<![CDATA[Land cover classification from UAV imagery has become increasingly important for environmental monitoring, especially in peatland ecosystems where vegetation density is closely related to land openness, degradation risk, and fire vulnerability. This study proposes a hybrid approach that combines CNN-based feature extraction with machine learning classification for vegetation density-based land cover classification. A total of 3,000 UAV images collected from Block 1 of the Liang Anggang Protected Forest, Banjarbaru, were used in this study and categorized into three classes: bare, moderate, and high vegetation density. The images were preprocessed through cropping, resizing, and labeling prior to feature extraction. ResNet-50 and DenseNet-121 were employed as feature extractors, while ten machine learning classifiers were evaluated, namely CalibratedClassifierCV, SVC, NuSVC, LogisticRegression, PassiveAggressiveClassifier, SGDClassifier, LinearSVC, XGBClassifier, Perceptron, and LGBMClassifier. The results show that ResNet-50 generally outperformed DenseNet-121 as a feature extractor. The best and most balanced performance was achieved by the ResNet-50 + SVC combination, which obtained 84% accuracy, 84% F1-score, 91% precision, 77% recall, and a computation time of 8.58 minutes. Although CalibratedClassifierCV achieved the same accuracy, it required substantially longer processing time. These findings indicate that classification performance is influenced not only by the classifier used, but also by the compatibility between feature representation and classification mechanism. Therefore, the combination of ResNet-50 and SVC is recommended for UAV-based vegetation density land cover classification in peatlands.]]>
Copyrights © 2026
