CNN 모델과 Transformer 조합을 통한 토지피복 분류 정확도 개선방안 검토

심우담,이정수 한국지리정보학회 2024 한국지리정보학회지 Vol.27 No.1

본 연구는 Transformer 모듈을 기반으로 다양한 구조의 모델을 구성하고, 토지피복 분류를 수행하여 Transformer 모듈의 활용방안 검토를 목적으로 하였다. 토지피복 분류를 위한 딥러닝 모델은 CNN 구조를 가진 Unet 모델을 베이스 모델로 선정하였으며, 모델의 인코더 및 디코더 부분을 Transformer 모듈과 조합하여 총 4가지 딥러닝 모델을 구축하였다. 딥러닝 모델의 학습과정에서 일반화 성능 평가를 위해 같은 학습조건으로 10회 반복하여 학습을 진행하였다. 딥러닝 모델의 분류 정확도 평가결과, 모델의 인코더 및 디코더 구조 모두 Transformer 모듈을 활용한 D모델이 전체 정확도 평균 약 89.4%, Kappa 평균 약 73.2%로 가장 높은 정확도를 보였다. 학습 소요시간 측면에서는 CNN 기반의 모델이 가장 효율적이었으나 Transformer 기반의 모델을 활용할 경우, 분류 정확도가 Kappa 기준 평균 0.5% 개선되었다. 차후, CNN 모델과 Transformer의 결합과정에서 하이퍼파라미터 조절과 이미지 패치사이즈 조절 등 다양한 변수들을 고려하여 모델을 고도화 할 필요가 있다고 판단된다. 토지피복 분류과정에서 모든 모델이 공통적으로 발생한 문제점은 소규모 객체들의 탐지가 어려운 점이었다. 이러한 오분류 현상의 개선을 위해서는 고해상도 입력자료의 활용방안 검토와 함께 지형 정보 및 질감 정보를 포함한 다차원적 데이터 통합이 필요할 것으로 판단된다. This research aimed to construct models with various structures based on the Transformer module and to perform land cover classification, thereby examining the applicability of the Transformer module. For the classification of land cover, the Unet model, which has a CNN structure, was selected as the base model, and a total of four deep learning models were constructed by combining both the encoder and decoder parts with the Transformer module. During the training process of the deep learning models, the training was repeated 10 times under the same conditions to evaluate the generalization performance. The evaluation of the classification accuracy of the deep learning models showed that the Model D, which utilized the Transformer module in both the encoder and decoder structures, achieved the highest overall accuracy with an average of approximately 89.4% and a Kappa coefficient average of about 73.2%. In terms of training time, models based on CNN were the most efficient. however, the use of Transformer-based models resulted in an average improvement of 0.5% in classification accuracy based on the Kappa coefficient. It is considered necessary to refine the model by considering various variables such as adjusting hyperparameters and image patch sizes during the integration process with CNN models. A common issue identified in all models during the land cover classification process was the difficulty in detecting small-scale objects. To improve this misclassification phenomenon, it is deemed necessary to explore the use of high-resolution input data and integrate multidimensional data that includes terrain and texture information.

RapidEye 위성영상과 Semantic Segmentation 기반 딥러닝 모델을 이용한 토지피복분류의 정확도 평가

심우담,임종수,이정수 대한원격탐사학회 2023 大韓遠隔探査學會誌 Vol.39 No.3

The purpose of thisstudy wasto construct land cover maps using a deep learning model andto select the optimal deep learning model for land cover classification by adjusting the dataset such asinput image size and Stride application. Two types of deep learning models, the U-net model and theDeeplabV3+ model with an Encoder-Decoder network, were utilized. Also, the combination of the twodeep learning models, which is an Ensemble model, was used in thisstudy. The dataset utilized RapidEyesatellite images as input images and the label images used Raster images based on the six categories ofthe land use of Intergovernmental Panel on Climate Change as true value. This study focused on theproblem of the quality improvement of the dataset to enhance the accuracy of deep learning model andconstructed twelve land cover maps using the combination of three deep learning models (U-net,DeeplabV3+, and Ensemble), two input image sizes (64 × 64 pixel and 256 × 256 pixel), and two Strideapplication rates (50% and 100%). The evaluation of the accuracy of the label images and the deeplearning-based land cover mapsshowed that the U-net and DeeplabV3+ models had high accuracy, withoverall accuracy values of approximately 87.9% and 89.8%, and kappa coefficients of over 72%. Inaddition, applying the Ensemble and Stride to the deep learning models resulted in a maximum increaseof approximately 3% in accuracy and an improvement in the issue of boundary inconsistency, which isa problem associated with Semantic Segmentation based deep learning models.

