KR20220168169A - Apparatus and method for detecting abandoned farmland using harmonic analysis and machine learning - Google Patents

Abstract

An embodiment of the present invention relates to a method and apparatus for detecting abandoned farmland using harmonic analysis and machine learning. It relates to a method and apparatus for detecting abandoned farmland to classify frequency coefficients obtained by harmonic time-series analysis of vegetation indices of satellite imagery according to a machine learning classification method and classify abandoned farmland, paddy fields, and fields.

Description

Method and apparatus for detecting idle farmland using harmonic analysis and machine learning

Embodiments of the present invention relate to a method and apparatus for detecting idle farmland using harmonic analysis and machine learning, and more particularly, to a method for distinguishing farmland where crops are grown from farmland that is idle using harmonic analysis and machine learning will be.

Agricultural land is defined as land designated for agricultural production, such as rice fields, fields and orchards. Public or private entities may own farmland for the purpose of promoting agriculture, but individuals may only own farmland on the condition that they continue farming. In the case of the Republic of Korea, farmland must be transferred or leased to other farmers within one year if production of agricultural products is suspended.

Due to various reasons, such as aging of rural areas, decrease in agricultural population, exodus of existing farmers, inheritance of farmland by non-farmers, neglect after purchase, etc., idle farmland left idle is increasing. Therefore, whether farmland is an idle farmland where crops are currently grown or crops are stopped is a growing concern worldwide.

Idle farmland has both positive and negative aspects. On the positive side, restoration of natural biodiversity and increased uptake of water and carbon are representative examples. On the other hand, the negative aspects include soil erosion due to crop removal, desertification and increased potential for forest fires. In addition, if an alien species invades idle farmland, it may spread to nearby farmland used for agricultural purposes. Alternatively, some idle farmland is converted to building sites, or accommodates real estate development, further negatively impacting rural community production.

The condition of abandoned farmland (idle farmland) varies greatly, and it is not easy to identify common and representative characteristics. The existence and type of vegetation is heterogeneous, depending on the reason and duration of land abandonment, as well as local circumstances.

Some idle agricultural lands may experience a local ecological succession, turning into forests, bushes, grasslands or swamps. It can also be left in other barren conditions or cleared for human use.

Hereinafter, with reference to FIG. 1, the state of use of farmland parcels defined as idle farmland will be described.

1 is an exemplary view of the current state of use of farmland parcels defined as idle farmland.

Referring to FIG. 1, (a) is an example of a state in which idle farmland is used as a cemetery. (b) is an example of a state in which a forest is formed on idle farmland. (c) shows an example in which buildings are constructed on idle farmland and used as a city. The image of FIG. 1 can be obtained from high-resolution satellite images of Google Earth.

In this way, idle farmland is transformed into various forms, and the existence of vegetation and the type of vegetation are different. Therefore, it is difficult to derive common and representative characteristics of idle farmland from images.

In addition, idle farmland is judged to have been abandoned after a certain period without agricultural production has elapsed. Therefore, a cross-sectional analysis method using observed data for several variables at the same time or the same period is not suitable for detecting, classifying, or mapping farmland abandonment. Therefore, conventionally, in order to identify idle farmland, it is possible to more accurately identify changes only when a specific place is compared before and after for a long period of 10 years or more.

However, when comparing observations over a long period of 10 years or more to classify idle farmland as above, there is a problem in that it is not possible to identify idle farmland and take action immediately after it begins to be considered idle farmland.

An idle farmland detection method and apparatus according to an embodiment of the present invention is for detecting idle farmland in a short period of time after idle farmland is generated.

In addition, the method and apparatus for detecting idle farmland according to an embodiment of the present invention are for detecting idle farmland using frequency characteristics of vegetation indices.

However, the technical problem to be achieved by the present embodiment is not limited to the technical problem as described above, and other technical problems may exist.

As a technical means for achieving the above-described technical problem, a method for detecting idle farmland according to an embodiment of the present invention includes acquiring an image for a preset period of time in a search target area, and detecting idle farmland using the image. Inputting the input to a model to determine idle farmland within the search target area, wherein the idle farmland detection model is harmonized with respect to the vegetation index for each land unit in which the search target area included in the image is divided into preset land units. and a machine learning model learned to discriminate idle farmland in the search target area based on harmonic components derived by performing harmonic analysis.

In addition, the apparatus for detecting idle farmland according to an embodiment of the present invention includes a memory in which an idle farmland detection program is stored, and a processor that executes the idle farmland detection program, wherein the processor is configured in advance with respect to a search target area. An image for a period of time is obtained, and the image is input to an idle farmland detection model to determine idle farmland within the search target area. It includes a machine learning model learned to discriminate idle farmland within the search target area based on the harmonic component derived by performing harmonic analysis on the vegetation index for each land unit divided by .

In addition, a method for constructing an idle farmland detection model using an idle farmland detection apparatus according to an embodiment of the present invention includes acquiring an image for a pre-set period of time for a search target area, and detecting a search target area included in the image. Dividing into preset land units and deriving a vegetation index for each land unit, deriving a harmonic component by performing harmonic analysis on the vegetation index, and the search target based on the harmonic component and constructing an idle farmland detection model by learning to discriminate idle farmland in the zone.

In addition, in the idle farmland detection model building apparatus according to an embodiment of the present invention, it includes a memory in which an idle farmland detection program is stored, and a processor that executes the idle farmland detection program, wherein the processor, for a search target area , Acquiring images for a preset period, classifying the search target area into preset land units, deriving a vegetation index for each land unit, and performing harmonic analysis on the vegetation index to obtain a harmonic component. and learn to discriminate idle farmland in the search target area based on the harmonic component to build an idle farmland detection model.

The method and apparatus for detecting idle farmland according to an embodiment of the present invention can detect idle farmland in a short period of time after idle farmland is generated.

In addition, the method and apparatus for detecting idle farmland according to an embodiment of the present invention may detect idle farmland using frequency characteristics of vegetation indices.

