0463-B4
A.A. Darvishsefat, M. Abbasi, M.R. Marvie Mohadjer 1
In order to investigate the potential of forest type mapping in beech forests of North Iran with Landsat ETM+, data of a district (780 ha) of educational and reseaerch forest in Northren of Iran dating July 2000, were analyzed. First the quality of the image was evaluated. There were no obvious noticeable geometric and radiometric errors. Geometric registration was implemented using digital elevation model and ground control points. The RMS error was calculated to be much less than half of an ETM+ pixel. In order to estimate the accuracy of the map derived from classification, a ground truth map was prepared using strip sampling in the forest. These strips covered 42% of the total area. Image classification was performed using original and synthetic bands (rationing, principal component analysis and tasseled cap transformation) for following four beech types: pure beech, dominant beech, mixed beech and non-beech types. Classification was performed using maximum likelihood (ML), minimum distance to mean (MD) and parallelepiped (PPD) classifiers. The best band set is selected using the "Bahttacharrya distance" method. The highest accuracy was obtained by using ML classifier (32%). Similar reflection of pure and dominant beech types caused this undesired result. These two types were merged and classification was done again. The highest overall accuracy, considering three classes (merged class and the other classes) increased the classification accuracy up to 46%.
According to the results of this study, the spectral data of ETM+ do not have a large potential for beech type mapping in such a heterogeneous forest with a strong pronounced relief. A suitable radiometric correction needs to be conducted to remove illumination effect, and also advanced analysis methods should be considered. Integration of digital environmental data may provide a means for improving the ability to accurately classification of forest types.
Forest stand maps is a very important and necessary tool to assess and manage the forest ecosystem. Mapping of forest stands in large areas is not easily through fieldwork or by means of arial photo interpretation. In contrast, satellite data with their own characteristics such as being able to cover large areas, their revisit frequency, their constant spatial resolution and finally their possibility of automatic analysis has created a high potential in forest type mapping.
Therefore there are many research efforts in the last years to investigate on the possibility of forest stand mapping in different forested areas by means of multispectral satellite data. White et al. (1996) evaluated the ability of Landsat-TM data to map forest types in California. This investigation points to the ability of TM data also to separate the major species. Bodmer (1993) also investigated forest stands mapping by Landsat-TM data in central Switzerland. The results of this investigation showed that the accuracy of the classification with eight forest types compared with the ground truth is 32%. A reduction to four types increases the accuracy up 60%. The previous investigation (Darvishsefat, 1994) indicated that Landsat-TM and SPOT data did not have very high potential for forest stand mapping in heterogeneous and mixed forests. An overall accuracy of 65.4% was then possible for a forest type classification with the summary class "seedlings/clearing" and other three further classes with varying degrees of mixture.
The beech forests in northern Iran are unique and very important. There is a need in achieving an accurate and update beech stand maps for forestry and environmental purposes. The objective of this investigation is to evaluate the potential of landsat ETM+ data for beech classification in a rugged terrain of Caspian forest.
The study site (780 ha) is located in mountainous forests in northern of Iran (Experimental forest of University of Tehran, (36° 32' N, 51° 39' E). The elevation of the study area ranges from 750 to 1510 meters above sea level. Stand diversity of this forest is noticeable. The main species of this natural and uneven aged forest is Fagus orientalis which composites with Carpinus betulus, Quercus castaneifolia, Fraxinus excelsior, Tilia begonifolia, Alnus subcordata and Acer velutinum.
A Landsat ETM+ image of the study area dated July 18, 2000 were analyzed. The image quality was excellent. There were no obvious noticeable geometric or radiometric distortions.
In order to evaluate the potential of Landsat ETM+ for forest stand mapping an accurate Ground truth must be available. Therefore a Ground truth was prepared through fieldwork and strip sampling. The width of the strips as well as the distance between them was 60 meters. These strips covered 42% of the total area. The forest types within the strips were defined and delineated qualitatively according to the estimation of beech canopy area percentage (BCAP) in the upper story (Gorji Bahri, 2000). The boundaries of the delineated forest types were digitized and used as the ground truth with four types: pure beech (BCAP>90%), dominant beech (BCAP 50-90%), mixed beech (BCAP <50%) and non-beech (BCAP <10%).
