|
|
|
Poster Session 3
|
Development of land cover classification method using Noaa Avhrr, Landsat TM and DEM images
3. Experiment
The test scene is the northern part of Thailand and almost equivalent to the Landsat TM image size. And data used in this research is geometrically corrected to overlay.
-
NOAA AVHRR NDVI (1992.4-1993.3)
- Landsat TM #58 (1989.1.20)
- Dem (1km)
The land use map produced in 1990 and visually interpreted Landsat TM image were used for picking up training data. The land use map has three classes, forest (evergreen forest and deciduous forest), paddy field, and mixed area of crop field and grassland.
As the first stage, we produced and validated three images, the image classified by multi-temporal analysis and elevation analysis in this paper, the only classified by maximum likelihood method (MLM) using 12-month NOAA AVHRR data, and the one by MLM using Landsat TM data (Figure 6, 7 and Table 1).
Figure 6. Land use map (1990)
The method proposed in this paper
Maximum likelihood method (12-month NOAA AVHRR NDVI)
Maximum likelihood method (Landsat TM band 2, 3 and 4)
Figure 7. Results classified into 3 classes (evergreen forest + deciduous forest, paddy, crop + grassland)
Table 1 : Validated result (numbers mean pixels)
| |
Classified result |
| Land use map |
|
Proposed method |
Maximum likelihood method (NOAA AVHRR) |
Maximum likelihood method (Landsat TM) |
| Paddy 49073 |
Paddy 34723 : 70.8% |
Paddy 35250 : 71.8% |
Paddy 32600: 66.4% |
| Crop etc. 13966 : 28.5% |
Crop etc. 10840 : 22.1% |
Crop etc. 14211: 26.6% |
| Forest 384 : 0.8% |
Forest 2983 : 6.1% |
Forest 2262 : 4.6% |
| Crop + 90433 grassland |
Paddy 18719 : 20.7% |
Paddy 21043 : 23.3% |
Paddy 31943 : 35.3% |
| Crop etc. 59228 : 65.5% |
Crop etc. 49416: 54.6% |
Crop etc. 46878 : 51.8% |
| Forest 12486 : 13.8% |
Forest 19974 : 22.1% |
Forest 11612 : 12.8% |
| Forest 65329 (evergreen + deciduous) |
Paddy 993 : 1.5% |
Paddy 2320 : 3.6% |
Paddy 9956 : 15.2% |
| Crop etc. 10347 : 15.8% |
Crop etc. 10915 : 16.7% |
Crop etc. 22022 : 33.7% |
| Forest 53985 : 15.8% |
Forest 52094 : 79.7% |
Forest 33351 : 51.1% |
| Other 23644 |
|
|
|
Result of texture analysis.
Visually interpreted image (Landsat TM)
(area where edge lines were detected . . . gray )
(area where edge lines were strongly detected . . . black)
Figure 8. Comparison of result of texture analysis with visually interpreted image.
Next, the result by texture analysis was compared with visually interpreted Landsat TM image (Figure 8).
Finally, the image classified into five classes was shown in Figure 9.
Figure 9. Classified result (5 classes)
4. Discussion
Table 1 shows that the result classified by MLM using temporal Landsat TM data tends to overlook forest (evergreen forest and deciduous forest).
forest and deciduous forest). Considering this data were detected in January, that's because the special characteristic of deciduous forest without leaves are similar in the one of crop and grassland. It can be said that there's a limitation of classification method using temporal data alone.
When comparing the result classified by the method in this research and the one classified by MLM using NOAA AVHRR data, we focus on the misclassified result, e.g. forest in land use map classified into paddy field. While the rate of that misclassification is high in the result of classification by the proposed method. It can be said that using the slope of the elevation prevented from misclassification, and it brought better accuracy.
Any seeing Figure 8, texture analysis is useful to discriminate crop field which show linear characteristic, especially sugar cane, from grassland. But, other crop field which show little linear characteristic, like bean, are not clearly identified. That's because in this research, crop is defined as the area affected by human impacts and which show linear characteristic. So, in this sense, it can be said texture analysis by wavelet transform analysis can be used to detect crop field which has linear characteristic.
5. Conclusion
In this paper, we have made a structure of classification which may be applied for large- scale land covers and could show the above.
-
To consider seasonal change in spectral characteristic, we focused on the relationship between the average of NDVI during dry season and the annual difference of NDVI and could use it as a criterion to classify evergreen forest, deciduous forest and paddy field.
- Wavelet transform analysis was done as a method of texture analysis and provide to be useful to discriminate crop field from grassland if crop is defined as the area which show linear characteristic.
- The information of the slope of the elevation could be useful to classify deciduous forest and paddy field.
References
-
Benedetti, R. Rossini, P., and Traddei, R., 1994. Vegetation classification in the Middle Mediterranean area by satellite data. International Jornal of Remote Sensing, 15, 583-596.
- Derek, R. P., and Steven E. F., 1991. Image Texture Processing and Data Integration for Surface Pattern Discrimination. Photogrammetric Engineering & Remote Sensing, 57, 413-420.
- Patrick, A. A., Egide, N., 1991. Comparisons between Spectral Mapping Units Derived from SPOT Image Textureandield Soil Map units. Photogrammetric Engineering & Remote Sensing, 57, 397-405.
- Sali, E., and Wofson, H., 1992. Texture classification in aerial photograph and satellite data. International Journal of Remote Sensing, 13, 3395-3408.
- Skole, D. L., Chomentowski, W. H., Salas, W. H., and Nibre. A. D., 1994, Physical and human dimensions of deforestation in Amazon. BioScience, 44(5), 314-32.
- Shri, Z., and Shibasaki, R., 1996. Wavelet Analysis and Its Application. Jornal of the Japan Society of Photogrammetry and Remote Sensing, 35, No.2, 3 and 4, 48-52, 51-55, 53-58.
|
|
|
|
|
|
|
