Scale is a key concept forunderstanding the complexity ofearth system. And it is regarded asone of themain challenges ofearth observation. It is crucial to select the optmi al spatial resolution ofremote sensing mi age according to its application field and its characteristics. Based on analyzing the scale characteristic of remote sensing mi ages, this paper analyses the scale selection and discusses geo-statistics based method of quantificationally selecting the optmi al spatial resolution of remote sensing mi age. Especially, this paper analyses traditional local variancemethod and its defects. As for local variancemethod, it is suggested tomeasure the relationship between the size of the objects in the scene and spatial resolution, and then calculate the mean value of the standard deviation by passing a n pixel by n pixelmoving window for each pixel on successively spatially degraded mi ages, and then takes the mean of all local variances of the successively spatially degraded mi ages as an indication of the spatial elementswithin the scene of the mi age, according to which the optmi al spatial resolution whose mean local variance ismaxmi um can be estmi ated. So the traditional local variancemethod computes the mean of all local variance on the different ground area, which results in that the local variance does not fall substantiallywith the successively degradation of spatial resolution of the mi age, consequently the computational results are non-comparable, and the traditionalmethod can notachieve satisfactory result. Breaking through the lmi itation, this paper proposes the modified local variance method based on variable window sizes and variable resolution to quantitatively select the optmi al spatial resolution of remote sensing mi ages, which are high spatial resolution\nmi ageswith largewindow size and low spatial resolution mi ageswith smallwindow size, so that the relevantground area is kept consistent. The actual process inevitably involves the ideal decmi alwindow size, which is proposed to be computed based on the spatial statistics theory. Consequently, the optmi al spatial resolution of remote sensing mi age can be computed by comparing themodifiedmean local variance. This paper takes three pieces of IKONOS mi ageswhich stand forbuilding district, farmland and forest individually as prmi ary expermi ental miage and themodified local variances are computed for the three kindsof landscape individually. The expermi ental results show that thisgeo-statisticsbasedmethod\nof quantificationally selecting the optmi al spatial resolution of remote sensing mi age has theoretical and instructional meaning: the spatial resolution of3—5m, 3—5m and 1—5m is respectively suitable for landscape of building district,farmland and forest; only the local variance based on variablewindow size and variable resolution can indicate the actual change of local variance with the degradation of spatial resolution of the mi age; local variance method adopts proper window size to reflect the change of landscape property, so it can reflect themicro-characters and is suitable for study on the fine scale landscape or the artificial landscape.