Extraction of mining disturbance-recovery trajectory types and temporal information in Shengli mining area based on NDVI time series classification

In order to reconstruct the time series information of the mining history of the grassland mining area, taking the Shengli mining area as an example, 32 Landsat TM/OLI images from 1985 to 2017 were applied to compare and analyze the classification accuracy and efficiency of three time series classification models, such as random forest (Random Forests, RF) and the nearest neighbor algorithm based on time dynamic distortion (DTW-kNN), and the residual network model based on convolutional neural network (Residual Network, Resnet). Finally, according to the change trajectory characteristics of the NDVI time series, the disturbance time, disturbance recovery time and disturbance duration of coal mining on vegetation are respectively extracted. The research results show that: ① The Resnet model has the best relative effect in the NDVI time series classification, with a classification accuracy of 90.1% and a Kappa coefficient of 0.898; ② The disturbed and undisturbed areas account for 16.5% and 83.5% of the study area, respectively. The unrestored area of disturbance accounted for 9.5% of the study area; ③ Disturbance mainly occurred during 2005-2008, with the largest disturbance area in 2008; disturbance restoration mainly occurred in 2008 and after, with the largest disturbance recovery area in 2009. The duration of continuous disturbance of the recovery type after disturbance is about 5 years, and the average disturbance duration is 6.7 years.