KR102316598B1 - Fabricating system and method of Landcover map, and Program recording media - Google Patents

Abstract

The present invention relates to a method for generating a landcover map and a manufacturing system for performing the method. A large amount of real-time data is formed through machine learning using a published street view image. Time-series data for each pixel of this large-scale real-time data and multi-channel spectral reflectivity data provided from a satellite are combined through machine learning. So, it is possible to manufacture a landcover map on a wide-area scale in an economical and reliable way.

Description

A recording medium containing a method for producing a land cover map, a production system, and a program for performing the method {Fabricating system and method of Landcover map, and Program recording media}

The present invention relates to a method for generating a land cover map and a manufacturing system for performing the method.

The land cover map is a kind of thematic map, and it is a spatial information DB (Database) that is expressed in the form of a map after color indexing the areas with homogeneous characteristics by classifying the shapes of the topographical features on the earth's surface according to certain scientific standards. Because it best reflects the phenomena of the ground surface, calculation of the load of nonpoint pollution sources based on the permeability of the ground surface, urban planning based on biotope map preparation, simulation of submergence damage in the downstream area at the time of discharge from the dam sluice, climate and atmospheric prediction modeling, environmental impact assessment It can be widely used, and is currently being used as a scientific basis for the establishment of environmental policies by the central and local governments and as various research data in related academia.

The land cover map has three hierarchies according to spatial resolution: large classification (resolution 30M class), medium classification (resolution 5M class), and subclassification (resolution 1M class). It is defined as arid area, agricultural area, forest area, grassland, wetland, bare area, and water area), and the middle classification is a total of 22 areas (residential area, industrial area, commercial area, cultural/sports/recreation area, transportation area, public facility) Area, paddy field, field, facility cultivation area, orchard, other cultivated land, broadleaf forest, coniferous forest, mixed forest, natural grassland, artificial grassland, inland wetland, coastal wetland, natural bare area, other open land, inland water, marine water) 41 areas (detached residential facilities, shared residential facilities, industrial facilities, commercial/business facilities, mixed areas, cultural/sports/recreational facilities, airports, ports, railways, roads, other transportation/communication facilities, environmental basic facilities, education/ Administrative facilities, other public facilities, cultivated rice fields, cultivated rice fields, cultivated fields, cultivated fields, orchards, ranches/farms, other cultivated fields, broadleaf forests, coniferous forests, It is further subdivided into mixed forests, natural grasslands, golf courses, graveyards, other grasslands, inland wetlands, tidal flats, salt fields, beaches, riverbanks, rocks/rocks, mining areas, playgrounds, other open fields, rivers, lakes, and ocean water), and the central government Construction and renewal are being carried out on a national level by the Ministry of Environment, a ministry.

Since the first national large-class land cover map was constructed in 1998 using various satellite and aerial images, the construction of a nationwide mid-class land cover map was completed by 2010. Subdivision land cover maps are being produced.

However, it is difficult to produce a wide area unit (country unit) due to the nature of the production and update of this land cover map based on image analysis for a limited area unit. It requires a wide range of reference point data, causing a problem in that time and cost are greatly increased. In addition, since it takes a lot of cost to update information on a specific regional basis, the need for a more efficient production method and a method for generating a cover map that goes beyond regional limits are growing.

Korean Patent Registration No. 10-2009573 Korean Patent Registration No. 10-1724801

The present invention has been devised to solve the above needs, and an object of the present invention is to form a large-capacity live-action data through machine learning using a street view image, and multi-channel spectroscopy provided by this large-capacity live-action data and satellites The purpose of this study is to provide a method and system that can produce land cover maps on a wide-area scale in an economical and reliable way by combining time series data for each pixel of reflectivity data through machine learning.

As a means for solving the above-mentioned problem, in an embodiment of the present invention, (A) obtaining a street view image corresponding to the longitude and latitude information of the target area from image providing servers including the longitude and latitude information; (B) obtaining a street view image corresponding to the longitude and latitude information of the target area, and forming a massive ground truth (first data) through machine learning; (C) obtaining the spectral reflectivity data of the artificial satellite for the target area, and forming second data by deriving the time series spectral reflectance and vegetation index data for the target area; and (D) using the result of combining the first data and the second data as an input value to form a cover map of the soil of the target area using a convolutional neural network (CNN). To provide a method for making a cover map.

