MAPPING BENTHIC HABITAT FROM WORLDVIEW-3 IMAGE USING RANDOM FOREST CASE STUDY: NUSA LEMBONGAN, BALI, INDONESIA

Abstract

Benthic habitats are coastal ecosystems that provide many benefits and play an important role in the diversity of nature. The maps are developed using random forest method on the Worldview-3 image. Optically shallow water around Nusa Lembongan was selected as the study area. Sunglint and water column correction were applied to surface reflectance data to produce deglint, depth invariant index, and deglint-depth invariant index band for random forest classification. In addition, tuning parameters, including the number of trees and the function to determine the number of randomly selected, are used in the classification. The benthic habitats classification scheme was constructed based on the variations of in situ data, which consisted of coral reefs, seagrass, macroalgae, and substrate. The confusion matrix was used to analyze the accuracy, and the McNemar test to evaluate the level of statistical significance between different processing scenarios. The best benthic habitats map is determined based on the accuracy and spatial distribution of the object. Meanwhile, the random forest algorithm produced 62.72% - 73.00% overall accuracy and these accuracy variations were not statistically significant. According to the findings, surface reflectance data with the parameter setting comprising 500 trees and square root function yielded the best random forest scenario for mapping benthic ecosystems.