This study tested the ability of Sentinel-1 C-band to separate forest from other common land use classes (i.e., urban, low vegetation and water) at two different sites. The first site is characterized by temperate forests and rough terrain while the second by tropical forest and near-flat terrain. We trained a support vector machine classifier using increasing feature sets starting from annual backscatter statistics (average, standard deviation) and adding long-term coherence (i.e., coherence estimate for two acquisitions with a large time difference), as well as short-term (six to twelve days) coherence statistics from annual time series. Classification accuracies using all feature sets was high (>92% overall accuracy). For temperate forests the overall accuracy improved by up to 5% when coherence features were added: long-term coherence reduced misclassification of forest as urban, whereas short-term coherence statistics reduced the misclassification of low vegetation as forest. Classification accuracy for tropical forests showed little differences across feature sets, as the annual backscatter statistics sufficed to separate forest from low vegetation, the other dominant land cover. Our results show the importance of coherence for forest classification over rough terrain, where forest omission error was reduced up to 11%.
Evaluating tropical forest classification and field sampling stratification from lidar to reduce effort and enable landscape monitoring