In addition, in another embodiment of the present invention, as a system for implementing the above-described method for producing a land cover map, a street view image providing server that provides an image including longitude and latitude information; a ground reference data calculation module 100 for obtaining a street view image corresponding to the longitude and latitude information of the target area from the image providing servers, and forming massive ground truth (first data) through machine learning; A time series data calculation module 200 that acquires spectral reflectance data of artificial satellites for a target area, derives time series spectral reflectance data to form second data, and a convolutional neural network ( CNN) to provide a land cover map production system, including a land cover map formation module 300 that forms a cover map of the soil of the target area.

In addition, as another embodiment of the present invention, it is possible to provide a recording medium in which a program for performing a method of manufacturing a land cover map is recorded.

According to an embodiment of the present invention, a large-capacity actual data is formed through machine learning using a public street view image, and the time-series data for each pixel of the large-capacity actual data and multi-channel spectral reflectivity data provided from the satellite are machined. By combining through learning, the advantage of being able to produce land cover maps on a wide-area scale in an economical and reliable way is realized.

In addition, according to the present invention, it is possible to classify various types of land cover, such as crops, trees, buildings, and bedrock, and create a land cover map with high reliability through machine learning for this, so that land use or change of cover, land It can be applied to various application industries such as estimation of cover area and monitoring of crop production.

1 is a flowchart illustrating a method of manufacturing a land cover map according to an embodiment of the present invention.
FIG. 2 is a block diagram of a system for manufacturing a land cover map according to another embodiment of the present invention performing the process of FIG. 1 .
3 to 15 are diagrams illustrating an embodiment to which a method for generating a land cover map according to an embodiment of the present invention is applied.

Advantages and features of the present invention, and methods for achieving them, will become apparent with reference to the embodiments described below in detail in conjunction with the accompanying drawings. However, the present invention is not limited to the embodiments described herein and may be embodied in other forms. Rather, the embodiments introduced herein are provided so that the disclosed subject matter may be thorough and complete, and that the spirit of the present invention may be sufficiently conveyed to those skilled in the art.

The terms used in the present application are only used to describe specific embodiments, and are not intended to limit the present invention. The singular expression includes the plural expression unless the context clearly dictates otherwise. In the present application, terms such as “comprise” or “have” are intended to designate that a feature, number, step, operation, component, part, or combination thereof described in the specification exists, but one or more other features It is to be understood that it does not preclude the possibility of the presence or addition of numbers, steps, operations, components, parts, or combinations thereof.

Unless defined otherwise, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. Terms such as those defined in a commonly used dictionary should be interpreted as having a meaning consistent with the meaning in the context of the related art, and should not be interpreted in an ideal or excessively formal meaning unless explicitly defined in the present application. does not

1 is a flowchart illustrating a method of manufacturing a land cover map (hereinafter referred to as 'the present invention') according to an embodiment of the present invention. In addition, FIG. 2 shows a block diagram of a manufacturing system of a land cover map according to another embodiment of the present invention for performing the process of FIG. 1 .

1 and 2, the present invention comprises the steps of: (A) acquiring a Street View image corresponding to longitude and latitude information of a target area from image providing servers including longitude and latitude information; (B) the Street View image classifying by land cover type to form massive ground truth (first data) for the target area through machine learning, (C) acquiring the spectral reflectivity data of the satellite for the target area, and time series spectrum Forming second data by deriving reflectance data and (D) using a CNN (Convolution Neural Network) as an input value for the result of combining the first data and the second data of the soil of the target area It may be carried out as a process including the step of forming a cover map. In this case, steps (A) and (B) may be implemented sequentially or may be implemented simultaneously, so there is no restriction on the order of the steps.

1. Ground Reference Data Calculation Process

In the step (A) of the present invention, a street view image corresponding to the longitude and latitude information of the target area is acquired from the image providing servers including the longitude and latitude information. Therefore, the data applied in this step specifies the target area based on the longitude and latitude information of the target area to form the land cover map, and extracts and acquires a street view image corresponding to the specified target area.

