Vegetation cover mapping for the Kronotsky State Biosphere Reserve (Russia, Kamchatka peninsula) using artificial neural network.
Thematic maps (geobotanical, forest inventory, land use, etc.) are the basis for decision-making in environmental management and environmental protection. However the most thematic maps were made in the 60's and 70's of the XX century and outdated to the present time. The wrong decisions is possible as a result of the lack of relevant mapping information on the status of a territory.
The study area encompasses the Kronotsky State Biosphere Reserve in its current boundaries. It located on the eastern coast of Kamchatka (Fig. 1). Mountains occupy a significant part of the protected territory. The plain areas make up about 10 % of the reserve. These are swamped marine terraces, water-glacial and alluvial plains. The flat areas volcanic valleys and volcanic calderas happen on significant heights too. The rest of the territory is occupied by slopes of varying steepness, sometimes very steep. The middle relief predominates over the reserve .
Complex of the tundra, forest and maritime meadow ecosystems succeed regularly each other on this area . Large range of altitudes within the reserve (from sea level to 3528 meters above sea level Kronotsky volcano crater) is responsible for the formation of a zonality with Ermans birch (Betula ermanii Cham.) forests, Japanese stone pine (Pinus pumila (Pallas) Regel) and siberian alder (Alnus fruticosa Rupr.) elfin woods, alpine and coastal tundra, swamps in the lower reaches of rivers, bald and glacial forms of vegetation on the tops of mountain ranges.
Vegetation map is made on the basis of multispectral satellite imagery medium resolution Landsat-7/ETM+ (6 spectral bands in the visible, near and middle infrared ranges, the spatial resolution of 30 meters per pixel). To ensure complete coverage of the territory cloudless remote sensing data were used 4 images Landsat-7, each of which provides a cloudless cover its part of the study area (Fig. 2).
So far as the middle latitudes location and mountainous, territory of the Kronotsky reserve is not evenly lit even when summer. Southern slopes are excessive lighting, while the northern ones are in deep shadow. Topographic normalization algorithm (Lambertian reflectance model) implemented in the GIS Erdas Imagine, was used to reduce the differences in illumination due to the slope and aspect of terrain relative to the elevation and azimuth of the sun (Fig. 3). The digital elevation model for topographic normalization was created using vector topographic maps. The net result is an image with more evenly illuminated terrain.
Interpretation of remote sensing data was based on the field survey area, thematic and topographic maps, literature sources and verbal information of a specialists. The field data, collected during 5 years, was granted by Neshataeva V. Yu. and Neshataev Yu. N. (St. Petersburg Botanical Institute of the Russian Academy of Sciences and Department of Geobotany of St. Petersburg University). 2209 relevé points with the names of the plant association were geocoded using topographic maps, text description and authors' verbal explanations (Fig. 4). Relevés, thematic and topographic maps, literature sources and verbal information of a specialists were used to determine the distinctive features of different types of vegetation.
The spectral profiles were created for each class of legend (plant association) by means of the relevé points. The significance of each band for the classification was defined. The near and middle infrared bands with the maximal differences between classes of the legend had greater weight during classification (Fig. 5).
The interpretation of multispectral images was performed in the software NeRIS (Neural Raster Interpretation System), developed by the Research and Development Center "ScanEx". The basic advantage of the program is the capability to use adaptive algorithms based on the Kohonen's artificial neural networks for the detection of the spectral structure of the source image, enabling to understand spectral correlation between all classes of the image being classified.
Untrained neural network is a regular 2-dimensional rectangular grid of 15 x 15, in which the nodes are individual neurons. A training of the neural network is made using special "labels" vector polygons, contouring different types of land cover. "Labels" were placed in a homogenous (not mixed) areas of a vegetation or mineral surfaces, occupying a large area to avoid edge effect. The neural network's training is a self-organization process in a multidimensional feature space, revealing the internal structure of a source multispectral image. The trained neural network is a irregular grid, oriented in a multidimensional feature space, in which the individual neurons have spectral properties similar to ones of pixels inside "labels".
The classification of multispectral images was performed through the trained neural network. Initially the original image was classified on 225 intermediate classes (Fig. 6). Each pixel was associated with a particular neuron using a measure of the proximity "City Block". Conversely, each of the 225 neurons was associated with a group of pixels with similar spectral characteristics in the original multispectral image. A gradient palette was created by means of the "color seeds" in a trained neural network. "Seed" is a metaphor meaning an individual neuron with a color assigned by user. The most relevant neurons (neurons, which belong most confidently to one of the legend's end classes) were selected as "seeds" to assign a colors. The colors of other classes are generated automatically by the interpolation of "seeds'" colors using Sammon mapping. As a result, the closer are the intermediate classes in the original image's feature space, the closer are a colors in the palette. At the same time, even the colors of related classes are some different, that allows the user to analize visually the primary classification result.
