Journal of Applied Spectroscopy, Vol. 71, No. 2, 2004

CLASSIFICATION OF FOREST AREAS BY SPECTROZONAL IMAGES OBTAINED FROM ONBOARD A HELICOPTER B. I. Belyaev,a Yu. V. Belyaev,a L. V. Katkovskii,a∗ T. M. Kurikina,a A. A. Kazak,a and V. I. Shuplyakb

UDC 535.3+528.85+519.272

A technique of surveying forest territories from onboard the VSC-2 helicopter using a videospectral complex for forest monitoring is described. The results of processing the obtained spectrozonal images with separation of the areas occupied by different species and withering spruce trees are presented. Keywords: aerial photography, spectrozonal image, mosaic, color contrast, classification with training, withering spruce groves. Introduction. Within the framework of the State Scientific-Technical Program "Forests of Belarus and Their Rational Utilization," at the Institute of Applied Physical Problems of Belarusian State University the airborne VSC-2 videospectral complex has been developed, created, and tested. It is intended for controlling the state of forest resources from onboard aircraft in a timely manner [1, 2]. During the vegetation periods of the years 2001–2002, the device was subjected to laboratory and field testing. Distant measurements were carried out from onboard an Mi-2 helicopter of the "Bellesavia" State-owned Enterprise, and a technique of acquisition, processing, and analysis of the data has been developed. The VSC-2 simultaneously records three spectrozonal (spectropolarizational) images, surveying television (TV) images, and a high-resolution spectrum (1024 channels in the 350-1150-nm region) at separate points of the airway (4–6 spectra per image of a scene). As objects of investigation, we selected testing grounds of some forest ranger stations of the Stolbtsy and Molodechno regional forestries of Belarus. The choice was due to a number of reasons. First, the areas selected are rich enough in diverse species composition of stand. Second, according to the data of ground surveys and the first aerophotographs taken in 2000, within the compartments selected, plots with withering spruce groves damaged by the bark beetle (Ips typographus), and also centers of root rots and fungi have been discovered. Detection and mapping of withering spruce stands constitute the main concern of the program "Forests of Belarus and Their Rational Utilization." Surveying Technique. Surveying was done from a height of 200–1000 m to determine optimum heights for solving different problems. The speed of the helicopter was maintained almost constant at 90 km/h. The flight directions for plotting were selected to be North–South or West–East, depending on the configuration of the test objects. Detailed plottings were carried out over test objects conventionally called "Shpaki" (Rubezhevichi Forest Ranger Station, Stolbtsy Regional Forestry) and "Lebedevo" (Lebedevo Forest Ranger Station, Molodechno Regional Forestry). The distance between parallel airways over test objects was selected so that they could overlap by about 30–50%, which amounted to 300 m from a height of 600–800 m. To develop optimum techniques of aerial photography by the VSC-2 when recording spectrozonal images, different combinations of interferential polarizational filters of a spectral-polarizational surveying block (SPSB) for color and contrast separation of the objects of silva were selected. We used interferential light filters with the centers of transmission bands at 490, 560, 655, 720, 820, and 870 nm. To carry out polarizational surveyings, in all three channels of the SPSB three identical spectral filters were installed in each channel with the following set of wavelengths: *

To whom correspondence should be addressed. a

A. N. Sevchenko Institute of Applied Physical Problems, 7 Kurchatov Str., Minsk, 220064, Belarus; e-mail: [email protected]; bBelarusian State University, Minsk, Belarus. Translated from Zhurnal Prikladnoi Spektroskopii, Vol. 71, No. 2, pp. 241–247, March–April, 2004. Original article submitted July 2, 2003. 0021-9037/04/7102-0263 ©2004 Plenum Publishing Corporation