Here, the image providing server is a server of a commercialized portal site or website, for example, Google Street View, a live-action web map service provided by Google, or Korea's Naver or Daum ( daum), a server that provides an image view that is formed in parallel with longitude and latitude information for a specific area, such as Street View provided by portals such as Baidu in China and Street View provided by portals such as Baidu in China. Therefore, the Street View image applied to the present invention can use images from various web map services, which are open source and have no limitation in their usability. It can be used in the present invention as long as it contains information of

As such an example, in the present invention, an example in which image data provided by Google Street View (GSV) can be used will be described as an example. Google Street View provides a large amount of panoramic images including detailed geographic coordinates and time information on a global scale, and in an embodiment of the present invention, a large amount of ground truth data (massive ground truth) is performed through machine learning using image data. ) to be calculated. In the case of the machine learning method applied to this process, various machine learning methods such as random forest and CNN (Convolution Neural Network) can be applied, and in the embodiment of the present invention, CNN (Convolution Neural Network) Network) to make it possible to implement.

Next, in step (B), the street view image acquired in step (A) is classified by land cover type, and actual image data for the target area through machine learning (hereinafter referred to as 'first data'). is the process of forming

That is, in the present invention, extensive reference data (ground reference) is required for training and verification through machine learning, and these reference data generally include file surveys, census data, and visual interpretation of remote sensing products, but The method of calculating the reference data of . The present invention makes it possible to obtain, process, and utilize a street view image in forming such reference data, thereby eliminating time, cost, and limitation of scope. In the case of labeling of reference data, it is possible to implement a data set that can be applied by dividing the acquired data into data to be applied to training and validation.

To this end, in step (B) of the present invention, (b1) processing the acquired street view image, classifying it by land cover type, and implementing it as a patterned training dataset, (b2) using a convolutional neural network (CNN) By learning the information of the acquired training dataset, (b3) it is possible to form reference data through the process of forming massive ground truth (first data). At the same time, a validation process can be performed by applying the validation data set to the aforementioned reference data set.

2. Pixel-based time series data calculation process

Next, step (C) of the present invention is implemented as a process of obtaining the spectral reflectance data of the artificial satellite for the target area, and forming the second data by deriving the time series spectral reflectance and vegetation index data. In the case of the artificial satellite data acquired in this step, since it is based on satellite images, it has time series information and spatial series information at the same time. It is possible to acquire spatial time series data for a specific target area. (For example, for the entire Korean Peninsula, time series satellite data at 30m resolution are acquired at 2-week intervals, and spectral reflectivity data and vegetation index data are calculated.)

The massive ground truth (first data) for the target area is calculated as reference data, and the spectral reflectivity data of the artificial satellite for the same target area can be obtained and processed separately from this.

To this end, the step (C) is, (c1) acquiring the spectral reflectivity data of the artificial satellite for the target area, (c2) filtering the obtained spectral reflectivity data to remove inappropriate data by applying the Fmask algorithm, ( c3) may be configured including the step of deriving time series spectral reflectance data from the filtered data.

For example, for the spectral reflectance data of the satellite, surface reflectance is acquired from Landsat 7 or 8 Collection 1 Level-2, which can be used in the US Geological Survey database (https://earthexplorer.usgs.gov/). can be used by Of course, the source of the satellite data acquisition is not limited thereto.

In addition, in the case of process (c2), it is a process of filtering spectral reflectance data. As an example of the present invention, in the surface reflectance data acquired using QGIS3 software, data polluted by clouds, cloud shadows, and snow In order to remove it, it can be filtered by applying the Fmask algorithm.

In this way, from the filtered data, time series spectral reflectance and vegetation index data for the target area (hereinafter referred to as 'second data') are derived.