Closely spaced neurons were combined in the groups. Aggregation of the considerably excessive number of intermediate classes (225 intermediate classes versus 17 classes in the final legend) allowed to define the boundaries between the legend's end classes more precisely. At first the same color was appropriated to spectrally close objects. Then the aggregation of the objects with the same color in one class was made.
The results of the multiple images interpretation were combined into a general map. To eliminate statistically unreliable objects, whose width was 1 pixel, the raster image was processed by filter, which replaces the class of the small object to the class, that appears most often within a specified neighborhood. The floating window 3 by 3 with majority option was used for eliminating statistically unreliable pixels. Finally, the prepared raster image was vectorized. The numerical values of classes were replaced by the text descriptions in attributive table (Fig. 7).
The resulting vegetation cover map complete covers territory of the Kronotsky reserve in its present boundaries with a 3-kilometers buffer along the perimeter and reflects the state of the area at the beginning of the XXI century. A spatial resolution of the map is 30 meters per pixel and corresponds to the scale of approximately 1 : 200 000. The legend consists of 17 classes (Fig. 7).
The first column shows the percentage of relevé points, which came into polygons with the appropriate plant communities. The relevé points of the second column locate at a distance of no more than 30 meters (in the neighboring pixels) to the polygons with the appropriate plant communities; the third column at a distance of 30 to 60 meters (one pixel), etc. Based on possible geolocation error (100 meters), we guess, that the relevé points, which are at a distance of not more than 3 pixels from the polygon with appropriate plant community (left 4 columns) can be considered as relevant to the correct interpreted pixels. Thus, 72 % of the relevé points have the correct interpreted polygon in the nearest neighborhood.
28 % relevé points, which do not have a polygon with the appropriate plant community in a nearest neighborhood, have the following distribution by legend's classes (Fig. 9).
The forb meadow communities have the most number (more then 25 %) of interpretation errors. The meadows in Kronotsky Reserve rarely occupy vast areas. Often this is a small meadows surrounded by forest, or narrow ribbon-like strips along water streams. Small objects may be indistinguishable on medium-resolution images if they have a size comparable to the a pixel size and have spectral characteristics similar to surrounding objects. Moreover, some pixels interpreted as the meadows, were eliminated during the processing of the filter. Thus, the spatial resolution of the remote sensing data does not allow to reduce the error of interpretation, appearing as a result of presence of small objects.
The remote sensing data imposes some others restrictions on the result. Using satellite images does not allow to characterize a lower tiers in a dense forest communities. Therefore, this map does not contain legend's classes, which describe a herbaceous, moss, lichen tiers and undergrowth, closed by tree canopies. Legend to this map consists of 17 classes, which are detectable confidently on a middle resolution multispectral satellite images (Fig. 7).
Field survey allows to characterize the vegetation more detailed, including lower tiers. For example, the geobotanical map of the Kronotsky reserve  created in 1979 by Y.N. Neshataev contains 37 classes of the legend, which describe all tiers. This map was made on the basis of 5-year field surveys of the territory and using forest inventory data. Following the extrapolation was made by means of the selectively-statistical technique of the geobotanical mapping .
The relevance, spatial accuracy, high detail and GIS format (geotif) are the main advantages of the land cover map based on the multispectral medium-resolution images in compare with forest inventory data . All this makes it possible to use the result in a different analytical purposes without any additional processing.
The resulting vegetation cover map for the Kronotsky State Biosphere Reserve based on medium resolutiona remote sensing data is available for download: kz_vegetation_map2002.zip.
Participants of the project.
Project has been implemented by Biodiversity Conservation Center (BCC) with the participation of International Socio-Ecological Union (SEU) at the initiative and sponsorship of Wildlife Conservation Society (WCS) in the network of "Conserving the Volcanoes of Kamchatka World Heritage Site, Russia".
1. Vasiljev, N.G., Matjushkin, E.N., Kuptsov, Y.V. 1985. Kronotsky reserve. In: Sokolov, V.E., Syroechkovsky, E.E. (eds.) Reserves of the USSR. Reserves of the Far East. Mysl', Moscow.
2. The geobotanical map of the Kronotsky state reserve Kamchatka region. State 1976 y. Scale 1 : 100 000. Main Administration for Hunting and Nature Reserves under the Council of Ministers of the Russian Soviet Federative Socialist Republic.
3. The Geobotanical Map of the Kronotsky State Nature Reserve Compiled Using Materials of Reference Geobotanical Profiles of 1974-1978 and Forest Inventory Data of 1976 by Associate Professor Y. N. Neshataev, Department of Geobotany, Leningrad State University. Scale 1 : 100 000. 1979.
4. Neshataev, Y.N., Neshataeva, V.Y., Naumenko, A.T. Vegetation of the Kronotsky state reserve (East Kamchatka). Transactions of St. Petersburg Botanical Institute of the Russian Academy of Sciences. Issue 16. Saint Petersburg 1994. 232 p.
5. Shamshin, V.A. 1999. Erman's birch forests of Kamchatka: biology, ecology, structure of stands. GEOS, Moscow.