Land Cover Classification from Hyperspectral Data in the Water Ecological Space of Miyun Reservoir

With the acceleration of China’s urbanization process, the problem of the structure and function of water ecological space has become increasingly severe. Monitoring the detailed distribution of land cover types in the key water ecological space is critical for their health assessment and future ecological planning. This study investigated a hybrid feature selection algorithm and GF-5 hyperspectral data (with a spatial resolution of 30 m) to generate a fine land cover classification method for the water ecological space of Miyun Reservoir in Beijing. Firstly, the feature importance ranking was determined using the Random Forest (RF) algorithm and several feature subsets were generated with feature amount gradually carried out in a step size of 10. Then, the classification model was generated based on each subset using the RF algorithm. The feature subset that achieved the highest overall classification accuracy was determined as the initial feature subset. Next, the backward sequential selection algorithm was used to this initial subset to search for the best feature subset. Finally, the classification model of the water ecological space of Miyun Reservoir was generated based on the best feature subset and RF algorithm. To validate the advance of GF-5 hyperspectral data, this study also developed a classification model using Sentinel-2 multispectral data (with a spatial resolution of 10 m) for comparison. The results indicated that hyperspectral data achieved high classification accuracy (overall classification accuracy of 93.61%, and Kappa coefficient of 91.71%), especially in the accurate recognition of tree species (The producer's accuracy and user's accuracy of the chestnut forest are 81.25% and 73.03%, respectively). The reflectance of shortwave infrared bands of GF-5 data has increased the differentiation between chestnut forests and other tree species. By contrast, Sentinel-2 data-based model achieved lower classification accuracy with an overall accuracy of 85.91% and a Kappa coefficient of 82.00%. This result indicated that although Sentinel-2 data has higher spatial resolution than GF-5 data, it still has difficulty identifying chestnut forests due to a lack of fine band information. The classification algorithm proposed in this study can provide accurate basic data for supporting the rational planning and management of water ecological spaces.