Evaluation of polarization modes and spatial resolution of PALSAR-2 for canal detection accuracy in Indonesia peatlands

Drainage canal inventories are vital for understanding peatland hydrology and mitigating CO₂ emissions from drained peatlands. This study presents a deep learning approach to develop an ensemble classifier for canal segmentation using PALSAR-2, incorporating hyperparameter tuning, multi-polarization configurations, and input resolution assessment. The U-Net architecture combined with EfficientNet backbone yielded superior performance, and HH-HV polarization emerged as a scalable solution for large-scale mapping. The ensemble classifier further enhanced canal detection, achieving a test F1-score of 71.51%. Multi-resolution analysis revealed that while the 6-meter resolution provided the highest accuracy, the 10-meter resolution demonstrated only a marginal reduction, making it a promising alternative for national-level mapping due to lower computational demands. In contrast, the 25-meter resolution showed significant limitations in detecting fine-scale canal features. These findings highlight the importance of polarization selection and model design for accurate canal segmentation, while also revealing resolution–accuracy trade-offs relevant to peatland management efforts.