Probabilistic estimates of the distribution of near-surface (within 1m) permafrost in Alaska.

This medium resolution (i.e. 30-m pixels) spatial dataset provides an estimate of the occurrence of near-surface (within 1 m of the ground surface) permafrost in Alaska. The dataset was developed using a decision-tree model that statistically and spatially extended late-season field observations (n ~ 17,000) using remotely sensed imagery, climatic data, and thematic maps of a wide range of surface and subsurface biophysical characteristics. To circumvent the use of seasonal frost observations, only thaw-depth measurements taken during late-season months (late July to mid-September) or measurements designated to have no near-surface (within 1m) permafrost were used for model calibration and validation. Individual pixel values represent the probability of encountering near-surface permafrost (i.e. 0 to 100 %), as derived from a tree ensemble model, and pixel values greater than 100 correspond to masked land cover types (i.e. 101 = open water; 102 = perennial ice/snow; 103 = developed; 104 = barren, and; 105 = cultivated areas) as defined by the 2001 National Land Cover Database. Decision tree models and resultant maps were tested using independent field observations and f-fold cross validations, which indicated that the map product has an overall accuracy of approximately 85 % when using a probability threshold of 50 % (i.e. less than 50 % = near-surface permafrost absent; greater than or equal to 50 % = near-surface permafrost present).

本数据集为像素尺寸30米的中等分辨率空间数据，用于估算阿拉斯加地区近地表（地表以下1米范围内）多年冻土的分布情况。本数据集基于决策树模型构建，通过遥感影像、气候数据及涵盖多类地表与地下生物物理特征的专题地图，对季末野外观测样本（样本量约17000个）开展统计与空间拓展。为规避季节性冻土观测数据的使用，模型校准与验证仅选用7月下旬至9月中旬季末时段获取的融深测量数据，或是被标注为不存在近地表多年冻土的观测数据。单个像素值代表由树集成模型推导得到的近地表多年冻土出现概率，取值范围为0至100%；像素值大于100的则对应掩膜土地覆盖类型，依据2001年全国土地覆盖数据库（National Land Cover Database）的定义，具体包括：101=开阔水域，102=常年冰/雪，103=建设用地，104=裸地，105=耕地。本研究采用独立野外观测数据与f折交叉验证对决策树模型及生成的分布图开展验证，结果显示：当以50%作为概率阈值时（概率小于50%代表不存在近地表多年冻土，大于等于50%代表存在近地表多年冻土），该数据集的总体精度约为85%。