Abstract We have generated permafrost probability distribution maps (10 m resolution) for the north-eastern Himalayan region in Sikkim using remote sensing measurements and machine learning algorithms. Four machine learning algorithms, logistic regression, random forests, support vector machines and neural networks, and two different sets of input data set, were used to generate a total of 8 machine learning models and hence 8 permafrost probability distribution maps. The first set of input data set included surface reflectance from atmospherically corrected Sentinel-2A spectral bands, elevation and slope parameters while the second set of input data set included mean annul air temperature (MAAT) and potential incoming solar radiation (PISR). Permafrost probability distribution maps obtained from the 8 models show reasonable agreement in the total spatial extent of permafrost occurrence but dissimilarities in the pattern of probability distribution. Accuracy assessment results are more optimistic towards models developed using spectral reflectance, elevation and slope parameters as input data set. Nevertheless, 5 out of 8 models agree that around 60% of total area under observation is highly likely to contain permafrost. This congruence in outputs, despite the use of different machine learning algorithms and separate sets of input data set, establishes reliability in the application of machine learning models for the preliminary estimation of permafrost distribution for remote and data-scarce Himalayan region.

中文翻译：

---

使用逻辑回归、随机森林、支持向量机和神经网络对锡金喜马拉雅山永久冻土发生的空间预测

摘要 我们使用遥感测量和机器学习算法为锡金喜马拉雅东北部地区生成了永久冻土概率分布图（10 m 分辨率）。四种机器学习算法，逻辑回归、随机森林、支持向量机和神经网络，以及两组不同的输入数据集，被用来生成总共 8 个机器学习模型，从而生成 8 个永久冻土概率分布图。第一组输入数据集包括来自大气校正的 Sentinel-2A 光谱带的表面反射率、高程和坡度参数，而第二组输入数据集包括平均年空温度 (MAAT) 和潜在入射太阳辐射 (PISR)。从 8 个模型获得的永久冻土概率分布图在永久冻土发生的总空间范围上显示出合理的一致性，但在概率分布模式上存在差异。精度评估结果更倾向于使用光谱反射率、高程和坡度参数作为输入数据集开发的模型。尽管如此，8 个模型中有 5 个同意观测总面积的 60% 左右很可能包含永久冻土。尽管使用了不同的机器学习算法和不同的输入数据集，但这种输出的一致性在机器学习模型的应用中建立了可靠性，用于初步估计偏远和数据稀缺的喜马拉雅地区的永久冻土分布。