Spatial scale transformation is one of the basic and important scientific problems in quantitative remote sensing field.Spatial up-scaling has particularly drawn much attention

as it can effectively help solve difficult problems

e. g.

validation of quantitative remote sensing products. However

some issues remain concerning spatial up-scaling research.（ 1） The transformation formula established by statistical methods has no explicit physical meaning and its available range is limited.（ 2） The lack of reasonable retrieved physical models hampers the development of up-scaling based on these models. As an important retrieval method

the up-scaling of NDVI also faces these two issues. To address these problems using statistical and physical methods

continuous spatial scaling model（ CSSM） of NDVI on the basis of fractal theory was established. The CSSM exhibits a wide available scale range and partial physical meanings. However

the means of determining the most reasonable Level（ scale hierarchies） for establishing the model remains an important problem

which is studied in this research.In this research

a precise and rigorous method of determining the most reasonable Level was developed based on a five-index estimation system. The system integrates statistical estimation indices（ r

p

rlo

and rup） and an availability-in-validation index [largest error in validation

Max

o

f

a

bs（ Error） ]. It was computed as follows. First

the NDVI CSSM of an image was established on each of the different Levels. Second

the indices（ r

p

rlo

and rup） on each Level were compared and analyzed. Third

the most reasonable Level could be computed based on the defined Max

o

f

a

bs（ Error） to establish the widest scale CSSM.Shatian Byland（ Beihai City

Guangxi Zhuang Autonomous Region） was selected as the experimental area because of its variety of ground objects and high spatial heterogeneity. Taking the values（ r≥0. 8

p

<

0. 05

rlo≥r≤rup and Max

o

f

a

bs（ Error） ≤0. 05） as estimation system

the most reasonable Level（ Level = 267） was computed. On that Level

the model was log2 NDVI =1- 0. 0347log2 1/scale-1.1296 and its scale range was from 30 m to 8010 m. Within the range

validating the NDVI image on any scaleup-scale（ corresponding to the integral multiple of the 30 m resolution of ETM + image） could be implemented by the model.Furthermore

the sensitivity of the Level to values of the estimation system was analyzed. The Level would dynamically change when the threshold values of the five-index estimation system were different and the application purpose changed

which meant that the method in the research was steady and rigorous.In this research

the method of determining the most reasonable Level for establishing the CSSM of NDVI was developed based on a five-index [r

p

rlo

rup

and Max

o

f

a

bs（ Error） ] estimation system. This model quantitatively described the transformation relationships of NDVI on continuous scales. On the basis of this result

NDVI validation of different low-resolution images could be implemented rapidly and effectively. This work results in a more systematic research on modeling the CSSM of NDVI.