학술연구논문 포스터발표 : 1분과: 산림환경, 산림생태, 수목학, 분류, 조림 ; 공간스케일 변화에 따른 생물다양성평가 -강원도 백두대간 지역을 대상으로-

심우담,이정수 한국임학회 2013 한국임학회 학술발표논문집 Vol.2013 No.-

Development of a Semi-automatic Search Program for Crown Delineation Based on Watershed and Valley Following Algorithms

심우담,박정묵,이정수 강원대학교 산림과학연구소 2018 Journal of Forest Science Vol.34 No.2

This paper discusses the development of semi-automatic search program for crown delineation in stand level. The crown of an individual tree was delineated by applying the Watershed (WS) and Valley Following (VF) algorithms. Unmanned Aerial Vehicle (UAV) images were used in the semi-automatic search program to delineate the crown area. The overall accuracy and Khat were used in accuracy assessment. WS algorithm’s model showed the overall accuracy and Khat index of 0.80 and 0.59, respectively, in Plot 1. However, the overall accuracy and Khat of VF algorithm’s model were 0.78 and 0.51, respectively, in Plot 2.

공간스케일 변화에 따른 생물다양성 평가 - 강원도 백두대간 보호구역을 대상으로 -

심우담,박진우,이정수 강원대학교 산림과학연구소 2014 Journal of Forest Science Vol.30 No.1

This research was conducted on the conservation area of Baekdudaegan, Kangwondo under the purpose of evaluating bio-diversity according to the changes of spatial scale, using GIS data and spatial filtering method. The diversity index was calculated based on the information of species of The 5th forest type map using Shannon-weaver index (H'), evenness index (Ei) and richness index (Ri). The diversity index was analyzed and compared according to the changes of 12 spatial scales from Kernel size 3x3 to 73x73 and basin unit. As for H' and Ri, spatial scale increased as diversity index decreased, while Ei decreases gradually. H' and Ri was highest; each 1.1 and 0.6, when the Kernel size was 73x73, while Ei was 0.2, the lowest. When you look at according to the basin unit, for large basin unit, ‘YeongDong’ region shows higher diversity index than ‘YeongSeo’ region. For middle basin unit, ‘Gangneung Namdaecheon’ region, and for small basin unit, ‘Gangneung Namdaecheon’ and ‘Gangneung Ohbongdaem’ region shows high diversity index. When you look at the relationship between diversity index and Geographic factors, H' shows positive relation to curvature and sunshine factor while shows negative to elevation, slope, hillshade, and wetness index. Also Ei was similar to the relationship between H' and Geographic factor. Meanwhile, Ri shows positive relationship to curvature and sunshine factor, while negative to elevation, slope, hillshade, and wetness index. macro unit diversity index evaluation was possible through the GIS data and spatial filtering, and it can be a good source for local biosphere conservation policy making.

토지이용변화 매트릭스 구축을 위한 국가공간주제도 활용방안에 관한 연구

심우담 ( Woo-dam Sim ),박정묵 ( Jeong-mook Park ),이정수 ( Jung-soo Lee ) 한국산림경제학회 2017 산림경제연구 Vol.24 No.2

