| GISdevelopment.net ---> Technology ---> Image Processing |
Estimation of forest parameters through Fuzzy classification of TM Data
Maisam Toosi
M.sc. Student, Dept. of Remote Sensing Eng.
Email: maisam_toosi@yahoo.com
M. J. Valadan Zouj
Assistant Professor, Dept. of Remote Sensing Eng.
Email: valadan_zouj@kntu.ac.ir
Address: faculty of Geodesy and Geomatic Eng., K. N. TOOSI University of Technology
Vali Asr St., Vanak Sq., Tehran, Iran,
Post Box: 15875-15433
Fax: ++98 21 878 6213, Tel: ++ 98 21 877 9473-5
Abstract
Several studies have investigated the utility of Landsat 5 TM imagery to estimate forest parameters such as stand composition and density. Regression equations have generally been used to relate these parameters to the radiance responses of the TM channels. Such method is not feasible in highly complex landscapes, where forest mixtures and terrain irregularities may obscure the existence of simple relationships. In the current paper a fuzzy approach to the problem is presented. First, some typical forest plots with known features are spectrally identified. A maximum likelihood fuzzy classification is apply to the study image, so as to derive fuzzy membership grades for all pixels with respect to the typical plots. Finally, these grades are used to compute the estimates of the forest parameters by a weighted average strategy.
One TM scene and accurate ground references taken in summer 1987 was utilitilized for the testing. For comparison results first, with PCI Geomatica software (8.1 ver) with supervise classification with maximum likelihood approach were used. Also this image has investigated by fuzzy approach for the testing. The first results statistically are quite encouraging.
Introduction
In the last few years the high resolution multispectral data acquired by the landsat 5 thematic mapper (TM) have been used increasingly in forest monitoring and mapping. Several forest inventors have been obtained for the identification of forest dominant species even in irregular landscapes. On the other hand, the estimation of forest parameters such as actual stand composition and density has been successful only when there was not a complex combination of these parameters and other environmental factors, such as terrain irregularities. The usual estimation methods are based on maximum likelihood with PCI Geomatica software have been set up between forest parameters measured on the ground and the relevant spectral responses in the individual TM channels or combinations of these. The rationale for these analyses is that the variations in forest parameters affect the spectral behavior in a way which can be easily modeled by simple relationships. Unfortunately, this assumption does not always hold, especially when other factors affect the spectral behaviors of the forest examined. In effect, if the signatures of the forests are disturbed by mixtures in composition, terrain irregularities, under story influences, act. The existence of simple relationships between stand parameters and spectral responses can be seriously hampered, or even completely obscured.
It can be note, however, that even if simple relationships do not exist some spectral information is likely to be present in the remotely sensed scenes about the parameters examined.
Actually, the high resolution TM multispectral data acquired at different times of the year contain a great deal of information about the condition of vegetation. The extraction and processing of this information therefore becomes the fundamental issue.
A new methodology is proposed here for the estimation of forest parameters in highly complex landscapes based on a “fuzzy” classification approach. The theory of fuzzy sets is intrinsically suited to dealing with mixed, spectrally undefined cases, and can be adapted to the specific problem as will be shown below. The comparison between results of PCI system and fuzzy method and also by study on TM image, acceptable result can be derived which if can be used in future researches.
2: Study Area, Ground and Satellite Data
The study area, of about 15.3×15.3 km in size, is located at approximately 36ْ 32′ north latitude, 51 ْ 07′ East longitudes. This is part of the mazandaran forest on Alborz mountain chain in the north of IRAN.
There are several maps of this area, like Iranian geography organization map (scale: 1/50000) and national cartographic center map (scale: 1/25000).
The TM frame fore the present research was taken on the 28 of July, 1987. Also PCI Geomatica (8.1) was used for analysis that was powerful software.
Study Area
Figure 1: color composite of TM Image
3: Methodology:
A-Data processing with PCI software:
The processing of the ground and remotely sensed data was carried out on a P4 computer system. The software consists is PCI Geomatica (8.1) for general processing.