This second data is combined with the first data in second machine learning (CNN) to be described later so that it can be applied as an input variable. This is because the remote sensing surface reflectance data for the target area can implement a soil cover map in consideration of time and spectral characteristics. As an input to the CNN model, time series spectral reflectance data (blue, green, red, SWIR1, and SWIR2) can be applied. Furthermore, in the case of vegetation such as crops or trees, the classification of other trees or crop types with similar spectral information may be hindered. If necessary, the vegetation index may be combined with the second data to be an input value. In this case, the applied vegetation index is based on the filtered data, at least one or more of NDVI (Normal Vegetation Index), EVI (Enhanced Plant Index), ENDVI (Standardized Regular Vegetation Index), and LSWI (Land Water Index). may include

3. Land Coverage Map Formation Process

Next, in the present invention, the step (D) of forming a soil cover map of the target area is performed by using a convolutional neural network (CNN) as an input value of the result of combining the above-described first data and second data. can be

That is, the step (D) is, as an input of the CNN model, through the second machine learning unit of the system to be described later, the first data (a large amount of actual data) and the second data (time-series spectral reflectance data) and the vegetation index (VI) ) to perform learning by combining them to form a land cover map.

According to the above-described present invention, a large-capacity actual data is formed through machine learning using a public street view image, and multi-channel spectroscopy provided from this large-capacity ground truth (first data) and an artificial satellite. By combining the time series data for each pixel of the reflectivity data through machine learning, it is possible to produce a land cover map on a wide scale in an economical and reliable way.

4. Land cover map production system

Hereinafter, a configuration of a land cover map system in which a module for embedding an apparatus and a program for implementing the method for producing a land cover map of the present invention described above will be described.

As shown in FIG. 2, the system (hereinafter referred to as 'the present system') for implementing the method for producing the land cover map according to the present invention described above is a street view image that provides an image including longitude and latitude information. The providing server 10 and the image providing servers 10 obtain a street view image corresponding to the longitude and latitude information of the target area, and form a massive ground truth (first data) through machine learning. The ground reference data calculation module 100, the time series data calculation module 200 which acquires the spectral reflectance data of the artificial satellite for the target area, and forms the second data by deriving the time series spectral reflectance data, and the first data and the second data may be configured to include a land cover map formation module 300 that forms a cover map of the soil of the target area using a convolutional neural network (CNN) as an input value.

As described above, the street view image providing server 10 performs the process of step (A) of the present invention, and is a server of a commercialized portal site or website, and Google is a live-action web map service provided by Google. As described above, a server such as Google Street View may be used.

In addition, the ground reference data calculation module 100 acquires a street view image corresponding to the longitude and latitude information of the target area from the image providing servers 10, and performs massive ground truth (first truth) data through machine learning. data), and the process of step (B) in FIG. 1 described above can be performed.

To this end, the ground reference data calculation module 100 in the system includes a data acquisition unit 110 that acquires a street view image corresponding to the longitude and latitude information of the target area by referring to the longitude and latitude information of the target area, a machine The information of the training dataset acquired using the data processing unit 120 and convolutional neural network (CNN), which processes the street view image acquired for learning, classifies it by land cover type and implements it as a patterned training dataset. A first machine learning unit 130 that classifies by land cover type, learns, and forms first data, and a data verification correction unit 140 that verifies, corrects, and stores the accuracy of the first data. make it possible

In addition, the time series data calculation module 200 is to perform the process of step (C) described above, acquires the spectral reflectance data of the artificial satellite for the target area, derives the time series spectral reflectance data to form the second data let it be To this end, the time series data calculation module 200, the reflectivity data acquisition unit 210 that receives the spectral reflectivity data of the artificial satellite for the target area, from the acquired spectral reflectivity data, removes inappropriate data by applying the Fmask algorithm. It may be configured to include a reflectivity data calculation unit 220 for filtering and deriving time series spectral reflectance data, and a vegetation index calculation unit 230 for calculating the vegetation index of the target area based on the filtered data.

In addition, the land cover map formation module 300 in the present system performs the process of step (D) described above. For this purpose, as an input of a convolutional neural network (CNN) model, the first data (large land cover map using the results of the second machine learning unit 310 and the second machine learning unit 320 learning by combining the actual data) and the second data (time series spectral reflectance data) and the vegetation index (VI) It may be configured to include a land cover map forming unit 320 to form a.