263

490, 560, and 655 nm. The polarizational filters were installed with different orientations of the optical axis (0, 45, and 90o) in each channel of the SPSB. However, the results of the polarizational survey are not analyzed in this work. To obtain spectrozonal images, the following combinations of interferential light filters were used installed simultaneously in the three channels of the SPSB: 1) 490, 560, and 655 nm; 2) 560, 655, and 720 nm; 3) 560, 655, and 820 nm; 4) 655, 720, and 870 nm; 5) 490, 560, and 720 nm; and 6) 560, 720, and 820 nm. The orientation of the polarizational filters at this time in all three channels was the same (0o). Further investigations have shown that the best results, as concerns detection and staining of vegetable objects, are obtained when light filters are used with the transmission band centers 560, 655, and 820 nm, although results that were not bad were also obtained when the filters were used at 490, 560, and 720 nm. The selected spectral bands in obtaining spectrozonal images are applied in remote probing to isolate vegetation against the background of other underlying surfaces, different kinds of vegetation, and also different stages of vegetation of the same species [3]. The choice of these bands was also attributable to the results of studying the informativeness of spectral channels in the course of numerous laboratory investigations [4]. Preliminary Processing. Both videoimages recorded by a surveying TV camera in the three standard (R, G, B) channels and then recorded by a video magnetophone and spectrozonal photographies were subjected to processing. Preliminary processing of images included numbering of TV images and their selection, systematization, and cataloguing. Here, not all sequential frames were selected for numbering, but only those that ensure sufficient overlapping to obtain a mosaic of images along the airway. The surveying TV images not only played an accessory role, but were themselves of great value. As accessory images they are used to ease fixation of spectrozonal images to one another, to a chart, and to the grid (boundaries) of compartments and plots of the geoinformation system "Forest Resources." This function is ensured by the wider angular field of vision of the surveying TV camera and also by the fact that surveying images are recorded in natural colors and are easily discerned by the eye. The surveying TV images are valuable in themselves, because they, just as the spectrozonal ones, can be used for thematic processing and classification of forest objects. Special emphasis should be placed on the advantages and drawbacks of the images obtained from onboard a helicopter, which influence the quality and character of the resulting products (thematic diagrammatic maps) in comparison with the images obtained from space vehicles and airplanes. Some positive features are evident and generally known: high spatial resolution, promptness, high mobility in selecting the objects for investigation and surveying parameters, the absence of the distorting effect of the atmosphere, etc. At the same time, among the substantial drawbacks (unless a hydrostabilized platform is used) are such phenomena as instability of the direction of the optical axis of the device, which is rigidly bound with the helicopter, due to an accidental bank, pitching, and yawing of the helicopter, strong influence of wind (here the speed of flight may change in the case of different airways). This makes the trajectory of flight along the Earth’s surface curvilinear with unalterable (on the average) direction of the longitudinal axis of the helicopter (its heading). All of these reasons lead to the necessity of assigning a sufficiently large overlapping of both successive frames inside each airway and of the airways themselves and make the construction of areal mosaics of images difficult (especially automated construction using the data of the GPS block available in the VSC2). As a result, the areal mosaics may not cover the entire area of the scene; intraway and especially interway omissions are possible. We note that in this case even the availability of a rather exact geofixation (with the aid of GPS) does not solve the problem, because of the oscillation of the device axis. The geographic coordinates put out by the GPS could, of course, have eased the map and GIS control of images; however, the available GIS "Forest Resources" is not tied with the global coordinates and has only local (relative) coordinates within the limits of each forest ranger station, which makes the creation of large video diagrammatic maps difficult. The above-listed problems are absent (or have been solved) for images obtained from space vehicles; partially solved (or exert a relatively small influence) for images obtained from onboard an airplane, and yet remain unsolved for surveying from onboard a helicopter not furnished with hydrostabilized platforms. The different level of illumination of the Earth’s surface on various frames and surveying airways leads to the appearance of sharp luminance boundaries between neighboring frames inside mosaic images. This problem is typical of all the methods of remote probing, in which the absolute values of the spectral density of radiance (SDR) are measured. This requires carrying out of radiometric correction. To obtain images of the spectral luminance factor, it is necessary to record the spectrum of illuminating radiation, but this too does not entirely solve the problem occurring in 264

