논 토양 성분 예측모델 향상을 위한 전처리 기법 연구 = A study on preprocessing techniques to improve prediction models for paddy soil properties

다국어 초록 (Multilingual Abstract)

This study analyzed a regression model predicting soil properties in salt-paddy fields using spectroscopic analysis equipment. Soil samples were collected from four fields in the salt-paddy fields of Jangan-myeon, Hwaseong-si, Gyeonggi-do, and analyzed at a soil verification center. The spectral data of soil samples pre-dried and post-dried were measured for the same soil samples. The predictive regression model utilized Partial Least Squares Regression (PLSR), and various preprocessing techniques were applied. The preprocessing techniques used in this study were Savitzky-Golay smoothing (SG smoothing) and Standard Normal Variate (SNV), and the results of PLSR analysis were compared. The results are as follows.

1. In this study, soil samples were collected from four fields in salt-paddy areas for analysis of nine soil properties: pH, Electric Conductivity (EC), Extractable Calcium (Ca+2), Extractable Magnesium (Mg+2), Soil Organic Matter (SOM), Total Nitrogen (TN), Total Organic Carbon (TOC), and Soil Texture (Clay, Silt). EC showed a difference of less than 20% compared to the overall average, while the other soil properties, excluding EC, exhibited differences of less than 15% compared to the overall average.

2. The soil properties prediction regression model, applying the SNV preprocessing technique, demonstrated Good model results for Silt and Clay in post-dried soil samples, with respective RPD values of 2.25 and 2.21. The application of the SNV preprocessing technique resulted in an overall increase in RPD values for all soil properties, ranging from approximately 5% to a maximum of about 104% improvement.

3. The soil properties prediction regression model, applying the SNV preprocessing technique, demonstrated Good model results for Silt and Clay in post-dried soil samples, with respective RPD values of 2.25 and 2.21. The application of the SNV preprocessing technique resulted in an overall increase in RPD values for all soil properties, ranging from approximately 5% to a maximum of about 104% improvement.

4. The validation of the regression models was conducted on post-dried soil samples, which exhibited superior models compared to pre-dried soil samples. Validation was performed for the regression models generated from two preprocessing techniques (SG smoothing and SNV) applied to the post-dried soil samples. For validation purposes, spectral and soil property data from 120 soil samples analyzed in the previous year were utilized. The verification of the SG smoothing regression model showed that pH, Ca+2, and TN received the best evaluation in the VIS band area. For the SNV regression model, EC and SOM in the VIS-NIR band area, Mg+2 in the VIS band area, and pH, Ca+2, TN, and TOC in the NIR band area exhibited the best statistical values. The prediction regression model with SNV showed overall higher statistical values compared to SG smoothing, with TN [VIS-NIR] showing up to approximately 3.6 times higher predicted values. Considering the development of soil property prediction regression models through PLSR analysis, using SNV preprocessing appeared more suitable than SG smoothing preprocessing.

5. During the development of the predictive regression model for soil properties, post-dried soil samples showed statistically higher accuracy in the regression model compared to pre-dried soil samples. The preprocessing technique indicated that the regression model with the application of SNV achieved higher accuracy than the one with SG smoothing.

6. This study is considered more efficient in terms of time, labor, and economy compared to traditional chemical soil analysis methods. However, the insufficient predictive model results suggest that the accuracy of the predictive model could be improved by collecting diverse soil condition data and further researching various preprocessing techniques.

목차 (Table of Contents)

• Abstract Ⅴ
• 약어 Ⅷ
• Ⅰ. 서론 1
• Ⅱ. 연구사 4
• Ⅲ. 재료 및 방법 8
• 3.1 토양 시료 채취 및 성분분석 8
• 3.2 토양 스펙트럼 및 전처리 11
• 3.2.1 토양 스펙트럼 수집 11
• 3.2.2 토양 스펙트럼 전처리 13
• 3.3 분석 절차 17
• 3.3.1 교차 검증 17
• 3.3.2 부분 최소 제곱 회귀 18
• 3.3.3 회귀 모형 평가 20
• 3.3.4 피어슨 상관분석 22
• Ⅳ. 결과 및 고찰 23
• 4.1 토양 성분 데이터 23
• 4.2 피어슨 상관관계 분석 25
• 4.3 부분 최소 제곱 회귀 26
• 4.3.1 최적의 전처리 토양 성분 예측 회귀모형 개발 26
• 가. 건조 전 Savitzky-Golay smoothing 부분 최소 제곱 회귀모형 26
• 나. 건조 전 Standard Normal Variate 부분 최소 제곱 회귀모형 34
• 4.3.2 전처리에 따른 토양 성분 예측 회귀모형 검증 및 평가 42
• 4.3.3 전처리에 따른 토양 성분 예측 회귀모형 비교 47
• Ⅴ. 결론 및 요약 49
• 참고문헌 52