Forest Canopy Density Classification Using Texture Quantization of Panchromatic Aerial Images

By Hamed Ashoori, MasoudTaefi Feijani, Valadan Zoej

ABSTRACT

Forest canopy density is an important criterion for forestry application. Several methods were introduced to compute these criteria. None manual methods use multispectral images to determine the canopy density. It also could be extracted from Arial images manually. Arial images are valuable data with rather high special resolution, but some of them are panchromatic and are not suitable for spectral processing. Image texture which is valuable contextual information helps the interpreter to distinguish different canopy density areas. In this paper panchromatic Arial images were used to classify forest canopy density cover. Texture features were generated from image and use beside the image in classification. The results shows good improvement in classifying different canopy density covers.

INTRODUCTION

The importance of studying,Forest is a very complex ecosystem. The complexity of forested area in Iran is more than similar area in the world. The Caspian Hyrcanian mixed forest in the north of Iran has a very great diversity. UNESCO classifies the Forested area in Iran as a natural world heritage sites for their great age and diversity. Forest in Iran is habitat of many endemic and semi-endemic and relic species; these exclusive properties make difficult processing and analyzing of satellite images.

Different methods were introduced to estimate forest canopy density; some of them are based on multispectral images which are rather expensive for large areas. Arial panchromatic images have been captured approximately for all the parts of Iran. These images could not be used for spectral processing, but has valuable information which could be used for interpretation and classification of forest canopy density visually. Sotexture quantification could be used to generate new features from panchromatic image, then they are used with the source panchromatic band as input data for classification.

Forest Canopy Density Estimation Models
Since now many models have been used for estimation forest canopy density and biomass inventory from satellite images. Forest canopy density or FCD is a very important factor for forest management and assessment. Some of the general methods for this issue is explained in short term as below:

Artificial neural network (Boyd et al., 2002),Artificial neural networks are neurologically inspired statistical mechanisms also employed in classification of forest cover using various sensors (Boyd et al., 2002). A three layer feed forward error-back propagation artificial neural network implemented in Interactive Data Language (IDL) was used in order to predict forest canopy density as a continuous variable. The algorithm minimizes the root mean square error between the actual output of the multi-layered feed forward perceptron and the desired output (Skidmore et al., 1997). Following Atkinson and Tatnall (1997) to search for system parameters to increase the accuracy of the method and avoid overtraining of the neural network. The neural networks with the sub sample of 186 sites with canopy density and the seven ETM+ bands was trained and the best combination of optimum learning rate and momentum to minimize the root mean square error (RSME) was empirically established. The best results were obtained with a learning rate of 0.7, a momentum of 0.7 and four hidden nodes. The RSME stabilized after approximately 7000 epochs. Finally, 20 iterations of 7000 epochs were performed and the best one was selected based on root mean square error.

Multi Linear regression techniques (Iverson et al., 1989 ; Levesque and King, 2003),Multiple linear regression techniques have been used to model the relation between spectral response and closed canopy conifer forest cover (Ripple, 1994). In this study, a multiple linear regression model has been developed, which best described the relation between canopy density and the seven ETM+ spectral bands. The regression equation using n = 186 observations is:

Y= 3.32 + 0.021B1 – 0.002 B2 + 0.003 B3+0.024 B4+0.023 B5+0.021 B6-0.029 B7 (1)
Where Y is the predicted forest canopy density and B1–B7 is the reflectance value of bands 1–7 of Landsat ETM+ image.

Forest canopy density mapper (Rikimaru, 1996), Rikimaru introduced an alternative deductive approach, i.e. forest canopy density mapper to map forest canopy density using four indices (vegetation, bare soil, shadow and surface temperature) derived from Landsat TM imagery.
Based on these four variables, nine canopy density classes, namely 0, 1–10, 11–20. . . 71–80+ were obtained. This model involves bio-spectral phenomenon modeling and analysis utilizing data derived from four indices, namely:

• advance vegetation index (AVI)
• bare soil index (BI)
• shadow index (SI)
• Thermal index (TI).