A scene of 512 ×512 pixels was extracted from frame and georeferenced by a polynomial interpolation algorithm trained on ground control points. In this process supervise method classification was implemented.
Figure 2: PCI Analysis
B-Fuzzy Estimation:
The supervised classification method for estimating environmental parameters from spectral images is based on a fuzzy strategy already applied to marine investigations. A fuzzy set is characterized by a membership function that associates with each element areas number between 0 and 1 which represents the “fuzzy membership grade”. In this way, each element of the set may have partial and multiple membership and mixed, uncertain situations can be properly handled. When applied to the supervised classifications of remotely sensed data, the fuzzy membership grades can be seen as a measure of the extent to which a pixel belongs to all the cover categories. The membership grades have been also demonstrated to be related to the cover proportions of the categories within the pixels. These grades can there fore be used to weight the measured parameters of the known classes by a strategy similar to that adopted for the statistical estimation of Geological parameters. Once the spectral signatures for each reference class have been defined, the membership grade of each pixel with respect to class i (Fi) can be found as:
With:
Pi: maximum likelihood probability of attribution to the class considered.
n: Number of measurement variables
Ci: Variance – Covariance matrix of the class considered.
Mi: Mean vector of the class considered
Xi: Pixel vector
Pri: Prior probability of the class considered defined from the frequency histograms of the training sets.
In practice, all the reference categories (plots) can be considered as individuals having representative environmental and spectral features. In this way the spectral signatures are found assigning full membership functions to all the reference pixel of the categories, which corresponds to computing the fuzzy mean vectors and the fuzzy variance – covariance matrices as “hard” ones. Since known environmental parameters are associated to all the reference categories, the same parameters can be estimated for unknown pixels by using the membership grades as weighting factors.
Thus:
Where:
Ve: Estimated value for the pixel.
Vmi: Measured value of class i.
Clearly, the method is only valid if the entire range of variability of the study parameters is covered by a sufficient number of representative known categories. As was seen above, such an assumption can be considered realistic in the current case.
Figure 3: Fuzzy Analysis
4: Results:
The results of fuzzy classification and PCI classification show that the algorithm can extremely enhance the classification.
With comparison among RMSE of PCI and fuzzy classification, we can be seen that RMSE decreases from 3.84 m2/ha in PCI classification to 1.69 m2/ha in fuzzy classification, this show a promote in classification using fuzzy estimation.
5: Conclusion:
The proposed fuzzy methodology showed great promise for the estimation of forest parameters. However the procedure was only tested on a single data set. Further testing in different areas and with other imagery data would be useful to validate the results obtained. In general, the method should benefit from a higher density of references plots, even if this could involve an increasing burden of computational complexity.
From a methodological viewpoint, this approach should be most effective in reducing biases in the estimated parameters. Also, the use of fuzzy classification permits considerable flexibility in the estimation procedures. For probabilistically inserted into the process.
Theoretically, the fuzzy approach to estimation could be applied to an enormous series of problems of environmental monitoring, such as agricultural, oceanographic and geological investigations. In all these cases, the method can be expected to produce optimum performance thanks to its probabilistic, flexible nature. Research is currently directed towards the exploration of these possibilities.
6: Refrences:
Several studies have investigated the utility of Landsat 5 TM imagery to estimate forest parameters such as stand composition and density. Regression equations have generally been used to relate these parameters to the radiance responses of the TM channels. Such method is not feasible in highly complex landscapes, where forest mixtures and terrain irregularities may obscure the existence of simple relationships. In the current paper a fuzzy approach to the problem is presented. First, some typical forest plots with known features are spectrally identified. A maximum likelihood fuzzy classification is apply to the study image, so as to derive fuzzy membership grades for all pixels with respect to the typical plots. Finally, these grades are used to compute the estimates of the forest parameters by a weighted average strategy.
One TM scene and accurate ground references taken in summer 1987 was utilitilized for the testing. For comparison results first, with PCI Geomatica software (8.1 ver) with supervise classification with maximum likelihood approach were used. Also this image has investigated by fuzzy approach for the testing. The first results statistically are quite encouraging.
