Improving the Accuracy of Vegetation Classifications in Mountainous Areas A Case Study in the Spanish Central Pyrenees

In recent decades, mountainous areas that contain some of the best-preserved habitats worldwide are experiencing significant, rapid changes. Efficient monitoring of these areas is crucial for impact assessments, understanding the key processes underlying the changes, and development of measures that mitigate degradation. Remote sensing is an efficient, cost-effective means of monitoring landscapes. One of the main challenges in the development of remote sensing techniques is improving classification accuracy, which is complicated in mountainous areas because of the rugged topography. This study evaluated the 3 main steps in the supervised vegetation classification of a mountainous area in the Spanish Pyrenees using Landsat-5 Thematic Mapper imagery. The steps were (1) choosing the training data sampling type (expert supervised or random selection), (2) deciding whether to include ancillary data, and (3) selecting a classification algorithm. The combination (in order of importance) of randomly selected training data, ancillary data (topographic and vegetation index), and a random forest classifier improved classification accuracy significantly (4-11%) in the study area in the Spanish Pyrenees. The classification procedure includes important steps that improve classification accuracies; these are often ignored in standard vegetation classification protocols. Improved accuracy is vital to the study of landscape changes in highly sensitive mountain ecosystems.