본 연구는 주요 선진국의 국가 인벤토리 보고서 (National Inventory Report; NIR)를 통하여 우리나라의 토지이용 범주별 정의 및 토지이용 변화를 파악하기 위한 매 트릭스 구축 방법에 대하여 비교·분석하였다. IPCC 가이드라인에 의하면, LULUCF 분야의 온실가스 흡수 및 배출량 통계는 산림지, 농경지, 초지, 습지, 정주지, 기타토지의 6가지 범주로 구분하고 있으며, 각 국가 토지 상황에 맞게 토지이용 범주별 세부항목을 다르게 정의하고 있다. 일본의 경우, 산림지에 대해서 Approach(App.)3 수준의 토지이용 변화 매트릭스를 작성하고 있으며, 산림지를 제외한 5가지 범주는 App.1~2 수준으로 보고하고 있다. 또한, 독일, 핀란드, 뉴질랜드 등 주요선진국은 6가지 범주 모두 App.3 수준으로 보고하고 있으며, 범주에 따라서 App.1 - 3까지 적용방식이 다양하였다. 우리나라 NIR의 경우, LULUCF 분야 범주별 정의는 시계열 활동자료가 미비하여 6개의 토지이용 범주 중 정주지 및 기타토지는 산정하지 않고 있으며, 현재 정의하고 있는 산림지, 농경지, 초지, 습지는 세부항목에 대한 정의가 명확하지 않아 주요 선진국가의 사례를 바탕으로 IPCC 기준에 부합하는 정의에 대해 검토하였다. 토지이용 변화 매트릭스 구축의 경우, 주요 선진 국가에서는 Sampling 또는 Wall-to-wall 기법을 이용한 토지이용 변화 매트릭스를 구축하고 있지만, App.1 수준인 우리나라는 국제적인 LULUCF분야 통계로 인정받기 위해서는 국가 통계자료와 공간정보를 활용한 App.2- 3 수준의 토지이용변화 매트릭스를 구축이 필요할 것으로 사료된다. The objectives of this study were to define the land use categories of Korea and present the way to establish the land use change matrix by using the national inventory report(NIR) of major developed countries, according to the IPCC Guidelines, the GHG absorption and emission statistics of the LULUCF sector are classified land use into 6 categories: forest, agricultural, grassland, wetland, jungju, and other land. Major developed countries such as Japan, Germany, Finland, and New Zealand present the definitions of land use categories satisfying the standards of IPCC through the consultation among national agencies. Moreover, the details of land use category are defined differently according to each country's land situation. South Korea does not estimate the statistics of settlements and other land categories because there are not enough active data. Moreover, South Korea needs to add more sub-categories under forest land, cropland, grassland, and wetland, which are currently defined. In the aspect of establishing the land use change matrix, major developed countries have established the approach(App.)3 level matrix based on sampling and wall-to-wall techniques. However, South Korea is at the App. 1 level, based on the national statistics. Therefore, it is required to construct a matrix by using image data such as national thematic maps and forest aerial photographs.

Detection of Individual Tree Species Using Object-Based Classification Method with Unmanned Aerial Vehicle (UAV) Imagery

박정묵,심우담,이정수 강원대학교 산림과학연구소 2019 Journal of Forest Science Vol.35 No.3

This study was performed to construct tree species classification map according to three information types (spectral information, texture information, and spectral and texture information) by altitude (30 m, 60 m, 90 m) using the unmanned aerial vehicle images and the object-based classification method, and to evaluate the concordance rate through field survey data. The object-based, optimal weighted values by altitude were 176 for 30 m images, 111 for 60 m images, and 108 for 90 m images in the case of Scale while 0.4/0.6, 0.5/0.5, in the case of the shape/color and compactness/smoothness respectively regardless of the altitude. The overall accuracy according to the type of information by altitude, the information on spectral and texture information was about 88% in the case of 30 m and the spectral information was about 98% and about 86% in the case of 60 m and 90 m respectively showing the highest rates. The concordance rate with the field survey data per tree species was the highest with about 92% in the case of Pinus densiflora at 30 m, about 100% in the case of Prunus sargentii Rehder tree at 60 m, and about 89% in the case of Robinia pseudoacacia L. at 90 m.

딥러닝모델을 이용한 국가수준 LULUCF 분야 토지이용 범주별 자동화 분류

박정묵,심우담,이정수 대한원격탐사학회 2019 大韓遠隔探査學會誌 Vol.35 No.6

Land use statistics calculation is very informative data as the activity data for calculating exact carbon absorption and emission in post-2020. To effective interpretation by land use category, This study classify automatically image interpretation by land use category applying forest aerial photography (FAP) to deep learning model and calculate national unit statistics. Dataset (DS) applied deep learning is divided into training dataset (training DS) and test dataset (test DS) by extracting image of FAP based national forest resource inventory permanent sample plot location. Training DS give label to image by definition of land use category and learn and verify deep learning model. When verified deep learning model, training accuracy of model is highest at epoch 1,500 with about 89%. As a result of applying the trained deep learning model to test DS, interpretation classification accuracy of image label was about 90%. When the estimating area of classification by category using sampling method and compare to national statistics, consistency also very high, so it judged that it is enough to be used for activity data of national GHG (Greenhouse Gas) inventory report of LULUCF sector in the future. 신기후체제에 대응하여 정확한 탄소흡수 및 배출량을 산정하기 위해 토지이용 범주별 통계량 산출은활동자료로서 매우 중요한 자료이다. 본 연구는 효과적인 토지이용 범주별 판독을 위하여 산림항공사진(이하FAP)에 딥러닝모델을 적용하여 토지이용 범주별 자동화 판독 분류를 한 후 샘플링기법을 통해 국가단위 통계량을 산출하였다. 딥러닝모델에 적용한 데이터세트(이하, DS)는 국가산림자원조사 고정표본점 위치 기반 FAP의 이미지를 추출하여 훈련데이터세트(이하, 훈련DS)와 시험데이터세트(이하, 시험 DS)로 구분하였다. 훈련DS는 토지이용 범주별 정의에 따라 이미지별 레이블을 부여하였으며, 딥러닝모델을 학습하고 검증하였다. 검증 시 모델의 학습정확도는 학습 횟수 1500회에서 정확도가 약 89%로 가장 높았다. 학습된 딥러닝모델을 시험DS에 적용한 결과, 이미지 레이블의 판독 분류정확도는 약 90%로 높았다. 샘플링기법을 통해 범주별 분류결과에 대해 면적을 추정하여 국가통계와 비교한 결과 정합성 또한 높아 향후 LULUCF(Land Use, Land Use Change, Forestry)분야 국가 온실가스 인벤토리 보고서의 활동자료로 활용하기에 충분하다고 판단된다.

