무인기 기반 다중분광 영상을 이용한 콩 경태 추정 머신러닝 회귀 모델

초록

This study aims to develop machine learning regression models to estimate the stem diameter of soybean (Glycine max (L.) Merr.) using vegetation indices derived from reflectance values obtained from UAV-based multispectral imagery. The experiment was conducted using the Seonpung soybean cultivar, sown on June 20, 2022, and June 24, 2023, at the Department of Southern Area Crop Science, National Institute of Crop and Food Science, Miryang, Gyeongsangnam-do, Republic of Korea. The soybeans were cultivated in both control and treatment plots. Growth surveys were carried out on August 20 and September 20, 2022, and on August 21 and September 25, 2023, and mulitspectral images was acquired on August 22 and September 21, 2022, and on August 22 and September 20, 2023. From the acquired images, five spectral reflectance bands were extracted to calculate nine vegetation indices. Four regression models—Ridge Regression (RR), LASSO Regression (LR), Random Forest Regression (RFR), and K-Nearest Neighbor Regression (KNR)—were used, and a stepwise variable selection method was applied. The dataset was divided into calibration and validation sets with ratios of 8:2, 7:3, and 6:4, and model performance was evaluated using the coefficient of determination (R2), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE). For the single-year models, the models from 2023 were selected as the best models for both August and September. In the multi-year monthly models, cases in which clustering was observed due to biased environmental conditions (August) and cases in which no clustering was observed despite statistically significant differences (September) were identified. Therefore, although its performance was lower than that of the monthly models, the integrated model using all data was selected as the optimal model because it did not exhibit clustering and had a larger sample size. The integrated model showed calibration results of R² = 0.916, RMSE = 0.683 mm, and MAPE = 5.644%, and validation results of R² = 0.708, RMSE = 1.002 mm, and MAPE = 8.957%.

키워드