1 is an exemplary view of the current state of use of farmland parcels defined as idle farmland.
2 is a block diagram of an idle farmland detection device according to an embodiment of the present invention.
3 is a conceptual diagram illustrating functions of a processor according to an embodiment of the present invention.
4 is a flowchart of a method for constructing an idle farmland detection model according to an embodiment of the present invention.
5 is a flowchart of a method for detecting idle farmland according to an embodiment of the present invention.
6 is an exemplary view of a sentinel true color image in Gwangyang.
7 is a land cover map surveying idle farmland, fields, and rice fields.
8 is an exemplary diagram of an R2 value adjusted based on land classification of an idle farmland detection model according to an embodiment of the present invention.
9A to 9C are variable coefficient tables of an idle farmland detection method according to an embodiment of the present invention.
10 is a graph of vegetation indices according to land classification of an idle farmland detection method according to an embodiment of the present invention.
11 is a classification result table according to the regular vegetation index.
12 is a classification result table according to the normal moisture index.
13 is a classification result table according to the soil-corrected vegetation index.
14 is a classification result table according to three vegetation indices.
15 is a table of results of comparison of classification performance according to higher order harmonic analysis and vegetation index.

Hereinafter, embodiments of the present invention will be described in detail so that those skilled in the art can easily practice the present invention with reference to the accompanying drawings. However, the present invention may be embodied in many different forms and is not limited to the embodiments described herein. And in order to clearly explain the present invention in the drawings, parts irrelevant to the description are omitted.

Throughout the specification, when a part is said to be "connected" to another part, this includes not only the case where it is "directly connected" but also the case where it is "electrically connected" with another element interposed therebetween. . In addition, when a certain component is said to "include", this means that it may further include other components without excluding other components unless otherwise stated.

In this specification, a "unit" includes a unit realized by hardware, a unit realized by software, and a unit realized using both. Further, one unit may be realized using two or more hardware, and two or more units may be realized by one hardware. Meanwhile, '~unit' is not limited to software or hardware, and '~unit' may be configured to be in an addressable storage medium or configured to reproduce one or more processors. Therefore, as an example, '~unit' refers to components such as software components, object-oriented software components, class components, and task components, processes, functions, properties, and procedures. , subroutines, segments of program code, drivers, firmware, microcode, circuitry, data, databases, data structures, tables, arrays and variables. Functions provided within components and '~units' may be combined into smaller numbers of components and '~units' or further separated into additional components and '~units'. In addition, the components and '~units' may be implemented to regenerate one or more CPUs in the device.

In addition, the accompanying drawings are only for easy understanding of the embodiments disclosed in this specification, the technical idea disclosed in this specification is not limited by the accompanying drawings, and all changes included in the spirit and technical scope of the present invention , it should be understood to include equivalents or substitutes.

Terms including ordinal numbers, such as first and second, may be used to describe various components, but the components are not limited by the terms. These terms are only used for the purpose of distinguishing one component from another.

It is understood that when an element is referred to as being "connected" or "connected" to another element, it may be directly connected or connected to the other element, but other elements may exist in the middle. It should be. On the other hand, when an element is referred to as “directly connected” or “directly connected” to another element, it should be understood that no other element exists in the middle.

Singular expressions include plural expressions unless the context clearly dictates otherwise.

In this application, terms such as "comprise" or "have" are intended to designate that there is a feature, number, step, operation, component, part, or combination thereof described in the specification, but one or more other features It should be understood that the presence or addition of numbers, steps, operations, components, parts, or combinations thereof is not precluded.

Hereinafter, the configuration of the idle farmland detection device 1 according to the embodiment will be described with reference to FIG. 2 .

2 is a block diagram of an idle farmland detection device according to an embodiment of the present invention.

Referring to FIG. 2 , the idle farmland detection device 1 includes a communication module 110 , a memory 120 and a processor 140 .

The communication module 110 includes components for transmitting and receiving data with external devices. The communication module 110 may refer to a device including hardware and software necessary for transmitting and receiving a signal such as a control signal or a data signal through a wired/wireless connection with another network device.

The communication module 110 may use all types of wireless communication networks, such as a mobile radio communication network or a satellite communication network. In addition, examples of wireless communication networks include 3G, 4G, 5G, 3rd Generation Partnership Project (3GPP), Long Term Evolution (LTE), World Interoperability for Microwave Access (WIMAX), Wi-Fi, Bluetooth communication, infrared communication, Ultrasonic communication, visible light communication (VLC: Visible Light Communication), LiFi, and the like are included, but are not limited thereto.

In addition, the idle farmland detecting apparatus 1 includes a memory 120 capable of storing an idle farmland detection program and data, and a processor 140 that performs a series of mathematical processing related to detecting idle farmland.

The memory 120 stores an idle farmland detection program. The name of the idle farmland detection program is set for convenience of description, and the name itself does not limit the function of the program. The memory 120 stores at least one of information and data input to the communication module 110, information and data necessary for functions performed by the processor 140, and data generated according to execution of the processor 140. can

The memory 120 should be interpreted as collectively referring to a non-volatile storage device that continuously maintains stored information even when power is not supplied and a volatile storage device that requires power to maintain stored information. Also, the memory 120 may temporarily or permanently store data processed by the processor 140 . The memory 120 may include magnetic storage media or flash storage media in addition to volatile storage devices that require power to maintain stored information, but the scope of the present disclosure is not limited thereto. no.

The idle farmland detection device 1 may further include a database 130 . The database 130 may refer to a configuration in which data for the operation of an idle farmland detection program is stored. The database 130 may constitute a part of the memory, but may be located outside of the idle farmland detection device 1, not necessarily inside.

The processor 140 is configured to execute an idle farmland detection program stored in the memory 120 . The processor 140 may include various types of devices that control and process data. The processor 140 may refer to a data processing device embedded in hardware having a physically structured circuit to perform functions expressed by codes or instructions included in a program. The processor 140 includes a microprocessor, a central processing unit (CPU), a processor core, a multiprocessor, an application-specific integrated circuit (ASIC), and a field programmable gate array (FPGA). ), etc., but the scope of the present disclosure is not limited thereto.

