precipitation connected domain, according to the number of pixels in the precipitation connected domain, the number of pixels in the precipitation area, and the prediction of the precipitation connected domain. Set the number of precipitation samples to obtain the sampling number of the precipitation connected area; sample the pixels of the precipitation connected area according to the sampling number of the precipitation connected area and the number of pixels of the precipitation connected area to obtain the precipitation samples of the precipitation connected area; Random sampling is used to obtain non-precipitation samples; precipitation samples and non-precipitation samples in all precipitation connected regions are determined as training samples for precipitation inversion. The invention can satisfy the representativeness of the training samples and the universality of the inversion model at the same time, so as to improve the inversion precision of the inversion model.

Description

An adaptive selection method of training samples for remote sensing precipitation inversion guided by precipitation events

technical field

The invention relates to the field of precipitation inversion, in particular to a method and system for adaptive selection of training samples for remote sensing precipitation inversion guided by precipitation events.

Background technique

Precipitation plays an important role in the fields of hydrology, meteorology, ecology, and agricultural research, and is especially one of the main drivers of global-scale material and energy exchange. Therefore, it is very important to use satellite data for precipitation inversion to obtain precipitation data with high precision and high spatial and temporal resolution. Precipitation inversion is a mathematical process of estimating precipitation values using satellite observation data. This method often uses model methods, such as fitting formulas for parameters, machine learning, etc. Regardless of whether it is a fitting formula for parameters or machine learning, the model method needs to input the measured precipitation value and the eigenvalue (often cloud top brightness temperature) observed by the satellite as the training data. After the parameters are determined by a certain method, the trained Models can then be used for precipitation inversion, ie inputting new observations to obtain inversion estimates. Then the selection of training samples in this process has a very direct impact on the final inversion result, that is, the strategy of sample selection can feel the final training result of the model.

An excellent sample selection strategy should satisfy both the representativeness of the sample and the universality of the model. The so-called representativeness means that various special situations can be covered in the sample selection process, and each type of situation has a sufficient number of samples, so that the model can not lack effective information during the training process. Universality means that the model has high stability for different or new inputs, rather than being limited to the types and situations provided by the sample set.

The existing sample selection strategies mainly include three types: random selection method, numerical distribution control method and sample selection method based on similarity (consistency) judgment.

The random selection method is to select all samples according to a certain random number principle, or according to a certain interval step size in the case of shuffling the order, to ensure the randomness of sample selection. The random selection method is simple and easy to implement. The selected samples have certain universality when the number of selected samples is sufficient. That is, by selecting samples for model training, the obtained training model can cope with most situations and can identify most of the feature. However, it lacks a certain degree of representativeness, because random selection often selects a large number of duplicate feature samples or invalid feature samples, resulting in a sample set with the same number of samples, the ability to represent the overall event is greatly reduced, resulting in greater data redundancy. .

The numerical distribution control method is to select a certain number of samples in each numerical area and add them to the sample set by dividing the numerical area according to the numerical distribution law in the overall event set. The sample set selected in this way can have a strong representativeness. Since the samples come from different specified numerical ranges, there is a guarantee of a certain number of samples for each numerical range, so that the sample set can represent most of the occurrences. However, since the division of the numerical range and the allocation of the quota are a priori, this is a supervised sample selection, and human factors are inevitably added to the sample set, which will reduce a certain degree of universality, resulting in the final trained model can only be Reflects the artificially considered sample situation.

The sample selection method based on similarity (consistency) judgment is a more complex method. This kind of method usually judges the similarity between samples, and then refines the classification of the sample set through continuous iterative methods, so that the final samples have higher consistency. This type of method is more complicated, and there are many definitions of similarity, and there are more options for iterative methods. But all in all, it can increase the representativeness of the sample, that is, it can represent the occurrence and characteristics of the overall event through a smaller number of sample sets, and greatly reduce the proportion of invalid samples and duplicate samples. However, due to the elimination of inconsistent samples, it will lead to the loss of samples for complex situations, and it is easy to overfit in the process of model training, which ultimately leads to a decrease in the universality of the model.