Using these four indices the canopy density for each pixel was calculated in percentage
The method requires intervention by an operator to establish threshold values. The accuracy obtained in three SE Asian countries averaged 92% (Rikimaru, 1996).
Maximum likelihood classification (MLC),as a parametric classifier, maximum likelihood classification method calculates the probability that a given pixel belongs to a specific class and assigns the pixel to the class having the highest probability (Richards, 1999). The training set of 186 pixels into 10 canopy classes, namely 0, 1–10, 11–20, . . ., 71–80+ was classified. The Interactive Data Language (IDL 6.0) and ENVI 4.0 (ENVI, 2003) was used for image classification. (Chudamani Joshi, 2005)
They are the more general models used for forest canopy classification. Some of others methodologies are:

• Object based classification (Dorren et al., 2003)
• Decision tree classification (Souza et al., 2003)
• Spectral unmixing at pixel or subpixel scale (Cross 1991)

General Models Used in Operational and Research Project in Iran
In international scale many models have been used in operational projects. Great center and organizations such as USGS, CRC, ITTO& FAO used one of the mentioned models for their activities under operational large projects.
In IRAN only research project that applied by universities or research institutes used professional models and algorithms.
Administrative agencies and executive organization used very elementary and basic models for their simplicity of implementation. Our study shows that 3 model (or indices) is very popular in these projects:

• NDVI
• Principle Component Analysis
• Visual Interpretation

Along with simplification of these models and indices the main reason for this issue is lack of satellite and field data in Iran. Indeed high spatial resolution data of natural and forested area is very low besides related lack of simultaneous field and train data.

Aerial photo interpretation of natural resource is a common activity that is being followed from near 1960 till now. And mid resolution images are accessible from 1973 (1 year after launching landsat#1).although any model that can use the aerial photo as a source of data not face with lack of image data, because as mentioned above they are taken from 1960 by 5-10 year of intervals. According to these capacities and limitations we use Panchromatic Arial Images in our model as aninput image data.

DATA SPECIFICATIONS
Two panchromatic Arial stereo images from Zagros MountainsGavbarg region inYASUJ province area where used. Images were captured at 1999 in 1:40,000 scale. At first images was oriented using ground control points, then DEM were generated in overlap area. Ortho image were generated using one oriented image and generated DEM.
Train and check data were selected using manually classified image. (Figure 1)
Sixclasseswere selected based on their canopy density which was determined manually, and 500*500 subsets were selected around each class.Selected classes were very dense canopy to bare land; they were ranked based on the canopy density. The test image was generated by mosaicking the 6 subsets. (Figure 2)

Feature generation
Different methods were introduced by authors for quantifying image texture. These methods could be used to generate image base features. Generated features could improve image classification accuracy beside spectral features.
Several methods were used to generate images, mean and variance from first statistical methods, direct variogram and madogram from geostatistical methods, low-pass and high-pass ringing and slice filters from Fourier based filters. These features use the following equations to generate features.

First Order Statistical Features
If (I) is the random variable representing the gray levels in the region of interest, the first order histogram P (I) is defined as (Theodoridis, 1999):

Now different features can be generated by using the following equations.

Moment

Where Ng = number of gray levels.
is the simple mean of pixels. Also 2nd, 3rd and other moments can be used.

Central Moments
(3)

Geostatistical Features
Geostatistics is the statistical methods developed for and applied to geographical data. These statistical methods are required because geographical data do not usually conform to the requirements of standard statistical procedures, due to spatial autocorrelation and other problems associated with spatial data (http://www.geo.ed.ac.uk).
Semivariogram that represents half of the expectation of the quadratic increments of pixel pair values at the specified distance can quantify both spatial and random correlation between the adjacent pixels. (Goodenough, et al, 2003) It is defined as:
(4)



That is the classical expression of variogram (h) here represents a vectorial lag between pixels. In this study direct variogram, madogramvariogram have been used.

Direct Variogram

In this approach the following equation is used to estimate:

(5)



n(h) is the number of pairs that are in mask filter.
Madogram

This is similar to direct variogram except squaring differences, but uses the absolute value of differences.
(6)

Fourier Based Features
Fourier transformation, transforms a signal from space/time domain to frequency domain. The amplitude and phase coefficients are two outputs of a Fourier transformation. So different texture patterns could be identified by their Fourier coefficients but because in this research one value for each pixel is required, raw Fourier coefficients couldn’t be used. Several features can be generated using sum of the Fourier amplitude under different masks (Pratt, 2001). These are comprised ringing, sectorial, horizontal and vertical which are shown in figure 2.

(7)



(8)



Figure3. Different mask which can be used to generate features from Fourier coefficients
Different parameters can be set in each method; the main parameter is window size. Different window sizes were used to generate features in each method.All features were generated using
3*3, 9*9, 15*15, 21*21, 27*27 and 33*33 window sizes. In geostatistical method four lags were used, they are [1, 0], [1, 1],[1,1] and [-1,1] also in Fourier method, two masks which are shown in figure 2 were used and high frequency and low frequency features were generated using each mask.