Fig. 1. Images of the forest areas "Lebedevo" (a) and "Shpaki" (b) in the spectral channel at 820 nm; figures denote the positions of compartments and plots (compartment number/plot number), and the areas for which the results of computer classification are given in Fig. 2 are enclosed in frames. aviation (under clouds) measurements. The thing is that different areas of the Earth’s surface (even within the limits of one frame) can be darkened by clouds of different optical density which create different illumination of the corresponding areas, and this cannot be taken into account. Many programs for processing images (in particular, ENVI) have different (more or less successful) functions of equalization of images by brightness (for example, matching of the brightness-value distribution histograms). However, this smoothing may lead to distortion of the true values of spectral density of the radiance of objects, which, naturally, is inadmissible from the viewpoint of their classification. One way of solving the indicated problems is to use of such methods of classification, the result of the application of which is independent of the level of illumination (i.e., when the shape of the spectral curve is analyzed). One such method of classification implemented in the ENVI program is the SAM (spectral angle mapper) method, in which each spectrum is characterized only by the direction of the vector in the spectral space whose axes correspond to the spectral channels of the image. We suggest the following technique for reducing a set (chain) of overlapping images to one illumination. We seek a set of pixels, corresponding to the same object, in the overlapping parts of neighboring frames, and equalize 265

TABLE 1. Taxation Description of Some Plots of Standard-Calibration Areas No. of comArea partment (No. of plot) of plot, ha

Type of forest

Prevailing species, composition factor age/height/diameter/density

Second species, Third species, composition factor composition factor age/height/diameter age/height/diameter/density Underwood /density

"Lebedevo" (Molodechno forestry, Lebedevo forest ranger station) 142 142 142 143 143

(12) (13) (14) (10) (11)

0.4 2.9 0.4 1.5 2.4

long moss spruce, 7, 10/8/–/– sorrel spruce, 10, 45/22/22/70 bilberry treelike willow, 10, 10/6/4/70 sorrel spruce, 10, 60/27/28/60 sorrel spruce, 9, 65/28/28/60

birch, 3 – – – aspen, 1

– – – – – treelike willow, 1, – /15/12/– spruce, 5, –/27/28/– – –

– – hazel – –

143 (12)

2.7

sorrel

spruce, 4, 25/10/10/70

birch, 5, –/17/14/–

143 (17) 143 (18) 143 (20)

4.8 2 –

sorrel sorrel sedge bog

oak, 4, 65/24/26/60 spruce, 10, 45/22/24/60 no regeneration

birch, 1, –/24/20/– – –

126 (9)

21.9

sorrel

oak, 6, 70/23/28/60

spruce∗, 4, – /28/28/–

–

–

126 (16)

0.8

sorrel

spruce cutover, 10, 10/4/–/–

–

–

hazel, raspberry

126 (17)

6.1

sorrel

spruce,∗ 9, 75/30/30/70

pine, 1

–

–

– – – –

"Shpaki" (Stolbtsy forestry, Rubezhevichi forest ranger station) 3 (14)

4

sorrel

spruce, 4, 35/15/16/70

hornbeam, 3

birch, 3

3 (15) 3 (16)

3.9 1.3

sorrel sorrel

spruce cutover spruce, 5, 15/6/6/60

– birch, 3

– aspen, 2

3 (23)

6.6

sorrel

spruce,∗ 9, 85/27/32/60

aspen, 1

–

4 (4)

6.3

sorrel

spruce, 8, 70/23/26/60

birch, 1

aspen, 1

4 (6)

10.1

spiraea

black alder, 7, 50/20/20/70

birch, 1

spruce, 1

4 (7)

3.1

spiraea

black alder, 6, 10/5/6/60

birch, 3

spruce, 1

4 (8)