To sum up, the existing sample selection strategies cannot satisfy the representativeness of the sample and the universality of the model at the same time.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide an adaptive selection method of training samples for remote sensing precipitation inversion guided by precipitation events, which can satisfy the representativeness of training samples and the universality of inversion models at the same time, so as to improve the inversion accuracy of inversion models.

For achieving the above object, the present invention provides the following scheme:

An adaptive selection method of training samples for remote sensing precipitation inversion guided by precipitation events, comprising:

obtaining a precipitation binary map, where the precipitation binary map includes a precipitation area and a non-precipitation area;

The precipitation area is divided into connected domains by using a classification algorithm to obtain multiple precipitation connected domains;

For any precipitation connected domain, obtain the sampling number of the precipitation connected domain according to the number of pixels of the precipitation connected domain, the pixel number of the precipitation area and the preset number of precipitation samples;

sampling the pixels of the precipitation connected domain according to the sampling quantity of the precipitation connected domain and the number of pixels of the precipitation connected domain to obtain the precipitation sample of the precipitation connected domain;

random sampling of pixels in the non-precipitation area to obtain non-precipitation samples;

Determine the precipitation samples of all precipitation connected domains and the non-precipitation samples as training samples of precipitation retrieval.

Optionally, the obtaining a binary image of precipitation specifically includes:

obtaining a precipitation situation map, where the pixel value of the precipitation situation map is the amount of precipitation, and the pixel position of the precipitation situation map is the geographic location corresponding to the pixel;

Threshold segmentation is performed on the precipitation situation map to obtain a precipitation binary map.

Optionally, before the use of a classification algorithm to divide the precipitation region into connected domains to obtain multiple precipitation connected domains, the method further includes:

Perform image mask processing on the precipitation binary image to obtain a processed precipitation area.

Optionally, the sampling number of the precipitation connected region is obtained according to the number of pixels of the precipitation connected region, the number of pixels of the precipitation region, and the preset number of precipitation samples, specifically:

Calculate the sampling quantity of the _{ith precipitation connected domain according to the formula Mi = (N i /} N _total )*M _total , where Mi is the sampling quantity of the _ith precipitation connected domain, and M _total is the preset precipitation sample quantity; Ni is the number of pixels in the _ith precipitation connected domain, and N _total is the number of pixels in the precipitation area.

Optionally, sampling the pixels of the precipitation connected domain according to the sampling quantity of the precipitation connected domain and the number of pixels of the precipitation connected domain to obtain the precipitation sample of the precipitation connected domain, specifically:

According to the formula S _i ={P _k |k=[(t-0.5)/M _i *N _i ], t∈N ⁺ , 1≤t≤M _i }, the precipitation samples of the i-th precipitation connected domain are obtained, where S _i is the precipitation sample of the i-th precipitation connected domain, P _k is the k-th precipitation pixel in the precipitation sample of the _i -th precipitation connected domain, t is the sample number of the precipitation Number of samples, Ni is the number of pixels in the _i -th precipitation connected domain.

Optionally, the random sampling of pixels in the non-precipitation area to obtain non-precipitation samples specifically includes:

Labeling each pixel in the non-precipitation area according to the position of each pixel in the non-precipitation area to obtain the pixel serial number of each pixel in the non-precipitation area;

The non-precipitation samples are obtained by randomly sampling the pixels in the non-precipitation area according to the preset number of non-precipitation samples and the pixel serial number of each pixel.