Introduction
In the last few years the high resolution multispectral data acquired by the landsat 5 thematic mapper (TM) have been used increasingly in forest monitoring and mapping. Several forest inventors have been obtained for the identification of forest dominant species even in irregular landscapes. On the other hand, the estimation of forest parameters such as actual stand composition and density has been successful only when there was not a complex combination of these parameters and other environmental factors, such as terrain irregularities. The usual estimation methods are based on maximum likelihood with PCI Geomatica software have been set up between forest parameters measured on the ground and the relevant spectral responses in the individual TM channels or combinations of these. The rationale for these analyses is that the variations in forest parameters affect the spectral behavior in a way which can be easily modeled by simple relationships. Unfortunately, this assumption does not always hold, especially when other factors affect the spectral behaviors of the forest examined. In effect, if the signatures of the forests are disturbed by mixtures in composition, terrain irregularities, under story influences, act. The existence of simple relationships between stand parameters and spectral responses can be seriously hampered, or even completely obscured.
It can be note, however, that even if simple relationships do not exist some spectral information is likely to be present in the remotely sensed scenes about the parameters examined.
Actually, the high resolution TM multispectral data acquired at different times of the year contain a great deal of information about the condition of vegetation. The extraction and processing of this information therefore becomes the fundamental issue.
A new methodology is proposed here for the estimation of forest parameters in highly complex landscapes based on a “fuzzy” classification approach. The theory of fuzzy sets is intrinsically suited to dealing with mixed, spectrally undefined cases, and can be adapted to the specific problem as will be shown below. The comparison between results of PCI system and fuzzy method and also by study on TM image, acceptable result can be derived which if can be used in future researches.
2: Study Area, Ground and Satellite Data
The study area, of about 15.3×15.3 km in size, is located at approximately 36ْ 32′ north latitude, 51 ْ 07′ East longitudes. This is part of the mazandaran forest on Alborz mountain chain in the north of IRAN.
There are several maps of this area, like Iranian geography organization map (scale: 1/50000) and national cartographic center map (scale: 1/25000).
The TM frame fore the present research was taken on the 28 of July, 1987. Also PCI Geomatica (8.1) was used for analysis that was powerful software.
Study Area
Figure 1: color composite of TM Image
3: Methodology:
A-Data processing with PCI software:
The processing of the ground and remotely sensed data was carried out on a P4 computer system. The software consists is PCI Geomatica (8.1) for general processing.
A scene of 512 ×512 pixels was extracted from frame and georeferenced by a polynomial interpolation algorithm trained on ground control points. In this process supervise method classification was implemented.
Figure 2: PCI Analysis
B-Fuzzy Estimation:
The supervised classification method for estimating environmental parameters from spectral images is based on a fuzzy strategy already applied to marine investigations. A fuzzy set is characterized by a membership function that associates with each element areas number between 0 and 1 which represents the “fuzzy membership grade”. In this way, each element of the set may have partial and multiple membership and mixed, uncertain situations can be properly handled. When applied to the supervised classifications of remotely sensed data, the fuzzy membership grades can be seen as a measure of the extent to which a pixel belongs to all the cover categories. The membership grades have been also demonstrated to be related to the cover proportions of the categories within the pixels. These grades can there fore be used to weight the measured parameters of the known classes by a strategy similar to that adopted for the statistical estimation of Geological parameters. Once the spectral signatures for each reference class have been defined, the membership grade of each pixel with respect to class i (Fi) can be found as:
With:
Pi: maximum likelihood probability of attribution to the class considered.
n: Number of measurement variables
Ci: Variance – Covariance matrix of the class considered.
Mi: Mean vector of the class considered
Xi: Pixel vector
Pri: Prior probability of the class considered defined from the frequency histograms of the training sets.
In practice, all the reference categories (plots) can be considered as individuals having representative environmental and spectral features. In this way the spectral signatures are found assigning full membership functions to all the reference pixel of the categories, which corresponds to computing the fuzzy mean vectors and the fuzzy variance – covariance matrices as “hard” ones. Since known environmental parameters are associated to all the reference categories, the same parameters can be estimated for unknown pixels by using the membership grades as weighting factors.