7

sedge

birch, 5, 35/13/12/70

black alder, 4

spruce, 1

4 (11)

6.9

goutweed

spruce, 9, 75/25/28/60

birch, 1

–

hazel, buckthorn – hazel buckthorn, hazel buckthorn, hazel buckthorn, willow buckthorn, willow willow bishes buckthorn

4 (13)

8.8

goutweed

spruce, 10, 65/25/26/40

–

–

–

4 (14)

25

sedge

black alder, 7, 50/20/20/70

birch, 1

spruce, 2

buckthorn, willow

4 (18)

1.7

brake

oak, 4, 2/–/-/70

spruce, 6

birch, 6, 2/1/–/50, aspen, 4

–

∗

∗

Spruce damaged by phytoinsects with signs of withering.

their luminances by multiplying the second image by the average ratio of the luminances of these pixels in the first and second images. Thereafter, this operation is repeated for the next pair, where the second image of the previous pair is the first image, and so on. This procedure does not distort the relative values of the spectral density of radiance in contrast to the matching of histograms. In processing images, photometric correction was not carried out, but rather the methods of classification were used which were almost insensitive to illumination, or the areas of a forest were taken for classification which were under identical illumination conditions when surveyed. The result of preliminary processing of the obtained images of forests are their areal mosaics fixed to the corresponding vector maps of the GIS "Forest Resources" with the boundaries of compartments and plots of forest ranger 266

stations. Mosaic images of overlapping frames and airways in which identical objects are determined visually were constructed interactively using the program packets Corel Draw, Corel Photo Paint, MS Photo Draw, and also the specialized packet ENVI. For nonoverlapping (or overlapping in separate frames) airways, first each airway was fixed to the grid of the boundaries of the plots in the GIS "Forest Resources" and then an areal mosaic of individual airways already "geofixed" (relatively, i.e., within one forest ranger station) was constructed. Thematic processing of the data of aviation surveyings consists of sedimentation (isolation of homogeneous as to their spectral properties) objects of images (groups of pixels) and their assignment to one predetermined class or another. The initial (requiring classification) data are either areal "geofixed" mosaics of images (also individual images) obtained with the aid of a surveying television camera or areal mosaics of spectrozonal images obtained with the aid of a spectral-polarizational photography block. To carry out thematic classification of images, a set of data are used which underlie the assignment of the unknown classes of forest vegetation and the formation of standard data for these classes (training samples): high-resolution spectra obtained from onboard a helicopter by an MS-09 spectrometer; taxation descriptions of plots contained in the database of the GIS "Forest Resources," taxation data of current examinations of forest territories. It should be noted that the videoimages themselves can also be used for approximate visual thematic classification of forest territories (forest areas with coniferous and deciduous species, forest meadows, underwood, crowns of dry spruce trees differ rather distinctly on the pictures taken, without making any other efforts). However, much better color display of species and the state of forest vegetation than on surveying pictures can be obtained on synthesized images obtained by output of spectroscopic images on a display, through the R, G, and B channels, in the spectrum bands selected by us: 560, 820, and 655, respectively. Examples of this kind of separation and classification of areas by species and phytopathological composition are given in Fig. 1. Taxation data of the designated plots of a forest taken from the GIS "Forest Resources" are listed in Table 1. Classification by Color. This is visual classification of pseudocolor spectrozonal images. Below, we give color characteristics of different classes which correspond to the color image synthesized through the R, G, and B channels from the spectrozonal images in the 569, 820, and 655 nm channels, respectively. Figure 1a presents the initial mosaic image of a portion of the Lebedevo forest ranger station with fixation to the boundaries of compartments and plots, which was obtained in the 820 nm channel (the filter halfwidth was 40 nm), and Fig. 1b gives a similar image for a portion of the Rubezhevichi forest ranger station (Shpaki area). It is seen from Fig. 1a that over the forest area with neighboring plot Nos. 12, 13, and 14 of the 142nd compartment, there are distinct color displays of a synthesized image. The reason for this is evident: the 12th plot consists of a mixture of young spruce and birch trees (the color is light brownish-green), the 13th plot consists entirely of mature spruce trees (dark brown), and the 14th plot is almost entirely covered with deciduous species (green). It is possible to name some other contrast differences of plots. Neighboring plot Nos. 17, 18, and 20 from the 143rd compartment also differ distinctly in color (light brownish-green, dark green, and light green, respectively) due to their species composition: mixed, spruce forest, and bog (see Table 1). Similarly, a mixed forest of plot No. 12 of the 143rd compartment is distinguished; it is green and borders almost entirely on spruce forest (plot Nos. 10 and 11, compartment 143). Moreover, the 10th plot, where spruce occupies 100% of the territory, certainly looks darker than the 11th one, where spruce occupies only 90% of the area. Moreover, the indicated plots have a notable reddish hue. As is known from the experience of processing these kinds of images, this points to the appreciable fraction of withering and damaged trees, which was confirmed by subsequent ground survey. Based on Fig. 1a, we give one more example of such contrasting. We refer to neighboring plot Nos. 9, 16, and 17 of the 126th compartment. Plot No. 16, where the spruce trees had been cut down, looks yellowish-green, whereas plot No. 9 (mixed forest) is light brownish-green, and No. 17 (spruce with some pine) is brown with a reddish hue, pointing to the withering of spruce (see Table 1). Thus, when putting out images in the considered 560-, 820-, and 655nm channels through the RGB channels, the withering spruce plantations have a characteristic reddish-lilac hue. Similarly, we will analyze Fig. 1b. For example, over the forest area with neighboring plot Nos. 4, 6, 7, 8, and 11 of the fourth compartment, different species and states of trees are sharply distinguished. Plot No. 4 consisting 80% of spruce and 20% small-leaved species (see Table 1) looks like a mix of regions with brownish-green and lightgreen colors. Plot Nos. 6, 7, and 8, which include mainly deciduous trees, have green and grayish-green colors, 267

