<![CDATA[This study evaluates and compares methods for extracting coastlines in the wetland coastal areas of northern East Java, specifically in Sidoarjo and Gresik Regencies, East Java, Indonesia. By analyzing eight scenarios that integrate bands from Landsat 8 imagery, the Modified Normalized Difference Water Index (MNDWI), and the SRTM Digital Elevation Model (DEM) under two machine learning classifiers (Random Forest and Classification and Regression Trees) on the Google Earth Engine (GEE) platform. Using a dataset consisting of 180 ground-truth points, we found that Random Forest (RF) combined with MNDWI and DEM (Scenario 5) achieved the highest overall accuracy (OA = 76%, Kappa = 0.71) in the flat wetlands of Sidoarjo, which are dominated by aquaculture ponds and mangrove vegetation. In contrast to the diverse Gresik coastal area, where some regions are dominated by industry, the pure CART model (Scenario 6) yielded the best results (OA = 66%, Kappa = 0.60), as the addition of DEM and MNDWI caused overfitting due to extreme local spectral and vertical disturbances from the complex coastal structure. These results indicate that ensemble-based classification (RF) with elevation filtering is crucial for distinguishing the microtopography of aquaculture ponds from natural seawater, whereas a single decision tree (CART) performs better with raw spectral inputs in structurally heterogeneous coastal environments.]]>
Copyrights © 2026
