Skoči na glavni sadržaj

Izvorni znanstveni članak

Comparison of GEOBIA classification algorithms based on Worldview-3 imagery in the extraction of coastal coniferous forest

Lovre Panđa orcid id ; Sveučilište u Zadru, Odjel za geografiju
Ante Šiljeg orcid id ; Sveučilište u Zadru, Odjel za geografiju
Ivan Marić orcid id ; Sveučilište u Zadru, Odjel za geografiju
Fran Domazetović ; Sveučilište u Zadru, Odjel za geografiju
Silvija Šiljeg orcid id ; Sveučilište u Zadru, Odjel za geografiju
Rina Milošević orcid id ; Sveučilište u Zadru, Ured za znanost, projekte i transfer tehnologija

Puni tekst: hrvatski pdf 2.583 Kb

str. 535-544

preuzimanja: 156



With their ecological, economic, aesthetic, and social function, coniferous forests represent an important part of European forest communities. The main objective of this paper is to compare the most used GEOBIA (Geographic Object-Based Image Analysis) classification algorithms (Random Trees - RT, Maximum Likelihood - ML, Support Vector Machine - SVM) for the purposes of the coastal coniferous forest detection on a high-resolution WorldView-3 (WV-3) imagery on the topographic basin of the Split settlement (Figure 1). The methodological framework (Figure 2) includes: (1)
derivation of a sharpened multispectral image (WV-3MS) (Figure 3); (2) testing of the user-defined parameters in segmentation process (Figure 4); (3) marking of test samples (signatures); (4) classification of a segmented model; (5) accuracy assessment of the classification algorithms, and (6) accuracy assessment of the final model. The developed ACP tool (Automated Classification Process) (Supplement figure 5) for speeding up the entire classification process, enabled the simultaneous generation of output results for three selected classification algorithms (RT, ML and SVM) (Figure 6). Metric indicators (correctness - COR, completeness - COM, and overall quality - OQ) have shown that RT is the most accurate classification algorithm for the coastal coniferous forest detection (Table 1; Figure 7). The iterative setting of segmentation parameters enabled the detection of the most optimal values &8203;&8203;for generating a segmentation model. It is found that shadows can cause significant problems if classification is done on high-resolution images (Figure 8). The solution may be to collect a larger number of samples in different areas for the purpose of more detailed class differentiation. The modified Cohen’s kappa coefficient (K) indicator shown the accuracy of the final model of 87.38% (Table 2; Figure 9). WV-3MS can be considered as very good data for the detection of coniferous forests using the GEOBIA method (Figure 10). According to this research, 31.36% of the Split topographic basin is covered by highly and extremely flammable vegetation.

Ključne riječi

GEOBIA, WorldView-3, Coniferous Forest, Random Trees, Maximum Likelihood, Support Vector Machine

Hrčak ID:



Podaci na drugim jezicima: hrvatski

Posjeta: 802 *