GISdevelopment.net ---> AARS ---> ACRS 1992 ---> Poster Session P

Ground target recognition: The operationally of Remote Sensing techniques in the tropics

Mazlan Hashim,Mohd Ibrahim,Smasudin Ahmad,Sarudin Awang
Center for Remote Sensing, Faculty of Surveying
University Teknologu Malaysia, Locked Bag 791
80990 Johor Bahru, Malaysia


Abstract
The successful extraction of information from satellite remotely sensed data are largely influenced by two dominant factors i.e. the atmospheric interferences inherent in the data and the target’s own spectral response behavior. The atmospheric effects can be compensated by using suitable atmospheric model. The target’s spectral signatures on the other hand are unique which enable their recognition. In reality, the target’s recognition from remotely sensed data is not trivial particularly in humid tropic regions.

This paper addresses the manipulation of digital data from the Landsat-5 TM and SPOT-1 multispectral data in order to obtain optimum classification accuracy for some selected land use categories. Summary of targets accuracies on classification of the manipulated and original data sets are tabulated. The results show that manipulation by spatial filtering and principal component analysis on the data sets have improved the classification accuracy of 6 land use categories by as much as 45 percent.

Introduction
Recognition of different land use classes from Remote Sensing data poses significant problem in humid tropic regions. This problem is mainly due to the minimum heterogeneity of the classes which will result in poor identification of classes to which they belong in the classification process. Some prominent factors discussed by Schowengerdt (1983) that cause variability within the classes are atmospheric scattering, topography, sun and view angles, class mixture, and within-class reflectance variability. Several processing techniques attempted by Hashim and Ahmad (1989) proved that best seprabilities among land use classes can be optimized using vegetation indices, statistical filtering and merging of data from different sources.

Paper presented at the 13th Asian Conference on Remote Sensing. 7 – 11 October 1992, Mongolia

This paper will further analyze the technique of statistical filtering of the data and utilize the transformed data together with the original data as input in the principal component analysis (PCA). The PCA method was utilized to reduce the number of images that are needed for classification.

Study Area and Data Acquisition
The study area is centered on the small town of Bedong, Kedah Peninsular Malaysia with approximately 80 square kilometers (10 km X 8 km) of various land cover types. The primary land cover is rubber, oil palm, forest, mangrove and a significant portion of mining area in the western part of Bedong. Water features include rivers and disused mining pools. The North and South Highway passes through the center of the study area with other transportation network that includes railways, two-way roads and tracks.

The data used for the identification of land use and cover types were the 1985 Ministry of Agriculture Land Use Map and 1988 Topographical Maps at a Scale of 1:50000. The satellite data used in the study were the Landsat -5 TM data of 12 February 1991 (WRS 128/56) on bands 1 to 5 and 7, and the SPOT-1 MLA data of 6 August 1988(K/J 266/339). Several land use classes based on the national legend (Wong, 1974) were adopted for the classification process in the study area which are indicated in Table 1. Water features including rivers and mining pools were also included in the legend since they feature prominently in the image.

Data Processing
The satellite data were geometrically corrected before spatial filtering and principal component analysis were carried out.

1 Spatial Filtering
The contribution of spatial filters in the processing, before any classification takes place reduce the internal variation within the land use categories and improve the texture properties of the image data (Cushnie and Atkinson 1985). In this study, a 3 x 3 moving window filter with the returning values of standard deviation, mean and variance of the original image data to its center pixels were employed. The transformation resulted in 3 new data files for each TM bands to give a total of 18 data files. A total of 9 data files were created for SPOT data.

2 Principal Component Analysis
Principal component analysis of remotely sensed image data has been used for various information extraction purposes. One of the major use of PCA is to reduce the number of images or variables that are needed for analysis. In this study 18 data files from the spatial filtering process together with the original 6 TM bands were applied in the PCA. For the SPOT data, the 9 data files created from the spatial filtering process together with the original 3 multispectral bands were used.

Table 1. The Land use Classification Legend

Supervised classification using maximum likelihood classifier were carried out on the first three components of the principal component image for each set of satellite data. These first three components will have a large percentage of the total variance present in the data sets. Classification on the original 6 TM bands and 3 SPOT multispectral bands were also performed so that they can be used as standard for comparison in this study.

