Accuracy Of High Spatial Resolution Satellite Images Classification For Forestry Needs

Osnovni cilj ovog rada je istražiti, usporediti i pronaći najbolji način interpretacije IKONOS satelitske snimke visoke rezolucije, koji će biti jednostavan i prihvatljiv za operativnu primjenu. Glavni materijal za izradu ovoga rada bila je satelitska snimka istraživanog područja visoke prostorne rezolucije IKONOS. Provedene su vizualna i digitalna interpretacija satelitske snimke te su ocijenjene točnosti klasifikacija. Vizualna interpretacija provedena je na izrađenom kolorkompozitu s odabrana tri spektralna kanala (4, 3, 2), dok je digitalna interpretacija provedena kroz šest različitih algoritama. Za svaki algoritam nadgledane klasifikacije i za vizualnu interpretaciju, izračunati su parametri točnosti same klasifikacije i Kappa statistika. Najbolji rezultati svih provedenih interpretacija (klasifikacija), dobiveni su vizualnom interpretacijom, odnosno tom interpretacijom dobiveni su rezultati koji najbolje odgovaraju stanju na terenu (terenska izmjera, podaci iz Osnove gospodarenja). Provedenim metodama nadgledane klasifikacije satelitske snimke IKONOS utvrđeno je da se najtočnijom metodom za operativnu primjenu pokazala klasifikacija pomoću prepoznavanja značajki (Feature extraction) s modulom prepoznavanja prirodnih oblika (natural features), gdje se neklasificirana područja pridružuju najsličnijoj klasi.Satellite images as a source of information are becoming more significant and more often used in Croatian forestry. Data in satellite images can be obtained in two ways – by visual and digital interpretation. Choice of adequate interpretation is dependant on set goals or, more precisely, on getting enough quality information needed for certain task. Main objective of this paper was to examine, compare and find the best way to interpret IKONOS high spatial resolution satellite images, that would be simple and acceptable for operational purposes. Research was conducted in Spačva forest basin area, where largest integral pedunculate oak (Quercus robur L) forest is situated. Interest array was defined with 1 IKONOS satellite image scene (11.3 x 11.3 km) and it covers the central part of Spačva forest complex, with four forest management units: Kragujna, Otočke šume, Slavir, Vrbanjske šume. The above mentioned shooting produced IKONOS satellite image of the Spačva basin area, 132 km2 large surface, in 5 spectral channels: PAN (1 x 1 m) and 4 MS Bundle (4 x 4 m). Visual interpretation was conducted on the created colorcomposit with three chosen spectral channels (4, 3, 2), while digital interpretation was conducted through six different algorithms, on the IKONOS satellite image. For each algorithm of the supervised classification, as well as the visual interpretation, Kappa statistics and accuracy parameters of the classification were calculated. Reference data from field research was compared to results from different satellite image classification methods to determine the level of accuracy for each classification. From these comparisons, the error matrix, which represents the base for accuracy verification, was created. Three accuracy indexes were derived from the error matrix: total, producers and users index. Last two refer to each class individually. Apart from the mentioned indexes, error matrix was also used for calculating the parameters of Kappa statistics, which enabled data generalization. Of all the conducted interpretations (classifications), visual interpretation produced best results – results obtained this way were the closest to the actual situation in the field (field research, data from forest management plans). Conducted methods of the IKONOS satellite image supervised classification determined that the most accurate method for operational application was the feature extraction classification with natural features recognition module, in which the unclassified areas are assigned to the most similar class. This research also confirmed the facts stated by Mas and Ramirez (1996) in their studies, concerning the claims that, besides being the most used, visual interpretation method also produced the most accurate results, primarily due to the human ability to identify objects/events of interest. On the other hand, process itself lasts relatively long, because each part of the image is analysed separately, which can significantly increase the time and the costs of the analysis