The idle farmland detection device 1 may build an idle farmland detection model that performs machine learning to detect idle farmland. Accordingly, the idle farmland detection device 1 may function as an idle farmland detection model building device. A specific function for constructing an idle farmland detection model and detecting idle farmland will be described in detail with reference to FIG. 3 to be described later.

3 is a conceptual diagram illustrating functions of the processor 140 . Hereinafter, the function of the processor 140 will be described with reference to FIG. 3 .

The processor 140 may build an idle farmland detection model that performs machine learning to detect idle farmland, and determine idle farmland within a detection target area by using the idle farmland detection model.

The processor 140 may include an image preprocessing module 141 , a vegetation index derivation module 142 , a harmonic analysis module 143 , and a machine learning module 144 . The processor 140 builds an idle farmland detection model through a data processing process using an image preprocessing module 141, a vegetation index derivation module 142, a harmonization analysis module 143, and a machine learning module 144, and an idle farmland can detect

The image preprocessing module 141 may obtain an image of a land area and perform preprocessing on the obtained image. In this case, the video may include a video, a still screen of the video, and a photo. The image may refer to a plurality of images that have a resolution equal to or higher than a preset standard and are accumulated and captured during a preset period. The higher the resolution of the image and the shorter the shooting period of the image, the higher the performance of identifying idle farmland.

For example, in the case of using satellite images, the satellite images have a resolution of 10m X 10m or more and may include satellite images captured every 10 days. However, the reference resolution and shooting cycle are not limited thereto, and may be set to various values according to the legal definition of idle farmland and the geographical characteristics of the search target area.

The image pre-processing module 141 may perform image pre-processing. For example, an image may be sampled using bi-linear interpolation. In addition, preprocessing may be performed to correct the resolution of the image or derive the vegetation index from the satellite image.

The vegetation index derivation module 142 derives a vegetation index (VI) for a search target area using the preprocessed image. The vegetation index is an index that derives vegetation characteristic information using signals in the wavelength range measured from satellites or aircraft, and is used to help monitor or understand the spatio-temporal change characteristics of vegetation.

The vegetation index derivation module 142 may divide the land area included in the image into preset land units and derive a vegetation index for each land unit. At this time, the land unit may mean a unit for dividing a land area into a certain standard (area), such as a parcel, a pyeong unit, a square meter (m 2 ) unit, and the like.

The vegetation index derivation module 142 selects one of a Normalized Difference Vegetation Index (NDVI), a Normalized Difference Water Index (NDWI), and a Soil Adjusted Vegetation Index (SAVI) for an image. abnormalities can be derived. Since abandoned farmland exists in various types, accuracy of detecting abandoned farmland can be improved by analyzing other vegetation indices together rather than using only the regular vegetation index.

The harmonic analysis module 143 may derive frequency features by performing harmonic analysis on the derived vegetation index. That is, the characteristics of each vegetation indicator type can be derived from the image using harmonic analysis. In addition, the harmonic analysis module 143 may derive frequency features by performing third or higher order harmonic analysis on the vegetation index.

Time series analysis of vegetation indices is commonly used to generate input data for classification. For analysis of vegetation indices, methods such as local asymmetric Gaussian function and double logistic can be used. In addition, a harmonic analysis method can be used for the analysis of the vegetation index.

Harmonic analysis corresponds to Fourier analysis in a narrow sense. Thus, harmonic analysis decomposes a time series into a sum of sinusoid terms of various frequencies to determine the morphological attributes of a vegetation growth trajectory or the functional shape of a vegetation index. (functional form) is derived mathematically.

Harmonic coefficients of each term estimated using harmonic analysis can be used as features for classification instead of commonly used phenology metrics. Examples of phenological metrics are the start, end, center, and length of the growing season and the basic and maximum Normalized Difference Vegetation Index (NDVI).

The advantage of features based on harmonic analysis compared to phenolic metrics is that they can be easily obtained without complex calculations. In addition, a variable number of features can be generated according to the observation interval, and most of them can be optional features for classification.

Studies related to conventional harmonic analysis have several commonalities. First, third-order or higher harmonic analysis was ignored because it simply showed noise or false deviation and was judged to have insignificant contribution to the overall systematic variability. Second, only limited types of vegetation indices were used for analysis.

The most frequently used of the vegetation indices is the normal vegetation indices. The normal vegetation index reflects the density of green space and photosynthetic activity. Therefore, the normal vegetation index is known as the most accurate indicator of ground-level biomass.

On the other hand, vegetation indices other than the normal vegetation indices among vegetation indices are not used for the purpose of this harmonization analysis.

For example, the Normalized Difference Water Index (NDWI) is often used as an indicator of water content in vegetation and is used to identify the presence of water bodies, water stress in plants, drought or land inundation. In addition, the Soil Adjusted Vegetation Index (SAVI) is used to determine plant health, particularly in arid regions where soil exposure is relatively high.

However, since abandoned farmland exists in various types as described above, the accuracy of detecting idle farmland can be improved by analyzing other vegetation indices together rather than using only the regular vegetation index.

For example, agricultural cultivation in temperate climates is complex, and double or multi-crop practices are common. In addition, as described above, idle farmland exists in various types depending on the type of vegetation, distribution, reason for abandonment of farmland, period, surrounding environment, and the like. Thus, lower-order harmonic components alone do not accurately capture more detailed growth trajectories.

In addition, during the normal rice growing process (irrigating the paddy field, transplanting the seedlings, and growing until harvest), not only the vegetation but also the soil and water content change simultaneously. Therefore, unlike many types of arable land, idle farmland cannot be explained by a single vegetation index, but can be explained only by analyzing other effects such as soil background reflectance and water content.

Accordingly, the processor 140 according to the embodiment may use various vegetation indices as well as regular vegetation indices and reduce a period for detecting idle farmland by using third or higher order harmonic analysis.

