Random forest classification and regression for seagrass mapping using PlanetScope image in Labuan Bajo, East Nusa Tenggara

Abstract

Random forest is a machine learning algorithm that can be used to improve the classification accuracy of mapping using remote sensing, especially for seagrass mapping in a complex optically water shallow. This research is aimed to map seagrass species composition and percent cover using random forest classification and regression using PlanetScope image. Optically shallow water around Labuan Bajo was selected as the study area. Sunglint and water column corrections were applied to the surface reflectance image. Principle Component Analysis (PCA) transformation was applied on surface reflectance bands, deDeglint bands, and depth-invariant index bands. These bands were used as the input band for random forest classification and regression algorithm, using field data to train the algorithm. Benthic field data was collected by the photo transect and seagrass field data was collected by the photo quadrat transect technique. Benthic habitat classification scheme was constructed based on the variation of benthic habitat insitu, which consisted of coral reefs, seagrass, macroalgae, and bare substratum. Seagrass species composition classification scheme was constructed following the variation of seagrass species insitu, which consisted Enhalus acaroides (Ea), Enhalus acaroides mixed Syringodium isoetilolium (EaSi), Enhalus acaroides mixed Thalassia hemprichii (EaTh), Halodule uninervis (Hu), Mixed species class, Thalassodendron ciliatum (Tc), Thalassodendron ciliatum mixed Enhalus acaroides (TcEa), Thalassia hemprichii (Th), Thalassia hemprichii mixed Cymodocea rotundata (ThCr), and Thalassia hemprichii mixed Syringodium isoetilolium (ThSi) class. Accuracy assessment using independent field data showed that random forest algorithm produced 63.57%- 72.09% overall accuracy for benthic habitat and 83.52%-85.71% overall accuracy for seagrass species composition. Random forest regression for seagrass percent cover produced R 2 between 0.78-0.81 with the error of prediction between 14.59-15.26.