That is, the present system may be configured as a system for implementing the method of manufacturing the land cover map of the steps (A) to (D) of the present invention described above in the embodiment of FIG. 1 . Accordingly, the present invention can provide a system for implementing a method of manufacturing a land cover map. Furthermore, the functional configuration and execution operation applied to the system executing the method of manufacturing the land cover map in the present invention may be represented by functional block configurations and various processing steps. These functional blocks may be implemented in any number of hardware and/or software configurations that perform specific functions. For example, the present invention provides integrated circuit configurations, such as memory, processing, logic, look-up table, etc., capable of executing various functions by means of the control of one or more microprocessors or other control devices. can be hired

Similar to how components of the present invention may be implemented as software programming or software elements, the present invention includes various algorithms implemented as data structures, processes, routines, or combinations of other programming constructs, including Python (Python). ), C, C++, Java, assembler, etc. may be implemented in a programming or scripting language. Functional aspects may be implemented in an algorithm running on one or more processors. Further, the present invention may employ prior art techniques for electronic configuration, signal processing, and/or data processing, and the like. Terms such as “module”, “part”, “mechanism”, “element”, “means”, and “configuration” may be used broadly and are not limited to mechanical and physical components. The term may include the meaning of a series of routines of software in association with a processor or the like.

According to the present invention, it is possible to classify various types of land cover, such as crops, trees, buildings, and bedrock, and generate a highly reliable land cover map through machine learning for this, so that land use or change of cover, land cover area It can be applied to various application industries such as estimation of crop production and monitoring of crop production. Hereinafter, in implementing the method for producing a land cover map according to the present invention described above in FIGS. 1 and 2, a method of forming a cover map implemented as a map by classifying the types of cover of crops will be described as an embodiment. decide to do

[Example: Fig. 2 system applied]

(1) Designation of target area

As shown in FIG. 3 , two areas are set as target areas as target areas in the present embodiment. This means that area I (119° 26" - 119° 17" W, 35° 26" - 35° 34" N) is located in Southern California, USA (FIG. 3(a)).

The rectangular area (total area 18,088 ha) posted in Fig. 3(a) is equal to 13.6 km Х 13.3 km. According to cropland data layer (CDL) data released annually by the U.S. Department of Agriculture (USDA), the region has a very heterogeneous landscape structure that includes a variety of crop types including alfalfa, almonds, corn, cotton, grapes, pistachios, and wheat.

Area II (89° 47" - 89° 24" W, 40° 08" - 40° 18" N) is located inside the state of Illinois (FIG. 3(b)), USA. The rectangular area (total area 62,748 ha) shown in Fig. 3(b) is equal to 33.2 km Х 18.9 km. This area is considered part of the US corn belt. The main crops are soybeans and corn.

(2) Acquisition of image data

Image data corresponding to the specified target area was acquired from Google Street View (hereinafter, 'GSV') as the Street View image providing server described above in the configuration of FIG. 2 . The GSV image was automatically downloaded from Google Maps using an HTTP URL request using the GSV API and a Python script, which was stored and processed by the ground reference data calculation module 100 described above in FIG. 2 .

(3) Calculation of reference data

The street view image corresponding to the longitude and latitude information of the target area acquired by the data acquisition unit 110 is implemented as a patterned training dataset by processing the street view image acquired for machine learning and classifying it by land cover type (data processing part: 120).

In this example, a CNN model is applied to classify all collected GSV images into 3 classes in Illinois (corn, soybean, etc.) and 7 classes in California (alfalfa, almond, corn, cotton, grape, pistachio, etc.) did. The main structure of such a CNN model is shown in FIG. 4 . The first machine learning unit 130 processes the street view image acquired for machine learning, classifies it by land cover type, and implements it as a patterned training dataset.

To train the CNN model, about 5,540 GSV images taken outside the target area were collected. In particular, for annual crops, images of various crop bioseasonal stages were included where possible.

Afterwards, images were randomly divided into three groups (60%, 20%, and 20%) for training (parameter fitting), validation (hyper-parameter tuning), and testing (performance evaluation). To prepare the training dataset for the CNN model, the GSV images were classified into 7 classes in California and 3 classes in Illinois (Fig. 5). Figure 5, (a) corn, (b) soybean, (c) almond, (d) cotton, (e) alfalfa, (f) grape, (g) pistachio, (h&i) other cases, CNN classification model A sample of GSV images used for training is shown.