In addition, the harmonic analysis module 143 may derive harmonic components suitable for classifying idle farmland among the harmonic components included in the harmonic analysis result for the vegetation index using a recursive feature elimination algorithm.

The machine learning module 144 performs learning of an idle farmland detection model to detect idle farmland within a search target area by applying a machine learning algorithm based on the result of the harmonization analysis. That is, the machine learning module 144 performs learning of the idle farmland detection model using the harmonic component. In addition, the idle farmland in the search target area may be determined by inputting the harmonic component for the search target area to the learned idle farmland detection model. At this time, the standard of idle farmland may be set differently according to laws or preset standards.

The machine learning module 144 may partition the training and validation sets for harmonic components and perform classification using a Support Vector Machine (SVM). The machine learning module 144 may classify the frequency coefficients included in the harmonic component using the support vector machine, and perform learning of an idle farmland detection model based on the classified frequency coefficients.

The machine learning module 144 may utilize only harmonic coefficients, which are relatively easy to extract, without using phenolic matrix elements in machine learning classification.

In addition, the machine learning module 144 uses a set of coefficients of higher order harmonic components to simulate more detailed phenological trajectories. In addition, three vegetation indices (regular vegetation index, normal moisture index, and soil-adjusted vegetation index) are used together to supplement the limitations of short-term observation.

Therefore, the idle farmland detection model according to an embodiment of the present invention makes full use of parameters derived from vegetation harmonization analysis and machine learning classification to identify phonological trajectories in the short term of 3 years to distinguish idle farmland from active cropland. In addition, the idle farmland detection apparatus 1 according to an embodiment of the present invention may acquire an image of a detection target area and input the acquired image to an idle farmland detection model to determine idle farmland within the detection target area.

A specific method for detecting idle farmland will be described in detail with reference to FIGS. 4 to 15 to be described later.

Hereinafter, a method for detecting idle farmland according to an embodiment of the present invention will be described with reference to FIGS. 4 to 15 .

4 is a flowchart of a method for constructing an idle farmland detection model according to an embodiment of the present invention.

4, the idle farmland detection model construction method according to an embodiment of the present invention includes an image collection step (S100), a vegetation index derivation step (S200), a harmonization analysis step (S300), and an idle farmland detection model learning step (S200). S400).

In the image collection step ( S100 ), images of the land area may be collected and pre-processing may be performed on the images. In this case, the image may refer to a plurality of images having a resolution equal to or higher than a preset standard and accumulated and captured during a preset period. The higher the resolution of the image and the shorter the shooting period, the higher the idle farmland discrimination performance.

For example, in the image collection step S100, a remotely sensed image of a sentinel-2A satellite may be used.

6 is an exemplary view of a sentinel true color image in Gwangyang.

The method for constructing an idle farmland detection model according to an embodiment of the present invention may use a remotely sensed image (image) of a sentinel-2A satellite. Satellite remote-sensing images of Sentinel-2A are multispectral images, repeated at 10 m, 20 m, and 60 m spatial resolutions over a 10-day cycle. This high spatio-temporal resolution is suitable for detecting urban small-scale idle farmland.

Specifically, all Sentinel-2A level 1C (L1C) 52SCD images from three years (e.g., January 2016 to December 2018) were used to construct time-series data using remote sensing images of Sentinel-2A. can

In addition, the method for constructing an idle farmland detection model according to an embodiment of the present invention may pre-process the Sentinel-2A image. The idle farmland detection method according to the embodiment may generate Sentinel-2A Level 2A (L2A) data by performing an image processing process.

Specifically, to generate Sentinel-2A L2A data, Sentinel 2A L1C data was atmospherically modified using the Sen2Cor (version 2.5.5) processing tool based on the LIBRADTRAN radiative transfer model. corrected) steps can be performed.

The LIBRADTRAN radiative transfer model can expedite the atmospheric correction process by using large look-up-table sensor-specific features that account for various atmospheric conditions and solar geometries.

It also uses internal Scene Classification Layer (SCL) images and atmospherically corrected clear pixels generated in Sen2Cor to remove cloud and snow-contaminated pixels. In order to equally use the multispectral image at 10 m, the band 11 image may be resampled with a grid size of 20 m to 10 m using bi-linear interpolation.

In the vegetation index derivation step ( S200 ), the land area included in the image may be divided into preset land units, and a vegetation index may be derived for each land unit. At this time, the land unit may mean a unit for dividing the search target area into a certain standard (area), such as a parcel, a pyeong unit, a square meter (m 2 ) unit, and the like.

In addition, in the vegetation index derivation step (S200), with respect to the image, a Normalized Difference Vegetation Index (NDVI), a Normalized Difference Water Index (NDWI), and a Soil Adjusted Vegetation Index (SAVI) are selected for the image. One or more of them can be derived.

For example, vegetation indices (normal vegetation indices, normal moisture indices, and soil-adjusted vegetation indices) are extracted from the preprocessed Sentinel-2A Level 2A (L2A) images. For the preprocessed satellite image as described above, normal vegetation index, normal moisture index, and soil correction vegetation were obtained using Band 4 (red), Band 8 (NIR: Near infrared), and Band 11 (SWIR: Shortwave Infrared). derive the index.

In the harmonic analysis step (S300), frequency characteristics may be derived by performing harmonic analysis on the derived vegetation index. That is, the characteristics of each vegetation indicator type can be derived from satellite imagery using harmonic analysis. In addition, the harmonic analysis module 143 may derive frequency features by performing third or higher order harmonic analysis on the vegetation index.

In addition, in the harmonic analysis step (S300), three types of vegetation based on R2 corrected using the average value of each vegetation index (normal vegetation index, normal moisture index, and soil-corrected vegetation index) for all image samples individually in each class. Identify the sinusoidal component that best fits the vegetation index. In addition, these phenomena can be visualized.

In addition, in the harmonic analysis step (S300), a recursive feature elimination algorithm is used to find the best subset of unique harmonic components of vegetation indices for classification of idle farmland. Then, the harmonic coefficient of each vegetation index per individual lot is extracted.