field deciduons forest spruse, 45 years old spruce, 10 years old withering spruce

field deciduons forest sound spruse felled area withering spruce Fig. 2. Computer thematic classification of the fragments of images (Fig. 1): a) "Lebedevo" area; b) "Shpaki" area. whereas "spruce" plot No. 11 has a grayish-brown color with small greenish spots (we note that here too areas with a red hue can be seen, pointing to withering of spruce trees). Next, neighboring plot Nos. 13, 14, and 18 from the fourth compartment also display substantial differences in color. The reason is that the 13th plot (reddish-brown) consists of dry and withering spruce and the 14th plot (green and grayish-green) has mainly deciduous trees, whereas the 18th plot (yellowish-green) consists of young two-year-old oak trees and other species (see Table 1), which evidently was unable as yet to screen the underlying bushes and grass. A similar classification can be made for the forest area with neighboring plot Nos. 14, 15, 16, and 23 of the third compartment: "spruce" plot No. 23 with a considerable fraction of damaged trees has a brown color with red spots, plot No. 14 mainly consisting of spruce and hornbeam has a green and dark-green color, whereas the 16th plot consisting of young spruce trees and small-leaved trees has a light green color, and finally, the 15th plot (tested area) has a yellow color. Thus, creation of synthesized images, when the spectrozonal images recorded with different spectral filters (in our case at 560, 820, and 655 nm) are fed into the standard TV R, G, and B channels, make it possible to carry out

