AREA ACCURACY ASSESSMENT FOR
CHINESE LAND-USE DYNAMIC DETECTION OF REMOTE SENSING PROGRAM
SUN Xiao-xia* ZHANG Ji-xian^" LI Hai-tao?
“Chinese Academy of Surveying and Mapping, Haidian District: Beijing 100039, sun.xiaoxia@163.com,lhtao@163.net
? Lab. Of Land Use, The Ministry of Land Resources, Haidian District Beijing, 100029, stecsm@public.bta.net.cn
KEY WORDS: Remote sensing, Land Use, Change Detection, Sampling, Accuracy , Monitoring
ABSTRACT:
A suitable area accuracy assessment scheme for change detection is presented. The technique consists of two parts which are the
necessary stages of accuracy assessment:(l)sampling method and (2)accuracy assessment. A area stratified random sampling is
applied based on the fact that area errors is related to a certain area range (called stratum in following sampling). The approach
involves subdividing the area range into strata,and within each stratum a spectific number of sample polygons are randomly chosen.
Three experiments about polygons of remote sensing monitoring have been done which aim at finding out the distribution law of
area error,the conclusions that the error of polygon area follows normal distribution and the mean is zero are drawed.This give
accuracy assessment of remote sensing-derived change detection a science foundation. Based on relative error theory and error
spread law ,three area accuracy indices including relative root mean square,relative error,and average relative error are built.Using
these indices,the formulas of single polygon error and the whole monitoring region error are produced. Experimental results and
practical work show that the method in this paper is simple, effective and applicable.
1. INTRODUCTION
frequence
0
1.1 General Instructions 0
El «10nu
ww : i a Bl 10- 20nv
The Ministry of Land and Resources P.R.China is conducting a 0 EU n
continuous Land use Dynamic Detection project using 0 050m
Thematic Mapper(TM) 30-meter and SPOT(10-meter or 5 -
: pn at nes ut 40 -25 -10 5 20 35 850 65
meter)data.The objective of this large project is providing the Lt
C : ze : S : relative error %
government a generalized, consistent, and reliable land use Y : à s
change data every year. The accuracy of product is very Figurel. The histogram of relative error.
important not only for users but also for makers, and how to
assess the accuracy of remote sensing-derived change detection
is the focus problem currently.
The rest of the paper is orgnized as follows. We carry out the
experiments about area error of polygons of remote sensing
monitoring in section 2.In section 3, We present area stratified
random sampling technique and an approach to calculate
sample size. In section 4, the area accuracy indices are built and
an example of area accuracy assessment is given in section 5.In n
section 6, We conclude with discussions. T- X—H, (1)
"Sdn
2.2 Mean Test
In order to test whether the mean of relative error is zero, or
whether the error is systematic, we make a hypothesis testing
on monitoring polygons from a variety of images. We use
normal population test technique and construct statistic:
2. EXPERIMENTS ON ERROR DISTRIBUTION
Three experiments are made as follows, aiming at finding out Original — hypothesis H p=p,=0 , alternative
0
the distribution law of area error. hypothesis H, : t 0, rejective range: IT| 2t 0-0, a
2.1 Histogram Experiment 0.01. The result is shown in Table 2.
196 polygons from the monitoring results using radarsat image image n u S ITI i {n=}
of Beiling region are sampled. We divided the sample polygons
into four stratums according to their areas there are 75 ones
: Ir SPOT(10m) 25 0.03683 15.9% 11553 2.7969
under 10mu , 43 ones within 10~20 mu, 50 ones within 20~50 IRS(S.8m) 26 0.03508 1 115% | 15552 27874
mu, 28 ones above 50 mu. In each stratum , relative error is IKONOS(1m) 40. | 001206 | 46% | 16444 | 2.7079
calculated. The histogram of relative error is shown in Figure |. SAR(6.25m) 31 0.04300 | 163% | 1.4688 2.7500
Clearly, Figl displays normal distribution law of area errors in
every area range. Table2. The result of mean test.
J
!
* | acre = 6.075 mu.
456
Intern
We c
thresh
That i
23 I
We mn
error «
First,
relati
Secon
calcul
probal
calcul
msn
——————
The 1
2
À 0.05
accept
normal
The ar
detecti
errors |
stratun
samplii
range i
sample
TM(30
whole
20-50
and rea
