ABSTRACT In this study, we investigated the potential of using decision tree Machine Learning (ML) algorithm and profiles of vegetation and moisture indices extracted from Land Remote-Sensing Satellite (System, Landsat) time series to identify the ages of rubber plantations in Thalang district, Phuket province, southern Thailand. The secondary Land Use and Land Cover (LULC) data and historical imagery from Google EarthTM were used to distinguish plantation boundary and the establishment year of rubber plantation (T 0). The inter-annual profiles of spectral indices for each rubber plantation were obtained from 129 Landsat images (summer period from October 1991 to April 2018). The predictors were generated from summary distribution values of spectral indices during summer, including their difference and ratio at two years before to six years T 0 for Recursive Partitioning (RP) supervised classification algorithm. Modelling dataset from ‘known T 0’ plantations was divided into the training and testing datasets with a 60:40 ratio. The training model was 30 times repeated training while cross-validation assessment was tested to optimize an appropriated hyperparameter based on F 1 score. Then, the best performance training model was applied to both modelling and predicting (‘unknown T 0’ plantations) datasets. The predicted T 0 for each plantation was selected based on results aggregation of 100 times repeated prediction. The results show that the RP model with a complexity parameter as 0.01 and the predictors generated only from Normalized Difference Vegetation Index (NDVI) profile acceptable achieves an accuracy of 84.4% and 84.7% for lowland and highland modelling datasets, respectively. The root means squared error (RMSE) of predicted establishment year were 0.83 years at plantation scale.

中文翻译：

---

使用机器学习算法和 landsat 时间序列识别橡胶种植园的建立年份：以泰国普吉岛他朗区为例

摘要 在本研究中，我们研究了使用决策树机器学习 (ML) 算法以及从陆地遥感卫星 (System, Landsat) 时间序列中提取的植被和水分指数剖面来识别他朗地区橡胶种植园年龄的潜力，普吉府，泰国南部。二级土地利用和土地覆盖 (LULC) 数据和来自 Google EarthTM 的历史图像被用来区分种植园边界和橡胶种植园的建立年份 (T 0)。从129张Landsat图像（1991年10月至2018年4月的夏季）获得每个橡胶种植园的光谱指数年际剖面。预测变量是根据夏季光谱指数的汇总分布值生成的，包括它们在两年前到六年 T 0 的差异和比率，用于递归分区（RP）监督分类算法。来自“已知 T 0”种植园的建模数据集以 60:40 的比例分为训练和测试数据集。训练模型是 30 次重复训练，同时测试交叉验证评估以优化基于 F 1 分数的适当超参数。然后，将最佳性能训练模型应用于建模和预测（“未知 T 0”种植园）数据集。每个种植园的预测 T 0 是根据 100 次重复预测的结果汇总来选择的。结果表明，复杂度参数为0的RP模型。01 和仅从可接受的归一化差异植被指数 (NDVI) 剖面生成的预测变量对于低地和高地建模数据集分别达到 84.4% 和 84.7% 的准确度。在人工林规模下，预测建立年份的均方根误差 (RMSE) 为 0.83 年。