Classifying Forest Types over a Mountainous Area in Southwest China with Landsat Data Composites and Multiple Environmental Factors

Abstract

Accurate information about forest type and distribution is critical for many scientific applications. It is possible to make a forest type map from the satellite data in a cost effective way. However, forest type mapping over a large and mountainous geographic area is still challenging, due to complex forest type compositions, spectral similarity among various forest types, poor quality images with clouds or cloud shadows and difficulties in managing and processing large amount data. Based on the Google Earth Engine (GEE) cloud platform, a method of forest types mapping using Landsat-8 OLI imagery and multiple environmental factors was developed and tested within Yunnan Province (about 390,000 km2) of China. The proposed approach employed a pixel-based seasonal image compositing method to produce two types of seasonal composite images, i.e., four 7-spectral-band composite images and four 5-VI-band composite images associated in spring, summer, autumn, and winter. Then, single-season feature bands and multi-seasonal feature bands were combined with the feature bands of topography, temperature, and precipitation, respectively, and resulting in 17 feature combinations. Finally, using a random forest (RF) classifier, 17 feature combinations were separately experimented to classify the forest type over the study area. The study area was firstly classified into the forest and the non-forest, and then the forest was sub-classified into five forest types (evergreen needleleaf forest, deciduous needleleaf forest, evergreen broadleaf forest, deciduous broadleaf forest, and mixed forest). The results showed that the pixel-based multi-seasonal median composite can produce a cloud-free image for the entire region and is suitable for forest type mapping. Compared with a single-season composite, a multi-seasonal composite can distinguish different forest types more effectively. The environmental factors also improve the accuracy of forest type mapping. With the ground survey samples as reference values, the classification performance of 17 feature combinations was compared, and the optimal feature combination was found out. For the optimal feature combination, its overall accuracy of the forest/non-forest cover map and the forest type map reached 97.57% (Kappa = 0.950) and 70.30% (Kappa = 0.628), respectively. The proposed approach has demonstrated strong potential of high classification accuracy and convenient calculation when mapping forest types over a national or global scale, and its product of 30 m resolution forest type map is capable of contributing to forest resource management.