Еmpirical method for estimation of the optimum size of random point samples for assessment areas of land cover from space images

Abstract

A promising fast method for estimating land cover areas from satellite imagery is the use of random point sampling. This method allows you to obtain area values without spatially continuous mapping of land areas. The accuracy of the area estimate by this method depends on the sample size. The presented work describes a method for empirically finding the optimal sample size. To use this method, you must select a key site for which a reference land cover exists. For the key site, we perform multiple generation of samples of different sizes. Further, using these samples, we estimate the area of land cover. Comparison of the obtained areas with the reference areas allows you to calculate the measurement error. Analysis of the mean and the range of errors for different sample sizes allows us to identify the moment when the error ceases to decrease significantly with an increase in the sample size. This sample size is optimal. We tested the proposed method on the example of the Kalach Upland. The size range from 100 to 3000 sampling points per key site is analyzed (the size of the sampling in the row increases by 100 points). For each element of this row, we created 1000 samples of the corresponding size. We then analyzed the effect of sample size on the overall relative error in area estimates. The analysis showed that for the investigated key site the optimal sample size is 1000 points (1.1 points/km2). With this sample size, the overall relative error in determining areas was 4.0 % on average, and the maximum error was 9.9 %. Similar accuracy should be at the same sample size for other uplands in the foreststeppe and steppe zones of the East European plain.