Combining GF-2 and RapidEye satellite data for mapping mangrove species using ensemble machine-learning methods

ABSTRACT

Mangrove forests are important constitutions for sustainable development of coastal ecosystems, and they are often mapped and monitored with remote sensing approaches. Satellite images allow detailed studies of the distribution and composition of mangrove forests, and therefore facilitate the management and conservation of the ecosystems. The combination of multiple types of satellite images with different spatial and spectral resolutions is helpful in mangrove forests extraction and mangrove species discrimination as it reduces sampling workload and increases classification accuracies. In this study, the 1.0-m-resolution Gaofen-2 (GF-2) and the 5.0-m-resolution RapidEye-4 (RE-4) satellite images, acquired in February 2017 and November 2016 respectively, were used with ensemble machine-learning and object-oriented methods for mangroves mapping at both the community and species levels of the Qi’ao Island, Zhuhai, China. First, the mangroves on the island were segmented from the GF-2 image on a large scale, and then they were extracted combining with their digital elevation model (DEM) data. Second, the GF-2 image was further processed on a fine scale, in which object-oriented features from both the GF-2 and RE-4 images were extracted for each mangrove species. Third, it is followed by the mangrove species classification process which involves three ensemble machine-learning methods: the adaptive boosting (AdaBoost), the random forest (RF) and the rotation forest (RoF). These three methods employed a classification and regression tree (CART) as the base classifier. The results show that the overall accuracy (OA) of mangrove area extraction on the Qi’ao Island with the auxiliary data, DEM, achieves 98.76% (Kappa coefficient (κ) = 0.9289). The features extracted by the GF-2 and RE-4 images were shown to be beneficial for mangrove species discrimination. A maximum improvement in the OA of approximately 8% and a κκ of approximately 0.10 were achieved when employing RoF (OA = 92.01%, κ = 0.9016). Ensemble-learning methods can significantly improve the classification accuracy of CART, and the use of a bagging scheme (RF and RoF) is shown as a better way to map mangrove species than adaptive boosting (AdaBoost). In addition, RoF performed well in mangrove species classification but it was not as robust as the RF, whose average OA and κκ were 80.59% and 0.7608, respectively, while the RoF’s were 77.45% and 0.7214, respectively, in the 10-fold cross-validation.