CHO-YING HUANGShen, Zih-YuZih-YuShenChen, Tzu-Hsin KarenTzu-Hsin KarenChen2024-06-252024-06-252023-01-01[9781713893646]https://www.scopus.com/record/display.uri?eid=2-s2.0-85191233518&origin=resultslisthttps://scholars.lib.ntu.edu.tw/handle/123456789/719451Tea is a popular drink worldwide and a major cash crop in mountainous agricultural regions in Taiwan. However, due to the rugged terrain, these areas are difficult to manage, and frequent fog, cloud cover, and spectral complexity hinder remote sensing applications. With only 92 observations in 221 satellite images (2019-2021) on average for each pixel in the Greater Ali-Mountain tea plantation region of Southern Taiwan (the study site), a systematic method may be necessary to effectively map tea farms. This study aims to evaluate the feasibility of classifying tea plantations in the study region using satellite imagery with machine learning. We utilized Sentinel-2 surface reflectance images to generate annual, seasonal, and continuous change detection and classification land use and land cover (LULC) maps. Two machine learning methods, Random Forests (RF), and U-net with Resnet-18 backbone, were employed to classify five types of LULC. After validating the results, we found that U-net had higher accuracy than RF with significantly higher efficiency in identifying tea farms compared to RF. Using U-net with the seasonal approach resulted in the highest overall accuracy of 0.949, with the tea farm producer's accuracy and user's accuracy being 0.916 and 0.939, respectively. Our findings suggest that U-net is suitable for identifying tea farms due to its ability to augment training data, to use an encoder-decoder structure, and to incorporate skip-connections, which capture image features more effectively and prevent the loss of critical information. This approach offers significant advantages in image interpretation. Moreover, the method shows promising potential for mapping other mountain evergreen crops, such as fruit and coffee trees.falsecontinuous change detection and classificationdeep learningsatellite imagestea farmtime seriesconference paper2-s2.0-85191233518