H. Yıldırım, E. Alparslan, B. Bilge, H. Kurar, O. Divan, S. Elitaş
TÜBİTAK-MAM Space Technologies Department
P.O. Box 21 41470 Gebze-Kocaeli / TURKEY
New possibilities for forest mapping using integration of geographic information systems (GIS) and remote sensing (RS) technologies are investigated. RS data processed by image classification methods is subsequently converted into a GIS data coverage, offering the possibility of analyzing the classified image together with other data levels such as digital terrain model (DTM) data existing in GIS. It was possible to analyze each forest type according to the altitude information in GIS, i.e. certain tree types would only grow in certain altitudes. As a result, the information extracted was tabulated.
The study area is located on Black Sea Shore, Turkey, being a small part of Ordu province covering 20 km by 22 km area with 40ø 50' N / 41ø 10' N latitudes and 37ø 44' E and 37ø 58' E longitudes. LANDSAT TM satellite imagery corresponding to the study area, acquired on 15.9.1987 was classified using ground truth information collected from Directorate of Forest Works (DFW) of that area. The classified satellite image was transferred into GIS as a data layer. Then, topographic map sheets of the study area obtained from General Commander of Mapping (GCM) at 1:25.000 scale were digitized at 50 m intervals to extract contours of constant elevation of the study area. Land cover was draped over the surface model generated from contours of constant elevation, creating a 3D perspective image of the area to facilitate further analysis.
Known UTM coordinates collected from map sheets of GCM were used as ground control points in geometrical correction of the satellite image corresponding to the study area in order to obtain an image georeferenced into its proper UTM coordinates. Maximum likelihood classifier was trained by ground truth information collected from DFW of the area and the training samples were validated using a minimum distance classifier. Validation served as some filter excluding the outliers in the training areas. The classifier assumed that each class would have normal (Gaussian) distribution and assigned each pixel to a class having the minimum Mahalanobis distance. Seven different land covers identified in the study area were: sea, alder forest, chestnut - alder, healthy mixed forest, unhealthy mixed forest, urban area and farm land.
The classified image was visually analyzed and some tiny class polygons in the form of pixel clusters were discovered with a center pixel being connected to neighboring pixels in horizontal, vertical and diagonal directions. Since these polygons are surrounded by larger polygons of some other class, eliminating these polygons will yield a smooth surface in the classified image. A sieve filter was used for this purpose. Due to formation of large class regions by sieve filtering, unit boundary vectors become only visible between pixels assigned to different classes .
The land cover map which is in raster data form was converted into vector form and integrated into ARC/INFO polygon coverage. Arcs forming the polygon coverage were analyzed and arcs sharing pseudo nodes were combined by removing these nodes. The distance between adjacent vertices on splined arc and curves within an added arc is set to 42.42 which is the hypotenus of a right triangle having sides of 30 m, which is the pixel size for TM images. Using this tolerance of 42.42 the shape of arcs building the polygon coverage were changed by building curves at the angles in an arc. Arcs were redrawn so that adjacent vertices are spaced by 42.42 tolerance apart. Vertices were added or deleted depending on this tolerance resulting in arc smoothing or arc generalization, respectively. Smoothed arcs were converted to polygon coverage, resulting polygons having no stair effects on their sides. This coverage enabled further analysis of classified image together with other additional data layers such as digital terrain model (DTM) data existing in GIS.
A surface model was generated from elevation contour lines digitized from topographic maps at 50 m intervals by creating a triangulated irregular network (TIN), which is a series of connected triangles accurately representing a surface with less data points than other data models. TIN creation errors were analyzed generating a descriptive listing about the TIN surface followed by its graphic display. Invalid flat triangles occurring along streams and ridges were eliminated adding new intermediate points along the ridges and streams between the input contours. Additional sample points were entered between the contours in order to increase the distance between vertices on each contour arc, resulting in removal of invalid flat triangles. Weed tolerance and proximal tolerance were also adjusted to remove excess vertices forming flat triangles. The created TIN surface was transformed back to contour arc coverage