WO2023079669A1

WO2023079669A1 - Learning method for automatically detecting regions of interest from radar image, and automatic detection device

Info

Publication number: WO2023079669A1
Application number: PCT/JP2021/040695
Authority: WO
Inventors: 将敬鈴木; 俊平亀山
Original assignee: 三菱電機株式会社
Priority date: 2021-11-05
Filing date: 2021-11-05
Publication date: 2023-05-11
Also published as: JPWO2023079669A1; JP7391278B2

Abstract

The learning method according to the technique of the present disclosure comprises: calculating a feature for detecting a region of interest from an optical satellite image; detecting the location of the region of interest on a map using the feature; generating a teacher label for the location; creating a training data set for a radar satellite image, which is paired with the optical satellite image; and performing machine learning using the training data set.

Description

LEARNING METHOD AND AUTOMATIC DETECTION DEVICE FOR AUTOMATIC DETECTION OF REGION OF INTEREST FROM RADAR IMAGE

The technology disclosed herein relates to a learning method and an automatic detection device for automatically detecting a region of interest from a radar image.

Conventionally, attempts have been made to interpret features using radar images from satellites or radars mounted on aircraft.
Radar that uses microwaves can use its characteristics to photograph a wide area in a wide area regardless of the weather, day or night. However, radar images are difficult to decipher features due to complex scattering mechanisms and noise effects.
It is also conceivable to use an optical image as a means for interpreting features. Since the optical image uses wavelengths in the visible range, the same information as that seen by the human eye can be obtained, and features are easy to read. However, optical images can only be captured in good weather.

For example, in Patent Document 1, radar image data from a synthetic aperture radar undergoes conversion processing to bring the reflection characteristics of the confusion intensity closer to the reflection characteristics of black-and-white panchromatic, which is familiar to humans, and is superimposed on an optical image. Therefore, a technique for obtaining a composite image that is closer to an image captured by an optical satellite has been disclosed.

JP 2009-47516 A

If it is possible to obtain a composite image that is similar in nature to an optical image, it may be possible to examine the composite image with human eyes and discover, for example, a disaster.
The disclosed technique goes further and aims to provide a learning method for automatic detection of regions of interest from radar images. Here, the region of interest may be, for example, a location of a disaster such as an oil spill on the sea surface, a landslide caused by heavy rain or an earthquake, a forest fire, a fallen tree, or a flood.

A learning method according to the technology disclosed herein calculates a feature value for detecting a region of interest from an optical satellite image, detects the position of the region of interest on a map using the feature value, and teaches the position. A label is generated, a learning data set is created for the radar satellite image paired with the optical satellite image, and machine learning is performed using the learning data set.

Since the learning method according to the disclosed technique includes the above procedure, it can solve the problem of automatically detecting a region of interest from a radar image.

FIG. 1 is a flowchart showing procedures of a learning method for automatically detecting a region of interest from a radar image according to Embodiment 1. FIG. FIG. 2 is a flowchart showing procedures of a learning method for automatically detecting a region of interest from a radar image according to the second embodiment. FIG. 3 is a flowchart showing procedures of a learning method for automatically detecting a region of interest from a radar image according to the third embodiment. FIG. 4 is a flowchart showing procedures of a learning method for automatically detecting a region of interest from a radar image according to the fourth embodiment. FIG. 5 is a flow chart showing procedures of a learning method for automatically detecting a region of interest from a radar image according to the fifth embodiment. FIG. 6 is a flowchart showing procedures of a learning method for automatically detecting a region of interest from a radar image according to the sixth embodiment.

Embodiment 1.
FIG. 1 is a flowchart showing procedures of a learning method for automatically detecting a region of interest from a radar image according to Embodiment 1. FIG. As shown in FIG. 1, the learning method according to Embodiment 1 is roughly divided into three procedures. The three procedures are a procedure for optical satellite images (ST1, ST2, and ST3), a procedure for radar satellite images (ST4), and a procedure for learning (ST5).

The procedure related to the optical satellite image shown in FIG. 1 includes a procedure of calculating a feature amount (ST1), a procedure of detecting a target position (ST2), and a procedure of creating a teacher label (ST3).
The procedure related to radar satellite images shown in FIG. 1 includes a procedure (ST4) of creating a learning data set.
The procedure for learning shown in FIG. 1 includes a procedure for performing machine learning (ST5).

