SAR and optical time series for crop classification

In this study, we compare the classification performance of time series with single polarized SAR data, optical data and fused optical and SAR data through the Gram-Schmidt transform. Different machine learning algorithms for crop classification were applied. Specifically Gradient Boosting Trees (GBT) yielded good results as they are known to perform well with imbalanced feature labels and datasets with different probability density functions. Surprisingly, the speckle-filtered time series SAR data performed worst, whereas the fused dataset yielded the highest overall accuracy score. Results showed that post-processing of the SAR classification improves accuracy scores.

Перевод пока недоступен