A study of land cover classification accuracy for MOS-1 MESSR high and low gain data.

A study of land cover classification accuracy for MOS-1 MESSR high and low gain data.

Yukio Mukai, yasunori Nakayama
Remote Sensing Technology center of Japan.

Korehiro Maeda
National Space Development Agency of Japan

Abstract
MOS-1 MESSR has a two-gain mode of high and low and is normally operated in low gain mode. This study examined the accuracies in land cover classification for both high and low gain mode data of MOS -1 MESSR An area within a side lap region between two paths is selected as a test site and two MESSR data which observed the test site only one day a part in high and normal gain were collected these two MESSR data covering the test site were registered using ground control points supervised maximum likelihood classifications were performed for both gain data selecting training areas at same cases using the confusion matrices of the training data. Percents correctly classified in the confusion matrix were higher for most of categories in high gain than in normal gain and the average values of the percents correctly classified for all categories were 97% in high gain and 95% in normal gain it van be seen from this study that the land cover classification accuracy is better in high gain than in normal gain..

Introduction
Japanese Marine observation satellite (MOS) was launched in Feb 1987 and it mounts three kinds of sensor MESSR (Multi spectral ) electronic self scanning radiometer VTIR visible and thermal infrared self scanning Radiometer 0m VRTIR scanning Radiometer ) MOS -1 is now successfully collecting the earth observation data and in Feb. 1990 MOS 1b mounting the same sensors as MOS-1 was launched which enabled to increase the chances to observe the earth by MOS stores.

MESSR has two gain modes of high and low and in normally in the low gain mode,. Therefore low gain is called normal gain. This study investigated the land cover classification accuracies of high and normal gain data.

The outlines of MESSR.
There are two systems in MOS-1 MESSR and they are mounted parallel to the direction of movement of the space craft as shown parallel to the direction of movement of the space craft as shown in fig 1 the system mounted in the east and west side is called each system 1 and system 2 the swath of each system is 100 km and their overlaps about 10 km the two systems are operated alternatively in normal operation mode but can be operated simultaneously in special mode MESSR data have four bands and the range of wave length and the spatial resolution of each band are shown in table 1

Fig. 1 Observation by System I and II of MESSR

Fig. 2 Input-Output characteristics of high and normal gain of MESSR

Band	Range of wavelength (mm)	Spatial Resolution (m)
1 2 3 4	0.51~0.59 0.61~0.69 0.72~0.80 0.80~1.10	50 50 50 50

The input out put characteristics of high and normal gain of MESS Rare shown in fig 2 the A/D converter output is from 0 to63 DN (digital number) for the input level from o to 100% in normal gain mode and the conversion characteristics of bands 1,2 are non linear but bands 3,4 are linear in high gain mode the coefficients of A/D conversion increase three times in band 1,2 and twice in band 4 compared to the normal gain mode there is no high gain mode in band 3.

Test site and used data.
There are about 50 km of side lap area between the two paths of MESSR scenes An area of 34 by 28 km in the side lap area between two MESSR scenes 20-69 west and 21-60 east was selected as the satellite in this study as the observation area of MESSR moves by ones path west ward every below which observed the test site in normal and high gain mode one day apart were used for this study.

20-69	west	SYSTEM I	normal gain	Dec 3. 1989
21-69	east	SYSTEM II	high gain	Dec.3 1989

a flat part in the side lap area was selected as the test site in order to avoid the stereoscopy caused by two data seen from two adjacent paths.

Fig 3 MESSR Scenes and test site

Study Method
Study method for the evaluation of land cover classification accuracies for high and normal gain data is shown in fig 4 at first a portion of images covering the test site was extracted each from the original MESSR data they were geometrically corrected using GCPs Ground control points 24 GCPs. Were selected and an affine transformation was used for the geometrical correction and the residual errors were about 0.6 pixel with both images which is considered that the two images register with an error less than 1 pixel training areas for land cover categories were likelihood classification was carried out supervised maximum likelihood respectively for high and normal gain data as the acquisition dates of two data are only on e day different the land cover conditions are considered to be almost same the land cover classification accuracies for high and normal gain data were evaluated by examining the confusion matrices of the training data.