고해상도 항공 영상과 딥러닝 알고리즘을 이용한 표본강도에 따른 토지이용 및 토지피복 면적 추정

이용규,심우담,이정수 한국산림과학회 2023 한국산림과학회지 Vol.112 No.3

This research assessed the feasibility of using high-resolution aerial images and deep learning algorithms for estimating the land-use and land-cover areas at the Approach 3 level, as outlined by the Intergovernmental Panel on Climate Change. The results from different sampling densities of high-resolution (51 cm) aerial images were compared with the land-cover map, provided by the Ministry of Environment, and analyzed to estimate the accuracy of the land-use and land-cover areas. Transfer learning was applied to the VGG16 architecture for the deep learning model, and sampling densities of 4 × 4 km, 2 × 4 km, 2 × 2 km, 1 × 2 km, 1 × 1 km, 500 × 500 m, and 250 × 250 m were used for estimating and evaluating the areas. The overall accuracy and kappa coefficient of the deep learning model were 91.1% and 88.8%, respectively. The F-scores, except for the pasture category, were >90% for all categories, indicating superior accuracy of the model. Chi-square tests of the sampling densities showed no significant difference in the area ratios of the land-cover map provided by the Ministry of Environment among all sampling densities except for 4 × 4 km at a significance level of p = 0.1. As the sampling density increased, the standard error and relative efficiency decreased. The relative standard error decreased to ≤15% for all land-cover categories at 1 × 1 km sampling density. These results indicated that a sampling density more detailed than 1 x 1 km is appropriate for estimating land-cover area at the local level.

Evaluation of Suitable REDD+ Sites Based on Multiple-Criteria Decision Analysis (MCDA): A Case Study of Myanmar

박정묵,심우담,이정수 강원대학교 산림과학연구소 2018 Journal of Forest Science Vol.34 No.6

In this study, the deforestation and forest degradation areas have been obtained in Myanmar using a land cover lamp (LCM) and a tree cover map (TCM) to get the CO2 potential reduction and the strength of occurrence was evaluated by using the geostatistical technique. By applying a multiple criteria decision-making method to the regions having high strength of occurrence for the CO2 potential reduction for the deforestation and forest degradation areas, the priority was selected for candidate lands for REDD+ project. The areas of deforestation and forest degradation were 609,690ha and 43,515ha each from 2010 to 2015. By township, Mong Kung had the highest among the area of deforestation with 3,069ha while Thlangtlang had the highest in the area of forest degradation with 9,213 ha. The number of CO2 potential reduction hotspot areas among the deforestation areas was 15, taking up the CO2 potential reduction of 192,000 ton in average, which is 6 times higher than that of all target areas. Especially, the township of Hsipaw inside the Shan region had a CO2 potential reduction of about 772,000 tons, the largest reduction potential among the hotpot areas. There were many CO2 potential reduction hot spot areas among the forest degradation area in the eastern part of the target region and has the CO2 potential reduction of 1,164,000 tons, which was 27 times higher than that of the total area. AHP importance analysis showed that the topographic characteristic was 0.41 (0.40 for height from surface, 0.29 for the slope and 0.31 for the distance from water area) while the geographical characteristic was 0.59 (0.56 for the distance from road, 0.56 for the distance from settlement area and 0.19 for the distance from Capital). Yawunghwe, Kalaw, and Hsi Hseng were selected as the preferred locations for the REDD+ candidate region for the deforestation area while Einme, Tiddim, and Falam were selected as the preferred locations for the forest degradation area.