An adaptive selection system for remote sensing precipitation retrieval training samples guided by precipitation events, comprising:

an obtaining module, configured to obtain a precipitation binary image, wherein the precipitation binary image includes a precipitation area and a non-precipitation area;

A precipitation connected domain determining module, used for dividing the precipitation region into a connected domain by using a classification algorithm to obtain a plurality of precipitation connected domains;

a quantity determination module, configured to obtain the sampling quantity of the precipitation connected domain according to the pixel count of the precipitation connected domain, the pixel count of the precipitation area and the preset precipitation sample quantity for any precipitation connected domain;

a precipitation sample determination module, configured to sample the pixels of the precipitation connected domain according to the sampling quantity of the precipitation connected domain and the number of pixels of the precipitation connected domain to obtain the precipitation sample of the precipitation connected domain;

a non-precipitation sample determination module, configured to randomly sample the pixels in the non-precipitation area to obtain non-precipitation samples;

The inversion training sample determination module is used to determine the precipitation samples of all precipitation connected domains and the non-precipitation samples as precipitation inversion training samples.

Optionally, the obtaining module includes:

an acquiring unit, configured to acquire a precipitation situation map, where the pixel value of the precipitation situation map is the amount of precipitation, and the pixel position of the precipitation situation map is the geographic location corresponding to the pixel;

The binary map determining unit is configured to perform threshold segmentation on the precipitation situation map to obtain a precipitation binary map.

Optionally, the precipitation event-guided remote sensing precipitation inversion training sample adaptive selection system further includes: a processing module configured to perform image mask processing on the precipitation binary image to obtain a processed precipitation area.

Optionally, the non-precipitation sample determination module includes:

a serial number determining unit, configured to label each pixel in the non-precipitation area according to the position of each pixel in the non-precipitation area to obtain the pixel serial number of each pixel in the non-precipitation area;

A non-precipitation sample determination unit, configured to randomly sample the pixels in the non-precipitation area according to the preset number of non-precipitation samples and the pixel serial number of each pixel to obtain non-precipitation samples.

According to the specific embodiment provided by the present invention, the present invention discloses the following technical effects: the present invention obtains the sampling number of each precipitation connected region according to the number of pixels in each precipitation connected region, the number of pixels in the precipitation region and the preset number of precipitation samples, so that each All precipitation connected domains can be sampled, and because each precipitation connected domain is covered, the number of samples extracted does not have a lot of repetition and redundancy, which can greatly compensate for the lack of representativeness of the random selection method. and the number of pixels to sample the pixels of the precipitation connected domain, so that the sample selection will not be too concentrated and have sufficient variability, which can effectively represent the precipitation of different situations and different characteristics, so as to improve the universality of the model trained according to it. Avoid overfitting.

Description of drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings required in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some of the present invention. In the embodiments, for those of ordinary skill in the art, other drawings can also be obtained according to these drawings without creative labor.

1 is a flowchart of a method for adaptively selecting training samples for remote sensing precipitation inversion guided by a precipitation event provided by an embodiment of the present invention;

Fig. 2 is the specific flow chart of the experiment using the method of self-adaptive selection of training samples for remote sensing precipitation inversion guided by precipitation events;

Fig. 3 is a comparison diagram of the results of different algorithms in the present invention, Fig. 3 (a) is a sampling result diagram of the random average sample sampling method, Fig. 3 (b) is a sample adaptive selection method based on precipitation events proposed by the present invention. Sampling result graph;

4 is a block diagram of an adaptive selection system for remote sensing precipitation inversion training samples guided by precipitation events provided by an embodiment of the present invention;

5 is a flowchart of a more specific method for adaptively selecting training samples for remote sensing precipitation inversion guided by a precipitation event provided by an embodiment of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

In order to make the above objects, features and advantages of the present invention more clearly understood, the present invention will be described in further detail below with reference to the accompanying drawings and specific embodiments.

The current training sample selection methods mainly include random method, numerical distribution control method and sample selection method based on similarity (consistency) judgment. The representativeness and universality of the models selected from different sample sets have their own emphasis, and their characteristics are listed as follows:

The representativeness of the models under different sample selection strategies is from strong to weak: the sample selection method based on similarity judgment, the numerical distribution control method, and the random selection method; while the universality of the models under different sample selection strategies is from strong to weak: Random selection method, numerical distribution control method, and sample selection method based on similarity judgment; and the degree of influence of human interference on the sample set due to a priori setting is from strong to weak: numerical distribution control method, similarity judgment-based method Sample selection method, random selection method.

