An Efficient Method for Predicting Soil Thickness in Large-scale Area Based on Cluster Sampling

After fast mean shift (FMS) clustering, the whole research area was divided to 10 subareas, so the new samples can characterize the geographical features of each subarea were collected through field investigations. Because of our limited human and material resources, it is difficult to conduct a mass of sampling in each subarea. In order to make the most of our limited resources, we need to conduct reasonable field sampling strategy. For the first two large subareas, we collected 70 field samples respectively, and labeled them as the first sample set and the second sample set that will be used to build their own GWR models for extend prediction of unobserved points in each area, i.e. local extension prediction; while the remaining 8 small subareas took moderate amounts of samples according to their size, if one subarea owns the size of raster points more than 5000, 16 samples will be collected from it, otherwise, take 12 samples. In this way, a total of 112 samples are put together as the third sample set, and the third GWR model is constructed to achieve the global extension prediction of 8 subareas. In addition, three sample sets were divided into training set and test set, respectively. For the first two sample sets, the ratio of sample size of training set and test set are all 5:2, i.e. training set contains 50 samples, test set has 20 samples. Because of the third sample set composed of samples from 8 subareas, we divided the samples of each subarea into training set and test set according to the ratio of 3:1. In the other word, the sample number of training set from third to tenth subarea is 12, 9, 9, 12, 9, 12, 12 and 9 respectively, and 84 training sample in total; and the sample number of test set from eight subarea is 4, 3, 3, 4, 3, 4, 4 and 3 respectively, a total of 28 samples.