268

thematic processing of the image and to classify forest areas by species (spruce or pine forest, fir or deciduous species, small-leaved or broad-leaved species) and by phytopathological state (damaged or sound spruce, etc.). Computer Classification with Training. The above-described visual classification is usually made at the first stage of analysis and plays an auxiliary role, allowing one to preliminarily determine the approximate number of classes and their spatial localization, which thereafter can be used for assigning the regions of interest (training samples). It does not allow one to determine, with a needed accuracy, the spatial boundaries of classes, their areas, and statistical parameters. The most precise thematic classification with determination of quantitative parameters is performed by computer methods of pixel-by-pixel classification with (or without) training. Here, different variants (depending on the structure data) of determining interclass distances (metrics) are used, which are implemented in certain specialized programs (ENVI, Erdas Imaging, etc.) and assignment of the rules (functions) of solutions, for example, parallelepiped, minimum distance, maximum likelihood, Mahalanobis distance, method of spectral angle, etc. The method of maximum likelihood yields good classification results. As an example, Fig. 2 presents the results of thematic computer classification with training by the method of maximum likelihood for the fragments of images shown in Fig. 1 in rectangular frames. To assign the regions on the images that were used to calculate training samples, we selected plots with a rather homogeneous species composition according to the database of the GIS "Forest Resources." Localization of areas with different species (for example, deciduous pine, etc., species) was made using the above-described color characteristics of a spectrozonal image and also characteristic spectra. In particular, for the Lebedevo forest ranger station, training samples were assigned using small homogeneous plot Nos. 12, 13, 14, and 16 of compartment 142. Five classes have been determined. Using the lower parts of plot Nos. 12 and 14, we prescribed the training part for the class "Deciduous trees" and using the upper parts of plot No. 12 — for the class "Young spruce trees" (10-years old). Based on a small number of pixels of the upper part of plot No. 13, we assigned the training part of the class "Maturing spruce trees" (45-years old), and using a part of plot No. 16 with a reddish-crimson hue, we assigned the training part for the class "Withering spruce trees." One other class corresponds to the field. According to the results of classification of a fragment of the Lebedevo forest ranger station (Fig. 2a), unclassified pixels are practically absent. Since a rather high threshold probability of assignment of a pixel to one class or another has been assigned (0.98), the results obtained point to the fact that the number of classes and their determined composition fit the forest territory well. Figure 2b presents a similar classified fragment of an image for the Shpaki area (Fig. 1b). In contrast to the previous case, here a somewhat different composition of classes has been determined, since there are felled areas on the territory considered, and spruce plantations differ greatly in age. Conclusions. For the user, most convenient are the thematic maps in which the original image (surveying or spectrozonal) is superimposed by one, two, or three individual classes colored in conventional contrast colors. Maps of precisely this kind are the final product of operational distant monitoring of forest areas. Thematic diagrammatic maps with isolation of the basic classes of forest territories (special attention was paid to the mapping of withering spruce plantations) have been obtained virtually for all of the forest ranger stations for which surveyings were made. The results of computer classification were compared with taxation descriptions on the basis of the "Forest Resources" database (percentage composition of different species). Correction of the results of classification and descriptions of the GIS "Forest Resources" is rather high in all of the cases and is equal to 75–95%.

REFERENCES 1. 2. 3. 4.

B. I. Belyaev, L. V. Katkovskii, and V. A. Sosenko, in: Selected Sci. Papers of Belarusian State University in 7 Volumes [in Russian], Vol. 4, Physics, Minsk (2001), pp. 103–125. B. I. Belyaev, Yu. V. Belyaev, V. V. Veller, et al., in: Proc. Sci. Conf. "Remote Probing of the Earth’s Mantle and the Atmosphere by Aerospace Means [in Russian], June 20–22, 2001, Murom (2001), pp. 193–196. S. V. Garbuk and V. E. Gershenzon, Space Systems of Remote Probing of the Earth [in Russian], Moscow (1997). B. I. Belyaev, Yu. V. Belyaev, V. M. Zaletnyi, et al., in: Proc. Sci. Conf. "Remote Probing of the Earth’s Mantle and the Atmosphere by Aerospace Means [in Russian], June 20–22, 2001, Murom (2001), pp. 197–202.

269