One aspect of the goals aimed at by the technique of the present disclosure is semantic segmentation (hereinafter referred to as “semantic segmentation”) for radar images. Semantic segmentation refers to supervised learning problems and techniques that associate labels or categories, such as what is seen, on a pixel-by-pixel basis in an image. In semantic segmentation, a data set consisting of an input image and a mask image is used, and learning is performed so that the mask image is output when the input image is input. A dataset consisting of the input image and the mask image is sometimes referred to as a dataset consisting of the original image and the labeled image.
If the scene assumed by the technology of the present disclosure is applied, the radar image is the input image, and the mask image is the location of the disaster indicated by the color of each type of disaster. That is, when a radar image is input, outputting a mask image in which disaster locations such as oil spills, landslides, and forest fires are masked with a color corresponding to the type of disaster, which is what the present disclosure is about. This is one of the goals of technology.

Semantic segmentation, like other supervised learning, requires huge amounts of training and validation datasets. A learning data set in the case where a radar image is used as an input image needs to be set by creating a mask image in which the disaster location is masked for the prepared radar image. However, as described above, the radar image has poor visibility and it is difficult to decipher the features. Therefore, it is not easy to create a mask image directly from a radar image.
Therefore, the technology disclosed herein pairs a radar satellite image with an optical satellite image in which at least a portion of the same area is captured, and creates a learning data set and a verification data set from the information of the paired images. think. That is, the technology disclosed herein uses a pair of a radar satellite image and an optical satellite image obtained by capturing the same area in the procedure of creating a learning data set. The procedures for optical satellite images (ST1, ST2, and ST3) and the procedure for radar satellite images (ST4) shown in FIG. 1 represent that the disclosed technique uses paired images.
Generally, for aerial photographs and satellite images, the shutter is pressed at the timing when half of the photographing area overlaps the next image. And the whole data is made up of multiple images taken. Therefore, the optical satellite image and the radar satellite image do not have to be a strict pair, but it is sufficient that they cover the same area as a whole.
Note that the learning data set is hereinafter simply referred to as a “learning data set”. Also, the verification data set is hereinafter simply referred to as a "verification data set".

A first procedure according to the technology disclosed herein is a procedure (ST1) for calculating a feature amount for an optical satellite image.
In the technology disclosed herein, semantic segmentation may be performed in advance on the optical image. Semantic segmentation for optical images may be realized, for example, by FCN (Fully Convolutional Networks). FCN is a kind of CNN, but considering that 1×1 convolution covering the entire region is performed, it has the property that the same result can be obtained even if the fully connected layer is replaced with the convolution layer. FCN has the advantage that it does not require resizing of the input image. The FCN includes a process of calculating a feature amount for detecting a region of interest from an input optical satellite image, that is, a feature amount calculation procedure (ST1).
Semantic segmentation for optical images may be performed by SegNet, U-Net, PSPNet, DeepLab, etc., in addition to FCN.

The technology disclosed herein may also use NDVI (Normalized Difference Vegetation Index) for this feature amount. NDVI is an index that indicates the distribution of vegetation and the degree of activity. NDVI is called Normalized Difference Vegetation Index in Japanese. NDVI is given by the following equation.

Here, NIR represents the reflectance of the near-infrared region (Band 5 of Landsat 8), and RED represents the reflectance of red light (Band 4 of Landsat 8). The NDVI is indicated by a normalized numerical value between -1 and 1, and the denser the vegetation, the larger the NDVI value. A pixel having an NDVI greater than or equal to a certain threshold may be determined to be a pixel in which vegetation is observed.
In addition, the technique of the present disclosure may compare a reference image of a certain period with the current image and images of two periods. Changes in vegetation can also be detected by comparing images from two periods and evaluating the amount of change in NDVI.
Details of the method using NDVI and other inter-band operations will become clear from the description of the third embodiment and subsequent embodiments.