The advantage of the random selection method is that it is simple to implement, but it lacks the representativeness of the positive samples; the numerical distribution control method can effectively improve the representativeness of the positive samples, but the supervision is greatly enhanced, and the results are easily affected by the a priori setting parameters. The sample selection method based on similarity judgment can significantly improve the intra-set consistency of the sample set, but the process requires constant iteration, the calculation process is slow, and at the same time, due to the improvement of the intra-set consistency, it may lead to overfitting. The performance is significantly enhanced, while overfitting can significantly reduce the generalizability of the trained model, possibly rendering the model unusable. Models with strong representativeness and weak universality can simulate the situation represented in the sample set better, but for the situation that does not appear in the sample set or occurs less frequently, singular value results will appear, which has a greater impact on the overall simulation results.

The sample determination method provided in this embodiment allocates sample selection quotas to each precipitation event according to actual precipitation events, and finally generates a relatively representative and universal precipitation sample set for training and learning of the precipitation inversion model. This embodiment overcomes the obvious defects of existing sample selection methods, selects suitable samples based on physical facts and phenomena, and finally makes the sample set highly representative and universal, and at the same time gets rid of the heavy dependence on prior parameters, As shown in FIG. 1 , the process of the method for adaptively selecting training samples for remote sensing precipitation inversion guided by precipitation events provided in this embodiment is as follows:

Step 101: Obtain a binary image of precipitation. The precipitation binary map includes a precipitation area and a non-precipitation area.

Step 102: Use a classification algorithm to divide the precipitation region into connected domains to obtain a plurality of precipitation connected domains.

Step 103: For any precipitation connected domain, obtain the sampling quantity of the precipitation connected domain according to the pixel count of the precipitation connected domain, the pixel count of the precipitation region, and the preset precipitation sample quantity.

Step 104: Sampling the pixels of the precipitation connected region according to the sampling quantity of the precipitation connected region and the number of pixels of the precipitation connected region to obtain a precipitation sample of the precipitation connected region.

Step 105: Randomly sample the pixels in the non-precipitation area to obtain non-precipitation samples. For non-precipitation areas, because they are often large in area and connected into pieces, the method is relatively simple, and only needs to be averaged according to the pixel position sequence number.

Step 106: Determine the precipitation samples and the non-precipitation samples in all precipitation connected domains as training samples for precipitation inversion.

The input precipitation situation map needs to be thresholded to generate the distribution map of precipitation events (precipitation binary map), because precipitation prediction often has a large number of small estimated values close to 0, whether such estimated values need to be judged as precipitation events need to pass the threshold value split to achieve. In practical applications, step 101 specifically includes:

A precipitation situation map is obtained, which is in a raster format, the pixel value of the precipitation situation map is the amount of precipitation, and the pixel position of the precipitation situation map is the geographic location corresponding to the pixel.

Threshold segmentation is performed on the precipitation situation map to obtain a precipitation binary map. Here, this method uses the internationally common value (0.1mm/hour) as the threshold value to judge the precipitation event: that is, if the pixel is greater than or equal to the threshold value is marked as 1, it is judged to be precipitation, and similarly, the pixel mark less than the threshold value is marked. If it is 0, it is judged that there is no precipitation.

In practical applications, the use of a classification algorithm to divide the precipitation area into a connected domain to obtain multiple precipitation connected domains is based on the connected domain principle, that is, adjacent and similar pixels will be judged as a subset under the same connected domain, There is no pixel independence in the adjacent connected domains extracted in this way, and the adjacent judgment used is 4-proximity judgment, that is, when a pixel is directly above, directly below, directly left, or directly right of another pixel In this case, these two pixels will be judged as adjacent. Therefore, in this way, a precipitation distribution map will be divided into multiple connected domains, and each connected domain will mark it as a geographically continuous precipitation event.