Geocoding of the ETM+ image was performed using ground control points method and a digital elevation model. The rectification included geometric correction of relief displacement. An affine transformation and the nearest neighbour resampling method were applied. The geocoded image was checked for reliability in comparison with a digital topographic map.
In order to improve information extraction from satellite image, suitable spectral transformations such as rationing, PCA and Tasseled Cap transformation were performed on the ETM+ data.
According to forest type definition, four beech types were determinated:
pure beech, dominate beech, mixed beech and non-beech. A supervised classification was conducted using three conventional classifiers (ML, MD, PPD), without any knowledge of a-priori possibilities. Three image classifications were investigated with three different combinations of classes. The training areas were taken through fieldwork and ground truth. The best band sets were selected using "Bahttacharrya distance" criterion and the defined training areas. In order to eliminate isolated classified pixels, the results of classifications were filtered by means of mode filter.
To evaluate the potential of the Landsat ETM+ data for forest type mapping, the results of the classifications with four classes were compared pixel by pixel to the ground truth map. The results, summarized in first row of table 1, demonstrate clearly that this kind of satellite data is not suitable for an operational beech classification in this forest area (overall accuracy = 32%).
According to the confusion matrix of the classification, the high spectral similarity of the pure beech and dominant beech classes caused the most misclassification. In contrast with this, the non-beech class could be classified as the best. Therefore, the two mentioned classes were merged together as beech type and classification was performed with three classes. This caused the improvement of the accuracy of the classification at 14.6%. Since the merged class showed a relatively high spectral similarity with mixed beech, the two classes merged further. The classification was done again only with two generalized classes, beech (BCAP>90%) and non-beech (BCAP<10%). The overall accuracy of the resulted map was 65.2% (third row of Tab.1) According to the error matrixes of the classifications, the user and producer accuracies of the classification with four classes were very different. But the merging of the classes resulted in more similar user and producer accuraies. This points to a better quality of the resulted map.
Based upon this, it is concluded that spectral data of ETM+ do not have a high potential for beech type classification in such a heterogeneous forest with a strong pronounced relief. Obviously this conclusion needs further validation, which are currently underway in other rugged areas. Furthermore a suitable radiometric correction needs to be conducted to remove illumination effect, and also advanced analysis methods should be considered. Integration of digital environmental data, may provide a means for improving the ability of satellite data in accurately classification of forest types.
Table 1. Forest type classifications accuracies
Classifier |
Overall accuracy % |
Kappa coefficient % |
Classes |
Producer accuracy % |
User accuracy % | |
four classes |
ML |
32.13 |
9.6 |
pure beech |
48.46 |
26.09 |
dominant beech |
20.07 |
39.96 | ||||
mixed beech |
21.6 |
14.01 | ||||
non-beech |
37.26 |
49.71 | ||||
three classes |
MD |
48.53 |
14.54 |
pure&dominant beech |
55.92 |
67.39 |
mixed beech |
20.86 |
11.45 | ||||
non-beech |
44.48 |
43.57 | ||||
two classes |
PPD |
65.24 |
19.73 |
beech |
74.77 |
74.34 |
non-beech |
44.9 |
45.46 |
Abbasi, M., 2002. Investigation on the possibility of beech type mapping using Landsat ETM+ data. M.S.c Thesis, Faculty of Natural Resources, Department of Forestry, University of Tehran. 116 pp.
Bodmer, H., 1993. Untersuchung zur forstlichen Bestandeskartierung mit Hilfe von Satellitenbildern. Ph.D. Thesis, Department of Forestry, ETH Zurich, 160 pp.
Darvishsefat, A., 1994. Einsatz und Fusion von multisensoralen Satellitendaten zur Erfassung von Waldinveturen. Ph.D. Thesis, Remote Sensing Laboratories, Department of Geography University of Zurich, 147 pp.
Gorji Bahri, Y, 2000. A study of Classification, Typology and Silvicultural Planning in Vaz Research Forest. P.h.D. Thesis, Faculty of Natural Resources, Department of Forestry, University of Tehran, 139 pp.
White, J.D. Kroh, G.C. Pinder, J.E.I. 1996. Forest mapping in Lassen Volcanic National Park, California, using Landsat TM data and a geographic information system, Photogrametric Engineering and Remote Sensing, P: 229-304.
1 Faculty of Natural Resources, University of Tehran, Iran
Tel: ++98 261 222 30 44
Fax: ++98 261 222 77 65