A second procedure according to the technology disclosed herein is a procedure (ST2) for detecting a target position on an optical satellite image. The term "subject" in the second procedure is synonymous with the aforementioned region of interest. Also, the term "target position" in the second procedure is synonymous with the position of the region of interest.
Each pixel in the satellite image can be associated with a location on the map, eg, latitude and longitude. The technique of superimposing a radar image and an optical image exemplified in Patent Literature 1 can also be realized because each image can be represented by a common coordinate system, which is a position on a map. A map may be considered as a representation of the real space, which is part or all of the earth, as information on a coordinate system.
Specifically, the procedure for detecting the target position (ST2) is to specify the pixels of the region of interest in the optical satellite image and obtain the position on the map associated with the pixels.

A third procedure according to the disclosed technology is a procedure (ST3) for creating teacher labels for optical satellite images. As described above, when performing semantic segmentation on an optical image, the procedure for creating teacher labels (ST3) is nothing other than creating the above-described mask image. As mentioned above, mask images are also referred to as labeled images. The creation of the mask image may be manually performed by a human using the optical satellite image, or may be performed using semantic segmentation that has been pre-learned on the optical image.

A fourth procedure according to the technology disclosed herein is a procedure (ST4) for creating a learning data set for radar satellite images.
The procedure for creating a learning data set for radar satellite images can be further broken down into detailed procedures. Detailed procedure 1 is to prepare a pair of optical satellite images for radar satellite images. Detailed procedure 2 is to refer to the mask image for the paired optical satellite image. Detailed procedure 3 is to convert the mask information into information on map coordinates from the positions on the map associated with the pixels in the mask image. Here, the mask information is nothing but label information given in units of pixels. Finally, the detailed procedure 4 is to create a mask image for the radar satellite image from the mask information on the map coordinates. A radar satellite image and its mask image form a training data set for the radar satellite image.

Data consisting of radar satellite images (hereinafter referred to as "radar satellite image data") may be processed into data suitable for learning. For example, a radar satellite image may be trimmed to a predetermined size, such as N vertical pixels and M horizontal pixels. Also, for example, a radar satellite image may be processed to have as many layers or attributes as there are polarizations available.

The fifth procedure according to the technology disclosed herein is the procedure for performing machine learning (ST5). A learning data set used for machine learning is a learning data set for radar satellite images created in the procedure for creating a learning data set (ST4). The machine learning performed here can be the same as the semantic segmentation performed on optical satellite imagery. Also, the machine learning model performed here may use a neural network having the same structure as the neural network used in the semantic segmentation performed on the optical satellite image.

The disclosed technology obtains a trained model that performs semantic segmentation on radar satellite images by performing the above first to fifth procedures.

As described above, since the learning method according to Embodiment 1 includes the above procedure, it is possible to automatically detect a region of interest from a radar image with poor visibility and difficult to interpret features.

Embodiment 2.
The method according to Embodiment 1 shows up to the stage of obtaining a learning model. A method according to Embodiment 2 includes the method according to Embodiment 1, and further includes a procedure for verifying the learning model.
In the second embodiment, explanations overlapping those of the first embodiment are omitted as appropriate.

FIG. 2 is a flowchart showing the steps of a learning method for automatically detecting a region of interest from a radar image, according to the second embodiment. As shown in FIG. 2, the method according to Embodiment 2 includes, in addition to the procedure shown in Embodiment 1, a procedure for creating a verification data set (ST6) and a procedure for confirming verification results (ST7). include.

A sixth procedure according to the technology disclosed herein is a procedure (ST6) for creating a verification data set for radar satellite images. A different radar satellite image is used for the verification data set than the radar satellite image for the training data set. Creation of the verification data set may be performed in the same manner as the creation of the learning data set shown in the first embodiment.

A seventh procedure according to the technology disclosed herein is a procedure (ST7) for confirming the verification result of the radar satellite image. Here, semantic segmentation is performed on radar satellite images prepared for verification using a model with advanced learning, and whether the results are correct or not is compared with the mask image of the verification data set.
If the result of the verification is correct within an acceptable range, the learning is terminated. If the result of verification is not correct within an acceptable range, add training data and continue learning.

As described above, since the learning method according to Embodiment 2 includes the above procedure, the learning model can be verified, and the region of interest can be automatically detected from a radar image with poor visibility and difficult to interpret features.