The combination of harmonized components of a vegetation index that fulfills the objective of best describing tree growth and the objective of best classifying idle farmland can be different.

For the purpose of best description of tree growth, a number of parameters can be used since an accurate description is important.

However, for the purpose of best classifying idle farmland, prediction is important, so fewer combinations of harmonic components can be used to avoid overfitting.

In addition, in the harmonic analysis step (S300), harmonic time-series analysis may be performed on the vegetation index derived using the image.

Specifically, harmonic regression analysis is a process of decomposing or expressing a time series as an additive sum of several sinusoidal components. Essentially, this time-domain approach regresses the current state of the cycle for multiple frequency sine and cosine terms, with the goal of finding the relative strength of each term.

These time-series analysis methods include reconstructing plant phenology, mapping and change detections of land cover, inventorying the national forest, and classifying plant types or land cover. (classification of plant types or land covers).

Since the Fourier frequencies in the Fourier transformation, the process of converting a time-domain function into a frequency-domain function, are often specified orthogonally, fitting is simpler with ordinary-least-square strategies. can The range fk of the frequency domain may be defined as in Equation 1 described below.

[Equation 1]

Here, n is the length or number of points sampled in a cycle, and k is an integer. In addition, each vegetation index can be expressed as an additional construct of a pair of sine and cosine terms with the time trend and intercept given by Equation 2 below.

[Equation 2]

는 시간 t에 기록된 식생지수를 의미하며, 본 발명의 실시예에서는 세 종류의 식생지수 (NDVI, NDWI, SAVI)중 어느 하나 이상을 의미할 수 있다. 식생지수의 시계열은 선형, 혹은 비선형 추세 (trend)와 계절성 (seasonality)로 이루어진다. 본 발명의 실시예에서는 식생지수의 비선형 추세에 해당하는 변수는 모형에 포함시키지 않을 수 있다. Time은 시간을 의미하며, 식생지수의 시계열적 변화에 선형 추세가 존재하면

계수의 통계적 유의성이 나타난다.

는 식생지수의 계절성을 의미하는 변수들의 조합이며, 다양한 주파수를 가진 사인 및 코사인 곡선의 합으로 이루어진다. 여기서 가장 높은 주파수는 1/2 가장 낮은 주파수는 1/n 이다. 각 주파수의 사인 및 코사인 곡선이 전체 시계열을 이루는 요소이면

의 통계적 유의성이 나타난다. here,

Means a vegetation index recorded at time t, and may mean any one or more of three types of vegetation indices (NDVI, NDWI, SAVI) in an embodiment of the present invention. The time series of vegetation indices consists of linear or non-linear trends and seasonality. In an embodiment of the present invention, the variable corresponding to the non-linear trend of the vegetation index may not be included in the model. Time means time, and if there is a linear trend in the time-series change of the vegetation index,

Statistical significance of coefficients is shown.

is a combination of variables meaning the seasonality of the vegetation index, and consists of the sum of sine and cosine curves with various frequencies. Here, the highest frequency is 1/2 and the lowest frequency is 1/n. If the sine and cosine curves at each frequency are elements of the entire time series, then

Statistical significance of is shown.

Since satellite imagery can be captured at 10-day intervals, a yearly cycle can be sampled with 36 viewpoints. In this case, the highest and lowest possible frequencies are 1/36 and 18/36, respectively, and the maximum number of sinusoidal pairs is 18. The highest frequency, the 18th-order, is deleted due to multicollinearity in the analysis, so that up to 17 pairs can be used for analysis.

In the idle farmland detection model learning step (S400), an idle farmland detection model is built by learning to discriminate idle farmland based on the result of the harmonization analysis. For example, in the idle farmland detection model learning step (S400), support vector machine classification is performed by inputting a combination of harmonic components. In addition, classification performance can be compared using overall accuracy, Kappa coefficient, and accuracy of producers and users for detecting idle farmland.

In the idle farmland detection model learning step (S400), a frequency coefficient classification step using a support vector machine may be included.

Specifically, support vector machines are one of several supervised classification strategies used with Decision Trees, Random Forests, Maximum Likelihood and Artificial Neural Networks.

Support vector machines classify data by applying a hyperplane that divides the data into multiple classes of homogeneous features. Among many possible hyperplanes, we choose the hyperplane with the largest margin for the closest data observations. By specifying the cost of erroneous classification of some noise and outliers, support vector machines can avoid overfitting.

Additionally, to classify linearly inseparable data, support vector machines use the kernel trick of forming hyperplanes for higher dimensional or Gaussian radial basis functions. Adopt. Support vector machines are used to detect idle farmland with various types of input attributes.

For training of the idle farmland classification model, the reference data can be split into training and test sets with a ratio of 7:3. First, we select a subset of features that maximize classification performance for each of the plurality of vegetation indices, and then tune the model with three kernel tricks: linear, polynomial, and radial functions using different sets of cost, gamma, and order.

The model performance of the idle farmland detection model generated as above can be investigated based on overall accuracy, kappa correlation coefficient, and accuracy of producers and users.

5 is a flowchart of a method for detecting idle farmland according to an embodiment of the present invention.

Referring to FIG. 5 , the method for detecting idle farmland according to an embodiment of the present invention includes a search target area image collection step (S110), an idle farmland detection model input step (S210), and an idle farmland discrimination step (S310).

In the search target area image collection step (S110), an image of the search target area is collected. In this case, the image may refer to a plurality of images having a resolution equal to or higher than a preset standard and accumulated and captured during a preset period. The higher the resolution of the image and the shorter the shooting period, the higher the idle farmland discrimination performance.

In addition, as described above in the image collection step (S100), in the search target area image collection step (S110), a remotely sensed image of a sentinel-2A satellite may be used.

In the idle farmland detection model input step (S210), the image of the search target region obtained in the search target region image collection step (S110) is input to the idle farmland detection model. In the idle farmland discrimination step (S310), an idle farmland discrimination result for the search target area is derived using the learned idle farmland detection model. That is, pre-processing, derivation of vegetation index, and harmonization analysis are performed on the input image of the search target area. In addition, idle farmland in the search target area is determined based on the harmonic component of the search target area derived as a result of the harmonization analysis.

