Accuracy Results Comparison of Machine Learning Classification for Mangroves on Worldview-2 Satellite Imagery

Mangrove forests are one of the highest carbon stores in tropical forests and the amount is very high compared to the average carbon storage in other types of forests in the world. High-resolution satellite imagery provides the ability to obtain finer details, allowing more accurate identification of different types of vegetation and mangrove structures. Worldview - 2 high resolution image can cover large mapping areas, making it more efficient to monitor changes on a large scale. The location of mangroves with limited access required remote sensing techniques which were relatively more cost and time efficient compared to field survey. Machine learning-based classification methods were compared to determine the most accurate and reliable performance results for mangrove classification. In this study, comparisons were made using accuracy test points to assess the level of accuracy of mangrove and another land cover classification generated from the Iso Cluster, Maximum Likelihood, Random Forest and Support Vector Machine methods. According to this study, the four machine learning displayed overall accuracy above 0.5. For kappa accuracy, only Iso Cluster Classification displayed value slightly below 0.5 (0.49). Four machine learning able to distinguish mangrove remarkably with user and producer accuracy 1.0. Worldview - 2 spatial resolution contribute to this accuracy, which enable classification to perform with 0.5 meter resolution. Random forest outperformed another machine learning classification with kappa accuracy 0.88 even without additional parameter and expected to overcome challenges in mapping mangrove vegetation which often has spectral values similar to other vegetation and located in quite heterogeneous ecosystem environment.