Embodiment 3.
It can be said that the methods according to Embodiments 1 and 2 are general-purpose methods without limiting the type of region of interest.
Embodiment 3 clarifies a learning method for automatically detecting a region of interest from a radar image, which is specialized for fallen trees, by limiting the region of interest to fallen trees.
In Embodiment 3, explanations overlapping those of the above-described embodiments are omitted as appropriate.

When a region in which a fallen tree occurs is set as a region of interest, optical satellite images and radar satellite images before and after the occurrence of a fallen tree are used as a learning data set. Before and after the occurrence of a fallen tree may be, for example, before and after a typhoon.
Optical satellite images are used that include multispectral data for at least red light and near-infrared light.
Radar satellite images are used in a wavelength band such as the L band, which allows the transmission of plant branches and leaves.

FIG. 3 is a flow chart showing the procedure of a learning method for automatically detecting a region of interest from a radar image, according to the third embodiment. As shown in FIG. 3, the method according to the third embodiment is divided into a procedure for obtaining a learning model and a procedure for verifying the learning model. The procedure for obtaining a learning model includes a procedure for reading multispectral data (ST31), a procedure for calculating NDVI (ST32), a procedure for calculating the amount of change in NDVI (ST33), and comparing the amount of change in NDVI with a threshold. (ST34), generating a teacher label (ST35), creating a learning data set (ST36), and performing machine learning (ST37). The procedure for verifying the learning model includes a procedure for creating a verification data set (ST38) and a procedure for confirming verification results (ST39).

A first procedure according to Embodiment 3 is a procedure (ST31) of reading multispectral data for an optical satellite image.
Optical satellite images containing multispectral data may be prepared by using a multispectral or hyperspectral camera on board the optical satellite. An image obtained by a multispectral camera is called a multispectral image. Reading multispectral data may be considered synonymous with obtaining a multispectral image.

Because multispectral images contain information in different wavelength bands, they can better capture the features of specific surface objects. Calculations using information in different wavelength bands are referred to as interband calculations. The NDVI described in Embodiment 1 is one of representative inter-band operations.

A second procedure according to Embodiment 3 is a procedure (ST32) of calculating NDVI for an optical satellite image or a multispectral image.
Here, NDVI is calculated from each of the multispectral images before and after the fallen tree. The method according to Embodiment 3 is based on the conjecture principle that ``areas where fallen trees have occurred will have lower NDVI due to reduced trunk exposure and chlorophyll uptake''.
In the method according to the third embodiment, it is conceivable to use the reflectance of short-wavelength infrared light instead of the reflectance of red light in the inter-band calculation. The details of NDWI (Normalized Difference Water Index), which uses the reflectance of single-wavelength infrared light for inter-band calculation, will become clear from the description of the fourth embodiment.

A third procedure according to Embodiment 3 is a procedure (ST33) for calculating the amount of change in NDVI for an optical satellite image or a multispectral image.
Here, for example, the difference in NDVI is calculated for image data before and after a typhoon that is estimated to have caused a fallen tree.

A fourth procedure according to Embodiment 3 is a procedure (ST34) of comparing the amount of change in NDVI with a threshold for the optical satellite image or the multispectral image.
Here, the NDVI obtained in the third procedure is compared with a threshold, and pixels exceeding the threshold are determined to be pixels in the region of interest where fallen trees have occurred. An empirically determined value may be used as the NDVI threshold for fallen trees. In the flowchart block of the procedure (ST34) for comparing the amount of change in NDVI with the threshold value shown in FIG. 3, ΔNDVI represents the amount of change in NDVI, and _σVΔ represents the threshold value.

A fifth procedure according to Embodiment 3 is a procedure (ST35) for generating teacher labels for optical satellite images or multispectral images. This procedure is the same as the third procedure according to the first embodiment, that is, the procedure for creating teacher labels (ST3).

The sixth procedure according to Embodiment 3 is a procedure (ST36) for creating a learning data set for radar satellite images. This procedure is also the same as the procedure (ST4) for creating the learning data set according to the first embodiment.

The seventh procedure according to Embodiment 3 is the procedure for performing machine learning (ST37). This procedure is also the same as the procedure (ST5) for performing machine learning according to the first embodiment.

The eighth procedure according to Embodiment 3 is a procedure (ST38) for creating a verification data set for radar satellite images. This procedure is the same as the procedure (ST6) for creating the verification data set according to the second embodiment.