Dilation and erosion of images is a fundamental step in graphics operations. In practical applications, before the use of a classification algorithm to divide the precipitation region into connected domains to obtain multiple precipitation connected domains, the method further includes:

Image mask processing is performed on the precipitation binary image to obtain a processed precipitation area, which can finally expand or narrow the range of the connected domain. Here, the threshold of graph expansion or erosion is selected as 4 connected domains, and only one operation is performed at the same time, first erosion and then expansion. The purpose of this step is to reduce the influence of a single precipitation pixel and the influence of random noise, so that the whole algorithm focuses on the larger scale. large precipitation events.

In practical applications, the sampling quantity of the precipitation connected domain is obtained according to the pixel count of the precipitation connected domain, the pixel count of the precipitation region, and the preset precipitation sample quantity, specifically:

According to M _i =(N _i /N _total )*M _total formula (1)

Calculate the sampling number of the ith precipitation connected domain, where Mi is the sampling number of the _ith precipitation connected domain, M _total is the preset number of precipitation samples; Ni is the number of pixels of the _ith precipitation connected domain, N _total is the number of pixels in the precipitation area.

In practical applications, sampling the pixels of the precipitation connected domain according to the sampling quantity of the precipitation connected domain and the number of pixels of the precipitation connected domain to obtain the precipitation sample of the precipitation connected domain, specifically:

According to S _i ={P _k |k=[(t-0.5)/M _i *N _i ], t∈N ⁺ , 1≤t≤M _i } Formula (2)

Obtain the precipitation sample of the ith precipitation connected domain, where S _i is the precipitation sample of the ith precipitation connected domain, P _k is the kth precipitation pixel in the precipitation sample of the ith precipitation connected domain, and t is the precipitation pixel of the ith precipitation connected domain. Sample serial number, Mi is the sampling number of the _ith precipitation connected domain, and _Ni is the number of pixels of the ith precipitation connected domain.

In practical applications, the random sampling of pixels in the non-precipitation area to obtain non-precipitation samples specifically includes:

A pixel serial number of each pixel in the non-precipitation area is obtained by labeling each pixel in the non-precipitation area according to the position of each pixel in the non-precipitation area.

The non-precipitation samples are obtained by randomly sampling the pixels in the non-precipitation area according to the preset number of non-precipitation samples and the pixel serial number of each pixel.

As shown in FIG. 2 , this embodiment also provides an experiment on the rainfall with the latitude and longitude of (105.0°-130.0°E, 18.2°-38.0°N) on June 1, 2018 by applying the above method, and finally obtained and preset There are 150 precipitation samples and 450 non-precipitation samples with the same quota. The specific steps are:

Step S210: Perform threshold segmentation on the precipitation map to obtain binary images of the precipitation area and the non-precipitation area.

Step S220: Perform an average sampling selection method according to geographic location for the non-precipitation area.

Step S230: Perform image erosion and image expansion operations on the precipitation area.

Step S240: Divide the precipitation event connected domain for the precipitation area.

Step S250: For the precipitation area that has been divided, perform precipitation quota allocation according to its area. Specifically, the precipitation quota is allocated according to formula (1), and the precipitation samples are selected according to formula (2).

Step S260: Integrate the precipitation event samples and the non-precipitation event samples into a final precipitation sample set.

The effect comparison is shown in Figure 3, in which Figure 3(a) is the result obtained by the common average random sampling method, and Figure 3(b) is obtained by applying the adaptive sample selection method based on precipitation events provided by the present invention. As can be seen from Figure 3, the method provided in this embodiment can replace the original common average random sample sampling method to generate precipitation event characteristics, high representativeness and high universality, and precipitation and precipitation can be adjusted according to the preset quota. The precipitation training sample set for the non-precipitation sample number. This dataset is then fed into a machine learning framework for precipitation inversion.