Thereafter, verification and correction are performed through the data verification and correction unit 140 in order to remove the image incorrectly classified by the CNN model.

Referring to what is shown in FIG. 6 , the coordinates of the image represent the position of the GSV vehicle and are not identical to the crop compartment of interest. Therefore, empirical coefficients x and y were used to move the coordinates from the GSV vehicle to the imaged compartment of interest.

Furthermore, in order to avoid the buffer zone between the road and the farmland and to extract the location information of the farmland, the spatial resolution of the Landsat pixel of 30 m information was used. The coordinates of the target GSV image were shifted back and forth by x at the edge of the parcel, 0.5y away from the vehicle (FIG. 6). x and y are the width and height in pixels, for Landsat, both x and y are 30 m. When using other satellite images, use the corresponding pixel resolution (eg 10 X 10 m for Sentinel).

If the reference point for a particular crop type was insufficient, the coordinates were shifted by an additional distance of 30 m from the road in the same direction as the cropland. A fixed empirical coefficient (x, y) value of 30 m between adjacent reference points is related to the spatial resolution of the Landsat 7 and 8 surface reflectance products used for mapping. Furthermore, to simplify the procedure, only images captured from four absolute cardinal directions: North, South, East and West were considered.

(5) Calculation of pixel-based time series data and formation of land cover map

Then, pixel-based crop classification is performed for the state of Illinois and California for the target area. It acquired surface reflectance data from Landsat 7 or 8 Collection 1 Level-2, which is available in Satellite Acquisition of Spectral Reflectance Data: US Geological Survey database (https://earthexplorer.usgs.gov/).

Thereafter, the reflectance data was filtered by processing the acquired reflectivity data. This is a time series spectral reflectance data derived from surface reflectance data acquired using QGIS3 software by applying Fmask algorithm to remove data polluted by clouds, cloud shadows, and snow.

In addition, the vegetation index was calculated. NDVI (Normal Vegetation Index), EVI (Enhanced Vegetation Index); ENDVI (Normalized Normal Vegetation Index); It was implemented through the calculation of the vegetation index calculation unit according to the LSWI (Land Water Index) equation.

[formula]

Thereafter, the land cover map can be formed through the second machine learning unit of the land cover map formation module 300 .

This was done using the CNN model described above in FIG. 4, and crop classification was performed by changing the withdrawal rates (0.1, 0.2, 0.3, 0.5) and input values. The temporal and spectral properties of remote sensing surface reflectance products are commonly used to map crop types. However, other crop types with similar spectral information may interfere with classification. Therefore, we combined the time series spectral reflectance data (blue, green, red, NIR, SWIR1, and SWIR2) with the vegetation index (VI) (NDVI, EVI, ENDVI, LSWI) as inputs to the CNN model.

To train the CNN model, we divided it into 80% of training and 20% of evaluation. Mapping results were evaluated using CDL products. The mapping accuracy of CDL products is reported to be 85-95% for major crop types.

7 is an image illustrating a driving process of downloading and acquiring a Google Street View image from the image data acquisition unit of the present system. 8 is an image for implementing a training dataset for calculating ground reference data in the data processing unit of the present system. 9 is an image of learning through a training dataset in the first machine learning unit of the present system. 10 is an image showing a result of generation of ground reference data. 11 is an image illustrating a driving process for calculating the Lansat vegetation index. 12 is an image illustrating a process of forming a map by performing machine learning in the land cover map formation module.

Figure 13 shows the number of areas of each land cover type in (a) California (b) Illinois using USDA CDL and CNN-based crop type classification as a result derived through the above-described process. Shows the results of map making.

14 shows the results of comparing crop type maps of Illinois using (a) USDA CDL data layer products in 2013 and (b) 2013 CNN and GSV images.

15 shows the results of comparing crop type maps of California using (a) USDA CDL data layer products in 2012 and (b) 2012 CNN and GSV images.

