부산대학교 도서관

Hyperspectral reflectance data in the VNIR-SWIR range (400-2500nm) are commonly used to non-destructively and contactless measure the chemical composition of the plants. Most traditional methods are based on non-spatial analysis, that method required the average spectral data to represent the entire image, resulting in the loss of spatial information. To address this issue, we utilize a U-Net network to preserve spatial information while also allowing for the identification and quantification of lactone content in the image. The resulting distribution map provides a clear visualization of lactone content throughout the field or crop, making it easy to identify areas with high or low lactone levels. The pre-processing method includes image registration, outlier removal, spectral smoothing, and normalization. These steps are designed to correct errors and improve the quality of the image and masking. According to the experimental results, the U-Net model achieved R 2 , RMSE, and DICE of 0.718, 11.66, and 89.92%, respectively. The results show that using hyperspectral images combined with the U-Net network can perform a reliable and accurate prediction model for determining lactone content in A. paniculata.