Thus:
Where:
Ve: Estimated value for the pixel.
Vmi: Measured value of class i.
Clearly, the method is only valid if the entire range of variability of the study parameters is covered by a sufficient number of representative known categories. As was seen above, such an assumption can be considered realistic in the current case.
Figure 3: Fuzzy Analysis
4: Results:
The results of fuzzy classification and PCI classification show that the algorithm can extremely enhance the classification.
With comparison among RMSE of PCI and fuzzy classification, we can be seen that RMSE decreases from 3.84 m2/ha in PCI classification to 1.69 m2/ha in fuzzy classification, this show a promote in classification using fuzzy estimation.
5: Conclusion:
The proposed fuzzy methodology showed great promise for the estimation of forest parameters. However the procedure was only tested on a single data set. Further testing in different areas and with other imagery data would be useful to validate the results obtained. In general, the method should benefit from a higher density of references plots, even if this could involve an increasing burden of computational complexity.
From a methodological viewpoint, this approach should be most effective in reducing biases in the estimated parameters. Also, the use of fuzzy classification permits considerable flexibility in the estimation procedures. For probabilistically inserted into the process.
Theoretically, the fuzzy approach to estimation could be applied to an enormous series of problems of environmental monitoring, such as agricultural, oceanographic and geological investigations. In all these cases, the method can be expected to produce optimum performance thanks to its probabilistic, flexible nature. Research is currently directed towards the exploration of these possibilities.
6: Refrences:
- F. Wang, “A knowledge-based vision system foe detection land changes at urban fringes.” IEEE Trans. Geosci. Remote Sensing, vol.31, pp.134-145, 1993.
- F. Maselli, A. Rodolfi, and C. Conses. “Fuzzy classification of spatially degraded TM data for the estimation of sub-pixel components, “submitted to Int. J. Remote Sensing.
- G. M. Foody, “Ordinal-level classifications of sub-pixel tropical forest cover,” Photogrammetric Eng. Remote Sensing. Vol 60, pp. 61-65, 1994.
- T. W. Anderson, An Introduction to multivariate statistical Analysis New York: Wiley, 1984, pp.451-479.
- Altman E.I., Corporate financial distress and bankruptcy: a complete guide to prediction & avoiding Distress and profiting from bankruptcy, John Wiley &Sons, 1993
- Bezdek J.C., “Pattern Recognition with Fuzzy Objective Function Algorithms,” Plenum Press, New York, 1981
- Bian H.Y., “Fuzzy-Rough Nearest Neighbor Classification: an Integrated Framework,” proceedings of IASTED International Symposium on Artificial intelligence and Applications, 2002, 160-164
- Bian H. Y., “Fuzzy-rough NN Approach and its Application to China Listed Companies Profit Prediction,” Master of Philosophy Thesis, City University of Hong Kong, 2002
- Cover T.M., and Hart P.E, (1967), “Nearest neighbor pattern classification,” IEEE Transactions on Information Theory, IT vol (13), January 1967, 21-27
- Dimitras A.I., Zanakis S.H. and Zopounidis C., “A survey of business failures with an emphasis on Prediction methods and industrial applications”, European Journal of Operational Research, 1996, 487-513
- Dubois D. and Prade H., (1990), “Rough Fuzzy Sets and Fuzzy Rough Sets,” International Journal of General Systems, 17, 1990, 191-209
- Duda R.O. and Hart P.E., Pattern Classification and Scene Analysis, New York: Wiley, 1973
- Estes W. K., Classification and cognition: Oxford Psychology series 22, Oxford University Press & Clarendon Press, 1994
- Keasey K. and Watson R., (1991), “Financial distress prediction models: a review of their usefulness”, British Journal of Management, Vol 2, 1991, 89-102
- Keller J.M., Gray M.R. and Givens J.A., “A Fuzzy K-Nearest Neighbor Algorithm,” IEEE Transactions on Systems, Man and Cybernetics, 15(4), 1985