The ninth procedure according to Embodiment 3 is a procedure (ST39) for confirming the verification result of the radar satellite image. This procedure is also the same as the procedure (ST7) for confirming the verification result according to the second embodiment.

As described above, since the learning method according to Embodiment 3 includes the above procedure, it is possible to verify the learning model and automatically detect areas where fallen trees have occurred from radar images with poor visibility and difficult to interpret features.

Embodiment 4.
Embodiment 3 is a learning method for automatically detecting a region of interest from a radar image, which is specialized for fallen trees, by limiting the region of interest to fallen trees.
Embodiment 4 clarifies a learning method for automatically detecting a region of interest from a radar image, which limits the region of interest to flooding and is specialized for flooding.
In the fourth embodiment, descriptions overlapping those of the previous embodiments are omitted as appropriate.

Flooding here specifically means that a land area is invaded by water when river flooding occurs due to heavy rain, typhoon, or the like.
When a flooded area is set as a region of interest, optical satellite images and radar satellite images before and after flooding are used as learning data sets. Before and after the occurrence of flooding may be, for example, before and after heavy rain, or before and after a typhoon.
Optical satellite images are used that include multispectral data of at least near-infrared light and short-wave infrared light.
It is generally known that the backscattering coefficient of a radar satellite image changes before and after flooding occurs. This property confirms that it is possible to identify a flooded area of interest from a radar satellite image.

FIG. 4 is a flowchart showing the procedure of a learning method for automatically detecting a region of interest from a radar image according to the fourth embodiment. As shown in FIG. 4, the method according to Embodiment 4 is divided into a procedure for obtaining a learning model and a procedure for verifying the learning model. The procedure for obtaining a learning model includes a procedure for reading multispectral data (ST41), a procedure for calculating NDWI (ST42), a procedure for comparing NDWI with a threshold (ST43), and a procedure for generating teacher labels (ST44). , a procedure for creating a learning data set (ST45), and a procedure for performing machine learning (ST46). A procedure for verifying the learning model includes a procedure for creating a verification data set (ST47) and a procedure for confirming the verification result (ST48).

The first procedure according to Embodiment 4 is a procedure (ST41) for reading multispectral data for optical satellite images. This procedure is the same as the procedure for reading multispectral data (ST31) according to the third embodiment.

A second procedure according to the fourth embodiment is a procedure (ST42) of calculating NDWI for an optical satellite image or a multispectral image. This procedure is similar to the procedure for calculating NDVI (ST32) according to the third embodiment. NDWI is specifically given by the following equation.

Here, NIR represents reflectance in the near-infrared region (Band 5 of Landsat 8), and SWIR represents reflectance in the short-wave infrared region (Band 6 or Band 7 of Landsat 8).

A third procedure according to Embodiment 4 is a procedure (ST43) of comparing the NDWI with a threshold for the optical satellite image or the multispectral image.
Here, the NDWI obtained in the second procedure is compared with the threshold value for each pixel. A condition is considered such that the NDWI is smaller than the threshold in the multispectral image before the occurrence of flooding, but the NDWI is larger than the threshold in the multispectral image after the occurrence of flooding. A pixel that satisfies this condition is originally a land area and is determined to be a pixel for a flooded area. An empirically determined value may be used as the NDWI threshold for flooding. In the flowchart block of the procedure (ST43) for comparing the NDWI and the threshold shown in FIG. 4, NDWI _pre is the NDWI in the multispectral image before the occurrence of flooding, and NDWI _post is the NDWI in the multispectral image after the occurrence of flooding. show. In the same flowchart block, σ _Wpre represents the NDWI threshold in the multispectral image before the occurrence of water inundation, and σ _Wpost represents the NDWI threshold in the multispectral image after the occurrence of water inundation. Note that σ _Wpre and σ _Wpost may be the same value.

The fourth procedure according to Embodiment 4 is a procedure (ST44) for generating teacher labels for optical satellite images or multispectral images. This procedure is the same as the procedure (ST3) for creating teacher labels according to the first embodiment.

The fifth procedure according to Embodiment 4 is a procedure (ST45) for creating a learning data set for radar satellite images. This procedure is also the same as the procedure (ST4) for creating the learning data set according to the first embodiment.