As shown in FIG. 4 , the present embodiment also provides a system for adaptive selection of training samples for remote sensing precipitation inversion guided by precipitation events corresponding to the above method, and the system includes:

The obtaining module A1 is configured to obtain a precipitation binary map, where the precipitation binary map includes a precipitation area and a non-precipitation area.

The precipitation connected domain determination module A2 is configured to use a classification algorithm to divide the connected domain of the precipitation area to obtain a plurality of precipitation connected domains.

The quantity determination module A3 is configured to, for any precipitation connected domain, obtain the sampling quantity of the precipitation connected domain according to the pixel count of the precipitation connected domain, the pixel count of the precipitation region and the preset precipitation sample quantity.

The precipitation sample determination module A4 is configured to sample the pixels of the precipitation connected area according to the sampling quantity of the precipitation connected area and the number of pixels of the precipitation connected area, so as to obtain the precipitation sample of the precipitation connected area.

The non-precipitation sample determination module A5 is configured to randomly sample the pixels in the non-precipitation area to obtain non-precipitation samples.

The inversion training sample determination module A6 is used to determine the precipitation samples of all precipitation connected domains and the non-precipitation samples as precipitation inversion training samples.

As an optional implementation manner, the acquisition module includes:

The obtaining unit is configured to obtain a precipitation situation map, where the pixel value of the precipitation situation map is the amount of precipitation, and the pixel position of the precipitation situation map is the geographic location corresponding to the pixel.

The binary map determining unit is configured to perform threshold segmentation on the precipitation situation map to obtain a precipitation binary map.

As an optional implementation manner, the precipitation event-guided remote sensing precipitation inversion training sample adaptive selection system further includes: a processing module configured to perform image mask processing on the precipitation binary image to obtain processed precipitation area.

As an optional implementation manner, the non-precipitation sample determination module includes:

A serial number determination unit, configured to label each pixel in the non-precipitation area according to the position of each pixel in the non-precipitation area to obtain the pixel serial number of each pixel in the non-precipitation area.

A non-precipitation sample determination unit, configured to randomly sample the pixels in the non-precipitation area according to the preset number of non-precipitation samples and the pixel serial number of each pixel to obtain non-precipitation samples.

FIG. 5 is a flowchart of a more specific method for adaptive selection of training samples for remote sensing precipitation inversion guided by precipitation events provided in this embodiment. For an actual precipitation map, a binary map is first obtained through threshold segmentation. The pixel value judges whether there is precipitation in this pixel, and obtains the precipitation area and the non-precipitation area. The non-precipitation area is sampled by simple average, and the image of the precipitation area is corroded and expanded, and then the independent but internally connected connectivity is obtained by dividing the precipitation events. Then, the precipitation samples are obtained by quantile selection, and the precipitation samples and non-precipitation samples are determined as the sample location set.

The present invention has the following advantages:

1. The present invention first divides the overall precipitation into various precipitation areas according to the concept of connected domain in graphics, and then selects samples in each precipitation area according to the assigned quota and quantile, which can effectively balance the disadvantages between different methods, because The sample selection is based on precipitation events and is allocated to precipitation events according to the number of places. There is not a lot of repetition and redundancy, so it can greatly make up for the lack of representativeness of the random selection method.

2. Since the sample set selected by this method is allocated to each precipitation event according to the precipitation quota, it will not make the sample selection too concentrated, so that it also has sufficient variability, and can effectively represent the precipitation events with different conditions and different characteristics. This improves the universality of the trained model and avoids overfitting, which is better than the similarity-based discriminant method.

3. There are not too many key parameters to input, only the precipitation quota needs to be input in advance, and this parameter only basically affects the volume of the data, and has no direct impact on the distribution and selection strategy of the data, so the relative numerical distribution control method, for artificial The prior parameter setting does not depend heavily, and it is a robust sample selection method.