Looking at the above results, the CNN model applied to the present invention selected 2,645 images suitable for classification of crops from 8,514 GSV images. These images were well categorized by crop types including alfalfa, almonds, corn, cotton, grapes, soybeans, and pistachios. Notably, the overall GSV image classification accuracy reached 93% in California and 97% in Illinois. Then, using fixed empirical coefficients, the image geographic coordinates were shifted to generate 8,173 crop control points, including 1,764 in Illinois and 6,409 in California. Evaluation of these new baselines using CDL products resulted in concordance rates of 94-97%. Machine learning (CNN) based cartography according to the present invention also captured a general pattern of crop type distribution, and the overall difference between CDL products and our cartographic results was 4% in California and 5% in Illinois. . Therefore, it is disproved that the land cover map system according to the present invention can act as a cost-effective alternative method.

As described above, the technical idea of the present invention has been specifically described in the preferred embodiment, but the preferred embodiment is for the purpose of explanation and not for limitation. As such, those of ordinary skill in the art will be able to understand that various embodiments are possible through combination of embodiments of the present invention within the scope of the technical spirit of the present invention.

10: Street View image providing server
100: ground reference data calculation module
200: time series data calculation module
300: Soil coverage map calculation module

Claims (10)

(A) acquiring a street view image corresponding to longitude and latitude information of a target area from image providing servers including longitude and latitude information;
(B) The street view image is classified by land cover type and implemented as a patterned training dataset, and the information of the training dataset acquired using a convolutional neural network is learned to convert large-scale due diligence data for the target area to the first data. forming with;
(C) forming second data by obtaining spectral reflectance data of artificial satellites for the target area and deriving time-series spectral reflectance data; and
(D) using the result of combining the first and second data and the vegetation index (VI) as an input value of the convolutional neural network model to form a soil cover map of the target area;
A method of making a land cover map.
delete The method according to claim 1,
The step (C) is,
(c1) Acquiring the spectral reflectance data of the artificial satellite for the target area,
(c2) From the acquired spectral reflectivity data, filter to remove nonconforming data by applying Fmask algorithm,
(c3) deriving time series spectral reflectance data from the filtered data; including,
A method of making a land cover map.
4. The method according to claim 3,
The step (C) is,
Based on the filtered data,
Deriving a vegetation index comprising at least one of NDVI (regular vegetation index), EVI (enhanced vegetation index), ENDVI (standardized regular vegetation index), LSWI (surface moisture index); further comprising,
A method of making a land cover map.
delete a street view image providing server 10 that provides an image including longitude and latitude information;
By acquiring the street view image corresponding to the longitude and latitude information of the target area from the image providing servers 10 and implementing it as a patterned training dataset, by learning the information of the acquired training dataset using a convolutional neural network, the machine a ground reference data calculation module 100 for forming first data that is a large-capacity actual data for a target area through learning;
a time series data calculation module 200 for obtaining spectral reflectance data of artificial satellites for a target area, and deriving time series spectral reflectance data to form second data; and
A land cover map forming module 300 that forms a soil cover map of the target area by using the result of combining the first and second data and the vegetation index (VI) as an input value of the convolutional neural network model; ,
Land cover map production system.
7. The method of claim 6,
The ground reference data calculation module 100,
a data acquisition unit 110 for obtaining a street view image corresponding to the longitude and latitude information of the target area with reference to the longitude and latitude information of the target area;
The data processing unit 120 that processes the street view image acquired for machine learning, classifies it by land cover type and implements it as a patterned training dataset
The first machine learning unit 130 that processes the street view image acquired for machine learning, classifies it by land cover type, and implements it as a patterned training dataset.
A data verification correction unit 140 that verifies, corrects, and stores the accuracy of the first data;
Land cover map production system.
8. The method of claim 7,
The time series data calculation module 200,
a reflectivity data acquisition unit 210 that receives the spectral reflectivity data of the artificial satellite for the target area;
The reflectance data calculation unit 220 that filters the acquired spectral reflectance data to remove non-conforming data by applying the Fmask algorithm, and derives time series spectral reflectance data
Based on the filtered data, the vegetation index calculation unit 230 for calculating the vegetation index of the target area; including;
Land cover map production system.
9. The method of claim 8,
The land cover map formation module 300,
As an input of a convolutional neural network (CNN) model, a second machine learning unit 310 for learning by combining the first data and the second data and the vegetation index (VI);
Containing;;
Land cover map production system.
A recording medium in which a program for performing the method of manufacturing a land cover map according to claim 1 is recorded.

Images