At this time, the preprocessing, vegetation index derivation, and harmony analysis may be performed using the same method as the method performed in the above-described image collection step (S100), vegetation index derivation method (S200), and harmony analysis step (S300).

Accordingly, the idle farmland detection model may sample images using bi-linear interpolation. In addition, preprocessing may be performed to correct the resolution of the image or derive the vegetation index from the satellite image.

The idle farmland detection model may generate Sentinel-2A Level 2A (L2A) data by performing an image processing process when a remote sensing image of a sentinel-2A satellite is input.

The idle farmland detection model uses the Sen2Cor (version 2.5.5) processing tool based on the LIBRADTRAN radiative transfer model to generate the Sentinel-2A L2A data, while the Sentinel 2A L1C data is queued. Atmospherically corrected steps can be performed.

In addition, the idle farmland detection model uses the LIBRADTRAN radiative transfer model to expedite the atmospheric correction process by using large look-up-table sensor-specific features to account for various atmospheric conditions and solar geometries. can be done

In addition, the idle farmland detection model uses internal Scene Classification Layer (SCL) images and atmospherically corrected clear pixels generated by Sen2Cor to remove cloud and snow contaminated pixels. do. In order to equally use the multispectral image at 10 m, the band 11 image may be resampled with a grid size of 20 m to 10 m using bi-linear interpolation.

The idle farmland detection model uses the input image of the search target area and the preprocessed image to determine the normalized difference vegetation index (NDVI), normalized difference water index (NDWI), and soil correction vegetation in the search target area. One or more of the indexes (SAVI: Soil Adjusted Vegetation Index) may be derived.

The idle farmland detection model may derive a harmonic component by performing harmonic analysis on the derived vegetation index. That is, the frequency characteristics of each vegetation indicator type can be derived from the image of the search target area using harmonic analysis. In addition, the idle farmland detection model may derive frequency features by performing third or higher order harmonic analysis on vegetation indices.

The idle farmland detection model determines idle farmland in a search target area based on frequency characteristics of the search target area. Hereinafter, result values of the method for detecting idle farmland according to an embodiment of the present invention will be described with reference to FIGS. 6 to 15 .

Referring to FIG. 6 described above, a district of Gwangyang is displayed using a sentinel image. In order to learn and test the idle farmland detection method according to the embodiment, survey data from Gwangyang, Jeollanam-do were used. Specifically, Gwangyang City is located at coordinates 34.9407°N, 127.6959°E at the southern tip of the Korean Peninsula. The urban area is 1,224,700 ha, and in Korea, which has a territory of about 10,030,000 ha, 1,621,000 ha (16.16%) was defined as farmland in 2018, a decrease of 1.4% from the previous year. Also, Jeollanam-do has more agricultural land than any other province.

To construct the main data set, Sentinel-2A L1C 52SCD images from Gwangyang City from January 2016 to December 2018 were used. Also, the cloud area of the image is less than 30% of the total image area. 16, 25, and 41 Sentinel-2A L1C images were acquired in 2016, 2017, and 2018, respectively. Since it is in the overlapped swath of images for two consecutive orbits, more than 36 images can be acquired in one year, and considering even intervals, a total of 68 images were used for analysis .

7 is a land cover map surveying idle farmland, fields, and rice fields.

Reference data for comparing idle farmland detection results and actual idle farmland according to an embodiment of the present invention are land cover maps published by the Ministry of Environment in 2013 and 2018, December 2018 at the Korea Spatial Information Portal The published continuous cadastral map and the survey of abandoned farmland by the Korea Rural Community Corporation are used.

As shown in FIG. 7, the mapping process of the land cover map is as follows. First, high-resolution satellite images (1m resolution, 0.7m resolution) and aerial photographs of 2016 and 2017 are collected, and the land cover status is identified using visual analysis. In addition, agency personnel visit unidentified parcels to confirm initial assessments.

Since this verification process is time consuming and budget intensive, since 2010 land cover mapping projects have been conducted annually for one state or a few cities. Annual data collection was not possible for the same reason.

The reference dataset contains 27,645 land parcels, of which 11,411 parcels, 6777 parcels and 9457 parcels are idle farmland, rice paddy, and agricultural (dry) fields, respectively. Rice fields and agricultural land are defined as designated by land cover maps in both 2013 and 2018. In addition, idle farmland is defined as parcels detected in the 2013 survey that still show no signs of agricultural production based on the land cover in 2018.

8 is an exemplary diagram of an R2 value adjusted based on land classification of an idle farmland detection model according to an embodiment of the present invention.

9A to 9C are variable coefficient tables of an idle farmland detection method according to an embodiment of the present invention.

Referring to Figure 8, it shows the R2 value according to land classification from the idle farmland classification method according to an embodiment of the present invention. Cells marked in black are the variables selected for the model. In this case, the variables refer to the optimal 35 variables with different sets of harmonic components. The y-axis represents adjusted R2. Detailed coefficients of variables describing the trajectory of each vegetation index are shown in FIGS. 9A to 9C.

First, referring to (1), (2), and (3) of FIG. 8, the harmonization analysis of the regular vegetation index shows that the maximum possible adjustment R2 for each grade (idle farmland, field, rice field) is 0.80, 0.76, and 0.86, respectively. suggests that Among the total of 35 variables included in the analysis (intercept + 17 pairs of harmonic components), higher-order sinusoidal coefficients exceeding the third are also included in the best-fitting model.

Referring to (4), (5), and (6) of FIG. 8, in the case of the normal moisture index, the highest average adjusted R2 among the three vegetation indices of 0.95, 0.95, and 0.90 for idle farmland, field, and rice paddy, respectively have

In addition, referring to (8), (9), and (10) of FIG. 8, the result of analyzing the soil-corrected vegetation index is similar to the regular vegetation index analysis result. As shown in the normal vegetation index, the soil-adjusted vegetation index is most useful for fitting paddy fields. R2 corresponding to idle land and field rice fields are 0.87, 0.82, and 0.90.

