초분광 영상을 이용한 콩 들불병 진단을 위한 주요 파장 선정 및 모델 개발 (Development of a Model with Key Wavelengths for Diagnosing Soybean Wildfire Disease Using Hyperspectral Imaging)

서지정보

· 발행기관 : 경상국립대학교 농업생명과학연구원
· 수록지 정보 : 농업생명과학연구 / 57권 / 6호 / 25 ~ 38페이지
· 저자명 : 김은리, 강예성, 류찬석, 박기수, 정종찬, 박진기

초록

이 연구는 초분광 영상으로 두 품종의 콩(청자 3호, 대찬)의 들불병을 진단할 수 있는 모델과 다중분광 영상센서를 개발하기 위해 수행되었다. 무처리구와 들불병 처리구에서 5 nm full width at half maximum (FWHM)으로 구성된 원시 초분광 중심파장들의 콩 식물 영역 반사율들을 추출하여 10 nm FWHM으로 병합한 후, t-test로 차이가 나타난 blue, green, red, red edge, NIR1 및 NIR2 각 영역에서 선정된 대표 밴드로 121개의 식생지수를 계산하였다. 식생지수를 입력변수로 support vector machine (SVM), random forest (RF), extra tree (EXT), extreme gradient boosting (XGB)의 머신러닝 기법과 shapley additive explanation 변수 선택 기법을 적용하여 들불병 진단에 가장 적절한 모델을 선정하고 사용된 식생지수와 파라미터를 나타내었다. T-test 결과 품종에 상관없이 blue 1개(420 nm), green 2개(500, 540 nm), red 1개(600 nm), red edge 2개(680, 700 nm), NIR1 2개(780, 840 nm), NIR2 1개(920 nm)의 총 9개 대표 밴드들이 선택되었고, 성능 평가를 통해 선정된 모델에 청자 3호의 경우 SVM모델(OA=0.86, KC=0.72, 10 VIs)이 선정되었으나 혼동행렬 분석 결과 정상오분류가 적은 RF모델이 선택되었다. RF모델(식 생지수 : RE/Blue, NSI, GDVI, Green/Blue, 파라미터 : max_depth=6, n_estimators=100)은 OA=0.81, KC=0.60, precision=0.86, recall=0.81, F1 score=0.80의 성능을 나타내었다. 대찬은 EXT모델(식생지수 : YVI, RE/Green, 2YVI, 파라미터 : max_depth=8, n_estimators=10)이 선정되 었고, OA=0.86, KC=0.72, precision=0.86, recall=0.86, F1 score=0.86의 성능을 나타내었다.

영어초록

This study was conducted to develop the diagnosis model of wildfire disease for two varieties of soybeans (Cheongja No. 3 and Daechan) based on hyperspectral images and to propose the bandpass filters for a multispectral sensor. The reflectances of the raw hyperspectral center wavelengths, which had a 5 nm full width at half maximum (FWHM), were extracted from the soybean plant canopies and then merged into 10 nm FWHM. To identify differences between the control and infected groups among the merged reflectances (blue, green, red, red edge, NIR1, and NIR2), a t-test was performed with an alpha level of 0.05. Subsequently, the representative bands were identified and applied to calculate 121 vegetation indices. The vegetation indices were utilized as input variables for developing machine learning models, including support vector machine (SVM), random forest (RF), extra tree (EXT), and extreme gradient boosting (XGB). These models were applied to the shaply additive explanation method for feature selection. The most suitable model for wildfiredisease diagnosis was selected by comparing the performance of various models, and the utilized vegetation indices and parameterswere presented in this manuscript. Nine representative bands were selected regardless of the variety: blue (420 nm), green (500 nm,540 nm), red (600 nm), red edge (680 nm, 700 nm), NIR1 (780 nm, 840 nm), and NIR2 (920 nm). As a result of the performanceevaluation, for Cheonja 3-ho 3, Even if the SVM model (OA: 0.86, KC: 0.72, 10 VIs) was better than the RF model, The RF modelwas selected because of the true negative problem on confusion matrix especially for disease diagnosis. The RF model (vegetation indices:RE/Blue, NSI, GDVI, Green/Blue, parameters: max_depth=6, n_estimators=100) exhibited performance with OA=0.81, KC=0.60,Precision=0.86, Recall=0.81 and F1 score=0.80. For Daechan, the EXT model (vegetation indices: YVI, RE/Green, 2YVI, parameters:max_depth=8, n_estimators=10) was selected, and the performance was OA=0.86, KC=0.72, Precision=0.86, Recall=0.86, and F1 score=0.86.

참고자료

· 없음