Selection of training and classification of the test site images.
Twelve categories of land cover as shown in table 2 were established in this study the training data which area are sample area representing each category were selected referring geographical maps of the test site mean values of training data of land cover categories are shown in fig 5 it can be seen from that the values of high gain are three times as higher as those of normal gain in band 1.2 and twice as higher as in band 4 and the values of band 3 are same, it shows that the input out put characteristics of high and normal gain shown in fig. 2 are realized correctly A supervised maximum likelihood classification was carried out using the training data selected above for each high and normal gain data the percentage of each category of the classified images through the sets site for high and normal gain data is also shown in table 2.

Table 2 land cover categories selected for the classification and percentage of each category of the classified image through the test site.

No.	Category	Percentage
No.	Category	Normal gain	High gain
1.	City	1.3	0.9
2.	Residential	18.8	16.6
3.	Factory	2.4	0.9
4.	Bare soil	3. 8	8.9
5.	Waste	23.1	2 9.1
6.	Lawn	5 .7	4.1
7.	Farm 1	18. 2	1 4.6
8.	Farm 2	2 .8	5.9
9.	Paddy field	18.0	13.5
10.	Forest	4.8	4.4
11.	Mountain	0.1	0.2
12.	River water	1.0	0.9

Evaluation of land cover classification accuracy for high and normal gain data.

Land cover classification accuracy by confusion matrix

It is at percent a general idea to use the confusion matrix for examiner the accuracy of multi spectral classification the confusion matrix is a table showing percent classified to each category of the training data representing a category table 3 and 4 show the confusion matrices of training data in the classification for normal and high gain data respectively the diagonal elements of the confusion matrix can be measure of the classification accuracy it can be seen from table 3 and 4 that the diagonal elements of high and normal gain are higher than those normal gain and the average classification accuracy of high and normal gain 97% and 95% respectively.

Table 3 confusion of training data in the classification for normal gain data

No.	Category	Percentage Classified to each category
No.	Category	1	2	3	4	5	6	7	8	9	10	11	12
1	City	93.8	4.6										1.5
2	Residential		96.4			2.7		0.9
3	Factory			95.7	4.3
4	Bare soil			2.3	93.2			2.3		2.3
5	Waste							1.7		0.8
6	Lawn				1.4	1.4	92.9		2.9	1.4
7	Farm 2			1.5	1.5	3.0		92.4		1.5
8	Farm 1						6.2		93.8
9	Paddy field				3.5	0.2	0.4	6.2	0.2	89.5
10	Forest					0.4					99.6
11	Mountain shadow											100
12	River water	1.1									0.7	1.8	96.3
Mean of diagonal element's		95.1%

Table 4 Confusion matrix of training data in the classification for high gain data.

No.	Category	Percentage Classified to each category
No.	Category	1	2	3	4	5	6	7	8	9	10	11	12
1	city	95.4	3.1		1.5
2	Residential	0.9	97.3		0.9					0.9
3	Factory			95.7	4.3
4	Bare soil				95.5			4.5
5	Waste					99.2	0.8
6	Lawn						95.7		4.3
7	Farm 2							97.0		3.0
8	Farm 1						3.1		96.9
9	Paddy field				3.9	0.2	0.2	0.4		95.3
10	Forest										100
11	Mountain water											100
12	River water	2.2									0.4		97.4
Mean of diagonal elements		97.7%

These results were obtained using normal gain data observed by system and high gain data observed by SYSTEM II therefore it is strictly speaking necessary to check if the land cover classification accuracies of data observed by the system 1 and 2 in same gain mode are identical.

Evaluation of the classified result of the test site.

The percentage of each category of the classified image through the test site are not necessary the same between the two data there are following differences.
- Artificial structure categories such as city residential and factory are larger in normal gain than in high gain.
- Bare soil and waste are larger in high gain than in normal gain than high gain.
- Paddy field and farm 1,2 are larger in normal gain than in high gain but farm 2 is larger un high gain than the normal gain.
More studies may be necessary in order to clarify the above difference between the two classified results.

Conclusion
An area within a side lap area between the two paths selected as a test site and to MSSR data which observed test site only one day a part in high and normal gain was collected. Land cover classification was carried out using the high and normal hair respectively by a supervised maximum likelihood method. The land cover classification accuracies were examined for the two data the diagonal elements of the confusion matrix of high gain are higher than those of normal gain and the average classification accuracy of high and normal gain are 97 % and 95 % respectively but it is necessary to consider that these results were obtained using normal gain data observed by system 1 and high gain data observed by system 2 the percentage of each category of the classified image through the test site are not necessary the same between the two results.

References

American society of photogrammetry: manual of Remote Sensing 2nd Edition pp. 800.801 1982.