Extended Data Fig. 10: Field validation of the near-surface permafrost map.

a, Probability of near-surface (≤1 m depth) permafrost estimated by Pastick et al.³². b, Zoom-in to our area of field observations, in which we collected n = 176 permafrost probe measurements in July 2018 (n = 137), June 2022 (n = 2) and October 2022 (n = 37). Blue dots indicate permafrost detected and red dots indicate no permafrost detected. c, A comparison between our permafrost ground-truth observations (b) and the permafrost probability estimates from Pastick et al.³² (an Alaska-wide permafrost map, calibrated using n = 16,786 statewide observations of near-surface permafrost, but no observations in the region shown in b). d, We explore the accuracy of the Pastick et al.³² permafrost map for the Koyukuk region based on applying a simple classification threshold (that is, classifying all pixels with a reported permafrost probability below the threshold as not permafrost and all pixels with a reported permafrost probability above the threshold as permafrost). We sweep through all possible threshold values, from 0% to 100%, and compute the true positive and true negative rates, as well as the total accuracy. e, The threshold value of 40% yields the highest total classification accuracy. The true-negative, false-negative, false-positive and true-positive values for this classification are shown in the confusion matrix in e. The satellite imagery in b is from Bing Maps Aerial, reprinted with permission from Microsoft Corporation.