Results, Discussions and Conclusions
In Table 2, the PCA Method decreases the classification accuracy of 3 land use categories namely mangrove (8), forest (7F) and mature oil palm (3Om) in comparison with classification on the original TM bans. The decrease in the classification accuracy is 8 percent for forest and mangrove and 3 percent for mature oil palm. Improvement in accuracy values of between 6 to 36 percent has been achieved with PCA Method for 6 land use categories. The largest increase in accuracy is obtained by the class pady (4P) with a value of 36 percent, followed by young rubber, 3Gy (39 percent), water, W (9 percent), mining area. 1X (8 percent), young oil palm, 3Oy (6 percent) and 5 percent of mature rubber (3Gm).

* Additional land use chosen for this study

With the SPOT data, (Table 3), the accuracy of 6 land use categories were improved by as much as 45 percent if the PCA method were used. The newly cleared area (7C) has improved accuracy of 45 percent, young rubber (42 percent) scrub forest, 7S (30 percent), mature rubber (25 percent), water (19 percent) and mature oil palm (6 percent). However, the other 3 land use categories have accuracy lowered by 1 to 19 percent i.e. mining area (1 percent), forest (11 percent) and mangrove (19 percent).

The effect of spatial filtering both TM and SPOT data has increased the overall accuracy of most the land use classes (targets) being tested. However, not all targets can be easily recognized, which has resulted in lower classification accuracy.

References

Cushnie, J.L. and Atkinson, P. (1985), Effect of spatial Filtering on Scene Noise and Boundary Detail in Thematic Mapper Imagery. Photogrammetric engineering and Remote Sensing. Vol 51 No. 9 PP 1483 – 1493.

Hashim, M. and Ahmad, S. (1989) Optimizing the Target Seperatability of Covertypes Inventorying of Humid Tropical Region. Proceedings of Global Natural Resource Monitoring and Assessments: Preparing for the 21st Century, Venice, Italy. PP 1307 – 1316.

Schowengerdt, R.A. (1983), Techniques for Image Processing and Classification in Remote Sensing, New York: Academic Press, 1983.

Wong, I.F.T. (1974), The Present Land Use of Peninsular Malaysia. Ministry of Agriculture, Kuala Lumpur.

Table 2. Classification Results Using the PCA Technique (in bold) and the Original TM bands
Theme No of Pixels % correct Commission Error
3Gy 3Gm 8 W 1X 7F 3Om 3Oy 4P
3Gy 517 98
69
509
458
0
0
0
0
0
0
0
53
0
0
0
0
0
0
0
0
3Gm 3668 98
93
0
0
3610
3412
0
0
0
0
0
0
0
0
0
0
0
0
0
0
8 1375 79
87
0
0
0
3
1081
1200
0
0
0
0
79
0
56
43
0
0
0
0
W 751 94
85
0
0
0
0
0
0
706
640
0
0
1
0
0
0
0
0
0
0
1X 544 67
59
0
0
0
0
0
0
0
0
363
325
0
0
0
0
0
0
0
0
7F 3021 82
90
0
0
0
0
233
0
15
0
0
0
2466
21727
5
0
0
0
0
0
3Om 4406 88
91
0
0
95
84
1
1
0
0
0
0
0
0
3875
4037
77
77
0
0
3Oy 1324 61
55
0
0
270
267
0
0
0
0
0
8
0
0
85
62
810
734
0
0
4P 1458 81
45
109
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1188
667

Table 3 Classification Results using the PCA Technique (in bold) and the Original SPOT bands.
Theme No of Pixels % correct Commission Error
3Gy 3Gm 8 W 1X 7F 3Om 7C 7S
3Gy 1615 72
30
1168
488
0
6
14
39
0
0
25
0
0
0
44
98
206
342
10
93
3Gm 5525 87
62
3
0
4797
3400
31
105
0
0
0
0
0
0
357
381
0
0
15
0
8 973 49
68
41
56
68
91
479
659
0
0
0
0
79
0
2
2
127
0
41
5
W 972 91
72
64
238
0
0
0
0
888
704
0
0
1
0
0
0
0
0
0
0
1X 782 83
84
53
21
0
0
0
0
0
0
652
654
0
0
0
0
5
7
024
7F 1457 48
59
29
8
261
294
313
106
0
0
7
0
2466
2727
16
43
53
0
28
0
3Om 1796 88
82
21
20
147
170
2
85
0
0
0
0
0
0
1584
1479
020 21
0
7C 1858 74
29
460
1273
0
0
0
0
0
0
9
4
0
0
0
0
1359
2539
0
0
7S 1285 75
45
92
36
5
84
33
4
0
0
0
0
0
0
33
0
112
316
961
573