4. It can search for precipitation events in a targeted manner, and divide and select the number of samples according to the precipitation events, which can effectively improve the representativeness of the sample set; and the prior parameters have little influence on the results of sample selection and have sufficient stability; The selection of the sample set has sufficient variability, which can effectively prevent overfitting and increase the universality of the sample set.

5. The present invention can apply image erosion technology to effectively reduce the influence of noise points on final sample selection.

6. The present invention can select samples according to the quantile in each precipitation event, so that the final selected precipitation sample set has universality, is not sensitive to the final sample generation result, and can get rid of the dependence on a priori parameters.

The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same and similar parts between the various embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant part can be referred to the description of the method.

In this paper, specific examples are used to illustrate the principles and implementations of the present invention. The descriptions of the above embodiments are only used to help understand the methods and core ideas of the present invention; meanwhile, for those skilled in the art, according to the present invention There will be changes in the specific implementation and application scope. In conclusion, the contents of this specification should not be construed as limiting the present invention.

Claims (8)

1. a remote sensing precipitation inversion training sample adaptive selection method guided by a precipitation event, is characterized in that, comprising: obtaining a precipitation binary map, where the precipitation binary map includes a precipitation area and a non-precipitation area; The precipitation area is divided into connected domains by using a classification algorithm to obtain multiple precipitation connected domains; For any precipitation connected domain, obtain the sampling number of the precipitation connected domain according to the number of pixels of the precipitation connected domain, the pixel number of the precipitation area and the preset number of precipitation samples; sampling the pixels of the precipitation connected domain according to the sampling quantity of the precipitation connected domain and the number of pixels of the precipitation connected domain to obtain the precipitation sample of the precipitation connected domain; randomly sampling the pixels of the non-precipitation area to obtain non-precipitation samples; Determine the precipitation samples of all precipitation connected domains and the non-precipitation samples as precipitation retrieval training samples; The sampling number of the precipitation connected region is obtained according to the number of pixels of the precipitation connected region, the number of pixels of the precipitation region and the preset number of precipitation samples, specifically: Calculate the sampling quantity of the _{ith precipitation connected domain according to the formula Mi = (N i /} N _total )*M _total , where Mi is the sampling quantity of the _ith precipitation connected domain, and M _total is the preset precipitation sample quantity; Ni is the number of pixels in the _i -th precipitation connected domain, and N _total is the number of pixels in the precipitation area; The sampling of the pixels of the precipitation connected domain according to the sampling quantity of the precipitation connected domain and the number of pixels of the precipitation connected domain is performed to obtain the precipitation sample of the precipitation connected domain, specifically: According to the formula S _i ={P _k |k=[(t-0.5)/M _i *N _i ], t∈N ⁺ , 1≤t≤M _i }, the precipitation samples of the i-th precipitation connected domain are obtained, where S _i is the precipitation sample of the i-th precipitation connected domain, P _k is the k-th precipitation pixel in the precipitation sample of the _i -th precipitation connected domain, t is the sample number of the precipitation Number of samples, Ni is the number of pixels in the _i -th precipitation connected domain. 2. The method for self-adapting selection of training samples for remote sensing precipitation inversion guided by a precipitation event according to claim 1, wherein the obtaining of a binary map of precipitation specifically comprises: obtaining a precipitation situation map, where the pixel value of the precipitation situation map is the amount of precipitation, and the pixel position of the precipitation situation map is the geographic location corresponding to the pixel; Threshold segmentation is performed on the precipitation situation map to obtain a precipitation binary map. 3. The method for adaptively selecting training samples for remote sensing precipitation inversion guided by a precipitation event according to claim 1, characterized in that, in the described precipitation region using a classification algorithm, a connected domain is divided to obtain a plurality of precipitation connectivity The domain also includes: Perform image mask processing on the precipitation binary image to obtain a processed precipitation area. 4 . The method for adaptively selecting training samples for remote sensing precipitation inversion guided by a precipitation event according to claim 1 , wherein the random sampling of pixels in the non-precipitation area to obtain non-precipitation samples, specifically comprising: 5 . : Labeling each pixel in the non-precipitation area according to the position of each pixel in the non-precipitation area to obtain the pixel serial number of each pixel in the non-precipitation area; The non-precipitation samples are obtained by randomly sampling the pixels in the non-precipitation area according to the preset number of non-precipitation samples and the pixel serial number of each pixel. 