Therefore, it is more effective to use high-order sine waves together than to use only low-order components to adjust the growth cycle of a heterogeneous mixture of plants in idle farmland. For example, in the case of the normal vegetation index for rice fields, components of 3rd or higher order are important to improve the fit because the adjusted R2 decreased from 0.86 to 0.59.

10 is a graph of vegetation indices according to land classification of an idle farmland detection method according to an embodiment of the present invention.

Referring to FIG. 10 , values displayed in black are estimated values using the idle farmland detection model in FIG. 8 , and values displayed in red are observed values. The general shape of the nine curves deviate greatly from simple sine and cosine curves, requiring higher order frequencies for simulation.

This may be due to complex cultivation procedures adopted in temperate climates. In addition, as can be seen from the above-mentioned results, the normal vegetation distribution and the phenomenon expressed as the soil-adjusted vegetation index are similar to each other, so the similarity of the two indices is confirmed.

In the case of the normalized vegetation index and the soil-adjusted vegetation index of idle farmland, the values of the normalized vegetation index and the soil-adjusted vegetation index of the other two land uses (fields and paddy fields) were overall higher. In addition, idle farmland tends to increase in value earlier and last longer than the other two lands, and falls during the cooler season.

Among them, the normalized vegetation index of rice fields draws the opposite trajectory in that most of them remain low, and high values appear only for a short period of time.

On the other hand, the normal moisture index shows a different pattern from the other two indices. The field's normal moisture index amplitude is the smallest among the nine amplitudes, and the confidence interval is relatively narrow. Thus, the division of fields can be well represented by normal moisture indices. This indicates that the characteristics extracted from the vegetation index can be useful for classification by representing characteristics unique to each class.

The accuracy of the three vegetation indices will be described with reference to FIGS. 11 to 15 .

Among the three types of kernel tricks, the radial kernel has the best performance for all vegetation indices, while the cost and gamma values vary slightly from model to model. In general, overall accuracy is highest when soil-adjusted vegetation indices are used and lowest when normal moisture indices are used.

In addition, the accuracy of producers and users is higher than that of fields in classifying idle farmland and paddy fields. The three models with the best subsets of harmonic components are implemented by Equations 3, 4 and 5 below. Equation 3 represents the normal vegetation index, Equation 4 represents the normal moisture index, and Equation 5 represents the harmonic analysis formula for the soil-corrected vegetation index.

[Equation 3]

[Equation 4]

[Equation 5]

Equations 3, 4, and 5 are results of fitting Equation 2 to three vegetation index (NDVI, NDWI, and SAVI) data, and are models in which only variables showing statistical significance among input variables are taken and recombined.

Commonly valid harmonized components for classification of paddy fields, fields, and idle farmland are first-order cosine, second-order sine, third-order sine, and fourth-order cosine.

The duration of the primary component is 36 time points, meaning that the cycle is 1 year. The duration of the secondary component is 18 time points, which means that the cycle is half a year. Similarly, the duration of the 6th component is a period of 6 time points, i.e. a period of 2 months.

11 is a classification result table according to the regular vegetation index.

Referring to FIG. 11, the support vector machine classification results using the normal vegetation index are shown. The best performance is obtained with a kernel trick cost of 6 and a gamma of 0.1. The overall prediction accuracy is 86.97% and the Kappa index is 0.799. The accuracy of classifying idle farmland, field, and rice field was 90.57%, 76.92%, and 89.55%, respectively, in terms of producer's accuracy.

12 is a classification result table according to the normal moisture index.

Referring to FIG. 12, the support vector machine classification results using the normal moisture index are shown. The overall prediction accuracy was slightly lower at 85.29%, with a kappa index of 0.775. However, the idle farmland discrimination performance is 90.67%, which is higher than the regular vegetation index result.

13 is a classification result table according to the soil-corrected vegetation index.

Referring to FIG. 13, the support vector machine classification results using the soil-corrected vegetation index are presented. In terms of overall accuracy and kappa index, the classifier using the soil-adjusted vegetation index performs better than the other two classifiers. The accuracy of producers and users for idle farmland is also the highest at 91.30% and 91.68%, respectively.

14 is a classification result table according to three vegetation indices.

Referring to FIG. 14 , values obtained by finding a subset with the best performance using all harmonic components for three vegetation indices and then performing classification again are shown. First, we take up to 6-order components from the three vegetation indices and then run a recursive feature removal algorithm to derive the best subsets. The final model is implemented using a function such as Equation 6.

[Equation 6]

The meaning of the variables and coefficients in Equation 6 is the same as in Equation 2.

The detection performance of idle farmland shows improved values in all criteria, such as overall accuracy of 120.72%, kappa index of 0.858, producer accuracy of 122.67%, and user's accuracy of 93.40%.

In addition, the unique properties contributed by the normal moisture index and the soil-corrected vegetation index result in improved performance compared to classification performed with the normal vegetation index alone.

15 is a table of results of comparison of classification performance according to higher order harmonic analysis and vegetation index.

Referring to FIG. 15, it is shown that the configuration of the higher harmonic component and the three vegetation indices improves the classification performance. Comparative models include only up to third harmonic components. As shown in FIG. 15, rather than discriminating idle farmland using only one of a regular vegetation index, a regular moisture index, and a soil-corrected vegetation index. In the case of discriminating idle farmland using all three vegetation indices, the accuracy is higher.

Unlike idle farmland, fallow land is cultivated land where production has been deliberately stopped to restore soil fertility. Therefore, fallow land is still managed and land classification is relatively difficult. However, the method for detecting idle farmland according to the embodiment can clearly distinguish idle farmland from idle farmland even in a short period of observation.

An embodiment of the present invention may be implemented in the form of a recording medium including instructions executable by a computer, such as program modules executed by a computer. Computer readable media can be any available media that can be accessed by a computer and includes both volatile and nonvolatile media, removable and non-removable media. Also, computer readable media may include both computer storage media and communication media. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data.

