Background and objectives: The Zagros forests, one of Iran's natural ecosystems, possess significant ecological diversity, and understanding their typology is crucial for sustainable management. The forests of Flard County in Chaharmahal and Bakhtiari Province, primarily covered by oak (Quercus brantii) and pistachio (Pistacia mutica), are part of the central Zagros region. These forests play a prominent role in ecological sustainability, biodiversity conservation, and providing environmental services. However, the lack of current and accurate typological maps has hindered their scientific management. Traditional ground-based survey methods are inefficient due to high costs and difficult terrain. This study aimed to use modern remote sensing technologies and evaluate how spectral indices derived from a spectral library improve the accuracy of supervised classification of Sentinel-2 satellite imagery for discriminating forest types.
Materials and methods: To investigate the impact of spectral indices derived from the spectral library on distinguishing the region's dominant species, this study was conducted in three forest areas: two pure types and one mixed type. These areas were Mishan (pure oak), Kate-Siah (pure pistachio), and Vandun (mixed oak-pistachio cover) in Flard County, Chaharmahal and Bakhtiari Province. Three different classification scenarios were used. In Scenario 1, only Sentinel-2 images from October 15, 2024, were classified for all three study areas. In Scenario 2, in addition to the sensor's main bands, general spectral indices—NDVI, SAVI, MSAVI, GNDVI, and EVI—were applied to the Sentinel-2 data and incorporated into the classification process. These indices, respectively, assess greenness, reduce soil background effects, measure chlorophyll health, and enhance sensitivity to dense canopies. In Scenario 3, besides the main bands, specialized spectral indices—PRI, CRI, MCRI, CAI, NDVI800, NPCI, RENDVI, and ARI—extracted from the spectral library data of the oak and pistachio species were included in the classification process. Data classification in all three areas was performed using three supervised algorithms: Support Vector Machine (SVM), Maximum Likelihood Classification (MLC), and Minimum Distance (MinDist).
Accuracy was assessed using a ground truth map, the Kappa coefficient, and overall accuracy criteria.
Results: Results indicated that in Scenario 1, the Maximum Likelihood Classification (MLC) algorithm performed best with overall accuracies and Kappa coefficients of 98.4% and 0.71 in Mishan, 77.5% and 0.32 in Kate-Siah, and 86.7% and 0.67 in Vandun, respectively. In Scenario 2, the same algorithm was best in Mishan (97.7% overall accuracy, 0.65 Kappa coefficient) and Kate-Siah (75.9% overall accuracy, 0.38 Kappa coefficient). However, in Vandun, the SVM algorithm was superior, achieving an overall accuracy of 77.9% and a Kappa coefficient of 0.57. In Scenario 3, accuracy assessment results demonstrated that the SVM algorithm reported the highest accuracy in all three regions: Mishan (99.2% overall accuracy, 0.81 Kappa coefficient), Kate-Siah (98.2% overall accuracy, 0.74 Kappa coefficient), and Vandun (89.0% overall accuracy, 0.67 Kappa coefficient). A comparison of the three scenarios showed that specialized spectral indices derived from the spectral library were more effective in increasing classification accuracy than general indices. Despite the high accuracy of the algorithms, limitations such as spectral similarity between soil and vegetation in low-density areas affected the results.
Conclusion: Based on these findings, integrating valid spectral data with machine learning algorithms is an effective tool for preparing detailed forest cover maps. Using a spectral library alongside Sentinel-2 imagery enables more accurate discrimination of forest. The SVM algorithm is a suitable option for classifying forest types in areas with dense cover and complex structures. The findings of this research are applicable to conservation-management planning and the continuous monitoring of Zagros Forest ecosystems.