5. A remote sensing precipitation inversion training sample adaptive selection system guided by a precipitation event is characterized in that, comprising: an obtaining module, configured to obtain a precipitation binary image, wherein the precipitation binary image includes a precipitation area and a non-precipitation area; A precipitation connected domain determining module, used for dividing the precipitation region into a connected domain by using a classification algorithm to obtain a plurality of precipitation connected domains; A quantity determination module, configured to obtain the sampling quantity of the precipitation connected domain according to the number of pixels of the precipitation connected domain, the pixel quantity of the precipitation area and the preset precipitation sample quantity for any precipitation connected domain; a precipitation sample determination module, configured to sample the pixels of the precipitation connected domain according to the sampling quantity of the precipitation connected domain and the number of pixels of the precipitation connected domain to obtain the precipitation sample of the precipitation connected domain; a non-precipitation sample determination module, configured to randomly sample the pixels in the non-precipitation area to obtain non-precipitation samples; The inversion training sample determination module is used to determine the precipitation samples of all precipitation connected domains and the non-precipitation samples as the precipitation inversion training samples; The sampling number of the precipitation connected region is obtained according to the number of pixels in the precipitation connected region, the number of pixels in the precipitation region, and the preset number of precipitation samples, specifically: Calculate the sampling quantity of the _{ith precipitation connected domain according to the formula Mi = (N i /} N _total )*M _total , where Mi is the sampling quantity of the _ith precipitation connected domain, and M _total is the preset precipitation sample quantity; Ni is the number of pixels in the _i -th precipitation connected domain, and N _total is the number of pixels in the precipitation area; The sampling of the pixels of the precipitation connected domain according to the sampling quantity of the precipitation connected domain and the number of pixels of the precipitation connected domain is performed to obtain the precipitation sample of the precipitation connected domain, specifically: According to the formula S _i ={P _k |k=[(t-0.5)/M _i *N _i ], t∈N ⁺ , 1≤t≤M _i }, the precipitation samples of the i-th precipitation connected domain are obtained, where S _i is the precipitation sample of the i-th precipitation connected domain, P _k is the k-th precipitation pixel in the precipitation sample of the _i -th precipitation connected domain, t is the sample number of the precipitation Number of samples, Ni is the number of pixels in the _i -th precipitation connected domain. 6. The remote sensing precipitation inversion training sample adaptive selection system guided by a precipitation event according to claim 5, wherein the acquisition module comprises: an acquiring unit, configured to acquire a precipitation situation map, where the pixel value of the precipitation situation map is the amount of precipitation, and the pixel position of the precipitation situation map is the geographic location corresponding to the pixel; The binary image determination unit is configured to perform threshold segmentation on the precipitation situation map to obtain a precipitation binary image. 7. The precipitation event-guided remote sensing precipitation inversion training sample adaptive selection system according to claim 5, further comprising: a processing module for performing image mask processing on the precipitation binary image Get the processed precipitation area. 8. The remote sensing precipitation inversion training sample adaptive selection system guided by a precipitation event according to claim 5, wherein the non-precipitation sample determination module comprises: a serial number determination unit, configured to label each pixel in the non-precipitation area according to the position of each pixel in the non-precipitation area to obtain the pixel serial number of each pixel in the non-precipitation area; A non-precipitation sample determination unit, configured to randomly sample the pixels in the non-precipitation area according to the preset number of non-precipitation samples and the pixel serial number of each pixel to obtain non-precipitation samples.

Images