Although the methods of the present invention have been described with reference to specific embodiments, some or all of their components or operations may be implemented using a computer system having a general-purpose hardware architecture.

The above description of the present invention is for illustrative purposes, and those skilled in the art can understand that it can be easily modified into other specific forms without changing the technical spirit or essential features of the present invention. will be. Therefore, the embodiments described above should be understood as illustrative in all respects and not limiting. For example, each component described as a single type may be implemented in a distributed manner, and similarly, components described as distributed may be implemented in a combined form.

The scope of the present invention is indicated by the following claims rather than the detailed description above, and all changes or modifications derived from the meaning and scope of the claims and equivalent concepts should be construed as being included in the scope of the present invention. do.

1: idle farmland detection device 110: communication module
120: memory 130: database
140: processor

Claims (22)

In the idle farmland detection method using an idle farmland detection device,
Obtaining an image for a preset period of time for a search target area, and
Including inputting the image to an idle farmland detection model to determine idle farmland in the search target area,
The idle farmland detection model,
Discrimination of idle farmland in the search target area based on the harmonic component derived by performing harmonic analysis on the vegetation index for each land unit in which the search target area included in the image is divided into preset land units A method for detecting idle farmland, which is a machine learning model trained to do. According to claim 1,
The vegetation index is,
A method for detecting idle farmland, including any one or more of Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), and Soil Adjusted Vegetation Index (SAVI). According to claim 1,
The harmonization analysis,
A method for detecting idle farmland, including third-order or higher-order harmonic analysis. According to claim 1,
The idle farmland detection model,
Idle farmland detection method of deriving a harmonic component for classifying idle farmland among the harmonic components included in the vegetation index using a recursive feature elimination algorithm. According to claim 1,
Acquiring the image is
Acquiring satellite images with a resolution equal to or greater than a preset minimum resolution and at period intervals less than a preset maximum duration
Including, idle farmland detection method. According to claim 5,
Acquiring the satellite image,
Sampling the satellite image using bi-linear interpolation
Including, idle farmland detection method. According to claim 1,
The idle farmland detection model,
Classifying frequency coefficients included in the harmonic component using a Support Vector Machine (SVM);
To perform learning of the idle farmland detection model based on the classified frequency coefficient, idle farmland detection method. In the idle farmland detection device,
A memory in which an idle farmland detection program is stored, and
Includes a processor that performs an idle farmland detection program;
the processor,
Obtaining an image for a preset period of time with respect to the search target area;
Inputting the image to an idle farmland detection model to determine idle farmland in the search target area,
The idle farmland detection model,
Discrimination of idle farmland in the search target area based on the harmonic component derived by performing harmonic analysis on the vegetation index for each land unit in which the search target area included in the image is divided into preset land units Idle farmland detection device, which is a learned machine learning model of the idle farmland detection model so as to do so. According to claim 8,
the processor,
Deriving any one or more of a Normalized Difference Vegetation Index (NDVI), a Normalized Difference Water Index (NDWI), and a Soil Adjusted Vegetation Index (SAVI) for the image, Idle farmland detection device. According to claim 8,
the processor,
To perform a third or higher order harmonic analysis on the vegetation index, the idle farmland detection device. According to claim 8,
the processor,
Idle farmland detection device to derive a harmonic component for classification of idle farmland among the harmonic components included in the vegetation index using a recursive feature elimination algorithm. According to claim 8,
the processor,
Having a resolution equal to or greater than a preset minimum resolution, and acquiring satellite images at period intervals less than a preset maximum period, an idle farmland detection device. According to claim 12,
the processor,
To sample the satellite image using bi-linear interpolation, idle farmland detection device. According to claim 8,
the processor,
Classifying frequency coefficients included in the harmonic component using a Support Vector Machine (SVM);
To perform learning of the idle farmland detection model based on the classified frequency coefficient, idle farmland detection device. In the idle farmland detection model construction method using the idle farmland detection model building device,
Obtaining an image for a preset period of time in a search target area;
Dividing the search target area included in the image into preset land units and deriving a vegetation index for each land unit;
Deriving a harmonic component by performing harmonic analysis on the vegetation index, and
Building an idle farmland detection model by learning to discriminate idle farmland in the search target area based on the harmonic component
Containing, idle farmland detection model construction method. According to claim 15,
Deriving the vegetation index,
Deriving any one or more of Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), and Soil Adjusted Vegetation Index (SAVI).
Containing, idle farmland detection model construction method. According to claim 15,
The harmonization analysis,
A method for constructing an idle farmland detection model, including third-order or higher-order harmonic analysis. According to claim 15,
The step of deriving the harmonic component,
An idle farmland detection model construction method of deriving a harmonic component for classifying idle farmland among the harmonic components included in the vegetation index using a recursive feature elimination algorithm. According to claim 15,
Acquiring the image is
Acquiring satellite images with a resolution equal to or greater than a preset minimum resolution and at period intervals less than a preset maximum duration
Containing, idle farmland detection model construction method. According to claim 19,
Acquiring the satellite image,
Sampling the satellite image using bi-linear interpolation
Containing, idle farmland detection model construction method. According to claim 15,
The step of building the idle farmland detection model,
Classifying frequency coefficients included in the harmonic component using a Support Vector Machine (SVM), and
Building an idle farmland detection model by learning to discriminate idle farmland in the search target area based on the classified frequency coefficient
Containing, idle farmland detection model construction method. In the idle farmland detection model building device,
A memory in which an idle farmland detection program is stored, and
Includes a processor that performs an idle farmland detection program;
the processor,
Obtaining an image for a preset period of time with respect to the search target area;
Divide the search target area into preset land units, derive a vegetation index for each land unit,
Deriving a harmonic component by performing harmonic analysis on the vegetation index,
An idle farmland detection model building apparatus for constructing an idle farmland detection model by learning to discriminate idle farmland in the search target area based on the harmonic component.

Images