The sixth procedure according to Embodiment 4 is the procedure for performing machine learning (ST46). This procedure is also the same as the procedure (ST5) for performing machine learning according to the first embodiment.

A seventh procedure according to the fourth embodiment is a procedure (ST47) for creating a verification data set for radar satellite images. This procedure is the same as the procedure (ST6) for creating the verification data set according to the second embodiment.

The eighth procedure according to Embodiment 4 is a procedure (ST48) for confirming the verification result of the radar satellite image. This procedure is also the same as the procedure (ST7) for confirming the verification result according to the second embodiment.

As described above, since the learning method according to Embodiment 4 includes the above procedure, it is possible to verify the learning model and automatically detect flooded areas from radar images with poor visibility and difficult to interpret features.

Embodiment 5.
In situations where flooded areas are automatically detected from radar images, if it is known in advance which locations on the map are water areas, false detections can be expected to be reduced.
Embodiment 5 is based on the learning method according to Embodiment 4, and clarifies a learning method that positively uses water area information.
In the fifth embodiment, descriptions overlapping those of the previous embodiments are omitted as appropriate.

FIG. 5 is a flowchart showing the procedure of a learning method for automatically detecting a region of interest from a radar image, according to Embodiment 5. As shown in FIG. 5, the method according to Embodiment 5 further includes a procedure for reading water area information (ST51) in addition to the procedure shown in Embodiment 4. FIG.

The geospatial information database shown in FIG. 5 is a database in which coordinates on the earth and land cover classifications are linked. Land cover classification refers to classification of land into types such as roads, rivers, forests, etc., or information summarizing the classification results. The method according to the technology disclosed herein positively uses the geospatial information database in situations where it can be used.
As mentioned above, if a geospatial information database can be used, false detections can be expected to be reduced. Furthermore, creating a geospatial information database is expected to reduce the number of satellite images required for learning.

As shown in FIG. 5, the second to ninth procedures according to the fifth embodiment are substantially the same as the first to eighth procedures according to the fourth embodiment. There are some differences between the fourth procedure according to the fifth embodiment and the third procedure according to the fourth embodiment.
Compared to the procedure (ST43) of comparing the NDWI and the threshold according to the fourth embodiment, the procedure (ST54) for comparing the NDWI and the threshold according to the fifth embodiment is different from the land area or The method of judging water areas is different. Embodiment 5 directly uses water area information obtained from a geospatial information database instead of comparing the NDWI of the multispectral image before the occurrence of flooding with a threshold value.

In the procedure (ST54) for comparing the NDWI and the threshold value according to Embodiment 5, the NDWI and the threshold value are compared only for the multispectral image after the occurrence of flooding. Furthermore, in the procedure (ST54) for comparing the NDWI with the threshold value according to Embodiment 5, the NDWI is compared with the threshold value only for the pixels corresponding to the areas other than the water area obtained from the geospatial information data. .

As described above, since the learning method according to Embodiment 5 includes the above procedure, the learning model can be verified, and the flooded area can be automatically detected from the radar image with poor visibility and difficult to interpret the features.
In addition, if a geospatial information database is available, the technology of the present disclosure can actively use it, reduce false detections, and reduce the number of satellite images required for learning.

Embodiment 6.
Embodiments 3 to 5 have clarified methods using NDVI and NDWI.
Embodiment 6 discloses yet another inter-band operation, NDSI (Noralized Difference Snow Index), and clarifies aspects of the method using multi-temporal observation data. Moreover, in Embodiment 6, the region of interest is not limited to the place where the disaster occurred.

FIG. 6 is a flowchart showing a procedure of a learning method for automatically detecting a region of interest from a radar image according to Embodiment 6. As shown in FIG. and a procedure for verifying the learning model. The procedure for obtaining the learning model includes a procedure for calculating the feature quantity (ST61), a procedure for calculating the amount of change in the feature quantity (ST62), a procedure for comparing the amount of change in the feature quantity with a threshold (ST63), and a teacher It includes a label creation procedure (ST64), a teacher label reading procedure (ST65), a learning data set creation procedure (ST66), and a machine learning procedure (ST67). The procedure for verifying the learning model includes a procedure for creating a verification data set (ST68) and a procedure for confirming verification results (ST69).