Implementation, Results and Conclusion
To evaluate the effect of using generated features in classification process, firstly the image was classified using dn slicing (parallel pipe classification) because the input is a gray-scale panchromatic image and couldn’t be classified using other methods.
All possible features using mentioned methods and selected window sizes and lag or mask parameter were generated. Then classification was done using each generated image beside panchromatic image as input data. Accuracy assessment was done, generating confusion matrix and computing overall, kappa and producer accuracies.
Table 1 shows parallelepiped results and best results obtained for each class user accuracy and overall and mean accuracy.

Results shows that using generated features could separate different forest canopy densities better and increase classification accuracy. It could be said that different methods for generating image based features should be used for different aims (e.g. Madogram for class 2); also some features could be used for general improvement (Mean with 33*33 window size).

References

References from Journals

Atkinson, P.M., Tatnall, A.R.L., 1997. Introduction neural networks in remote sensing. Int. J.Remote Sensing 18, 699–709

Boyd, D.S., Foody, G.M., Ripple, W.J., 2002. Evaluation of approaches for forest cover estimation in the Pacific Northwest, USA, using remote sensing. Appl. Geography 22, 375–392

Chudamani Joshi, Jan De Leeuw, Andrew K. Skidmore, Iris C. van Duren, Henk van Oosten, 2005, “Remotely sensed estimation of forest canopy density: A comparison of the performance of four methods”, International Journal of Applied Earth Observation and Geoinformation

Cross, A.M., Settle, J.J., Drake, N.A., Paivinen, R.T.M., 1991. Subpixelmeasurement of tropical forest cover using AVHRR data. Int. J. Remote Sensing 12, 1119–1129.

Haralick, R.M., Shanmugam, K., Dinstein, I., 1973. “Textural features for image classification”, IEEE Transactions on Systems, Man and Cybernetics, vol. 3, no. 6, pp 610-621.

Iverson, L.R., Cook, E.A., Graham, R.L., 1989. A technique forextrapolating and validating forest cover across large regions:calibrating AVHRR data with TM data. Int. J. Remote Sensing 10,1805–1812

Levesque, J., King, D.J., 2003. Spatial analysis of radiometric fractions from high-resolution multispectral imagery for modelling individual tree crown and forest canopy structure and health. Remote Sensing Environ. 84, 589–602.

Skidmore, A.K., Turner, B.J., Brinkhof, W., Knowle, E., 1997. Performance of a neural network: mapping forests using remotely sensed data. Photogrammetric Eng. Remote Sensing 63, 501–514.

Souza Jr., C., Firestone, C.L., Silva, L.M., Roberts, D., 2003. Mapping forest degradation in the Eastern Amazon from SPOT-4 through spectral mixture models. Remote Sensing Environ. 87, 494–506.

References from Books
John A. Richards, 1999, “Remote Sensing Digital Image Analysis an Introduction”, Springer-Verlag

Pratt, 2001,” Digital Image Processing”

SergiosTheodoridis, 1999, “Pattern Recognition”, Academic Press
References from Other Literature

Ashoori, H., Alimohammadi, A., ValadanZoej, M. J., Mojarradi, B., 2006. Generating Imagebased Features for Improving Classification Accuracy of High Resolution Images, May, ISPRS Mid-term Symposium, Netherlands.

Dorren, L.K., Maier, A.B., Seijmonsbergen, A.C., 2003. Improved Landsat-based forest mapping in steep mountainous terrain using object-based classification. Forest Ecol. Manage. 183, 31–46.

Goodenough, David, A.S. Bhogal, R. Fournier, R.J. Hall, J. Iisaka, D. Leckie, J.E. Luther, S.Magnussen, O. Niemann, and W.M. Strome, Earth Observation for Sustainable Development of Forests (EOSD), Victoria, B.C.: Natural Resources Canada, http://www.aft.pfc.forestry.ca,1998

P.S. Roy, S. Miyatake and A. Rikimaru, “Biophysical Spectral Response Modeling Approach for Forest Density Stratification”, ACRS 1997

Rikimaru, A., 1996. Landsat TM data processing guide for forest canopy density mapping and monitoring model. In: International Tropical Timber Organization (ITTO) workshop on utilization of remote sensing in site assessment and planning for rehabilitation of logged-over forest, Bangkok, Thailand, pp. 1–8.
References from websites

http://www.geo.ed.ac.uk

Acknowledgements
Visual classified image which have been used as training source is received from “Forests, Range and Watershed Management Organization (FRWO); Engineering and Evaluation Bureau”

Source: Coordinates