A first procedure according to Embodiment 6 is a procedure (ST61) for calculating a feature amount for an optical satellite image. Here, there may be a plurality of types of feature amounts, and NDSI shown in the following equation may be added in addition to NDVI and NDWI.

Here, GREEN represents the reflectance of the green region (Band 3 of Landsat 8), and SWIR represents the reflectance of the short wavelength infrared region (Band 6 or Band 7 of Landsat 8).
Optical satellite images to which the technology of the present disclosure is applied may be observation data from multiple periods. By calculating the NDSI with respect to the observation data of multiple periods, it is possible to specify the area and the period of snowfall.

The second procedure according to Embodiment 6 is the procedure (ST62) for calculating the amount of change in the feature amount for the optical satellite image. The optical satellite images for calculating the amount of change in the feature amount are the optical satellite image at the place and time of interest, the optical satellite image taken at the time just before the change in the land cover at that place, should be used. The time immediately prior to the change in land cover at the location may be, for example, one week prior, such as August 27, 2021 if the time of interest was September 3, 2021. .

The third procedure according to Embodiment 6 is a procedure (ST63) for comparing the amount of change in the feature quantity and the threshold for the optical satellite image. Here, a threshold value may be provided for each type of feature amount. As described above, an index such as NDVI is indicated by a normalized numerical value between -1 and 1. Therefore, the threshold for the amount of change in NDVI or the like may be set at 0.5, for example. Since the amount of change in NDVI or the like can be both positive and negative, the threshold values may also be determined on the positive side and the negative side, respectively.

The fourth procedure according to Embodiment 6 is the procedure (ST64) of creating teacher labels for optical satellite images or multispectral images. Here, the teacher label may be a label with content according to the nature of the feature quantity. For example, a pixel identified by comparing the amount of change in NDVI with a threshold value may be associated with a teacher label named "Vegetation". Similarly, a pixel specified by comparing the NDSI change amount and the threshold value may be associated with a teacher label named "Snow".

The fifth procedure according to Embodiment 6 is the procedure of reading teacher labels (ST65). Embodiment 6 uses multi-time observation data as described above. Therefore, the teacher label created in the procedure for creating the teacher label (ST64) is temporarily stored in the database shown in FIG. The procedure of reading teacher labels (ST65) is a procedure that is executed when machine learning is finally performed after sufficient multi-time observation data has been collected.

A sixth procedure according to Embodiment 6 is a procedure (ST66) for creating a learning data set for radar satellite images. This procedure is the same as the procedure (ST4) for creating the learning data set according to the first embodiment.

The seventh procedure according to Embodiment 6 is the procedure for performing machine learning (ST67). This procedure is also the same as the procedure (ST5) for performing machine learning according to the first embodiment.

The eighth procedure according to Embodiment 6 is the procedure for creating a verification data set (ST68). This procedure is the same as the procedure (ST6) for creating the verification data set according to the second embodiment.

The ninth procedure according to Embodiment 6 is the procedure for confirming the verification result (ST69). This procedure is also the same as the procedure (ST7) for confirming the verification result according to the second embodiment.

As described above, since the learning method according to the sixth embodiment includes the above procedure, the learning model can be verified, and automatic detection according to the type of land cover can be performed from a radar image with poor visibility and difficult to interpret features. can.

The disclosed technology can be applied to an automatic detection device for damaged areas using radar satellite images, and has industrial applicability.

Claims

Calculating features for detecting regions of interest from optical satellite images,
Detecting the position of the region of interest on a map using the feature amount,
generate teacher labels for the positions;
creating a training data set for the radar satellite image paired with the optical satellite image;
performing machine learning using the learning data set;
learning method.
Furthermore, using a verification data set different from the learning data set, verifying the learning model that has undergone the machine learning,
A learning method according to claim 1.
The feature quantity includes NDVI,
2. The learning method according to claim 1, wherein the area of interest is an area where fallen trees have occurred.
The feature quantity includes NDWI,
The learning method according to claim 1, wherein the region of interest is a flooded area.
5. The learning method according to claim 4, wherein the teacher label is generated using a geospatial information database.
The optical satellite image is multi-time observation data,
wherein the feature quantity includes NDSI;
A learning method according to claim 1.
Learned by the learning method according to any one of claims 1 to 6,
Automatic detection device.