TRANSFER LEARNING BASED LAND COVER CLASSIFICATION OF RELATIVELY CALIBRATED L-BAND DATA IN THE LOWER TAPAJÓS

This study explores transfer learning for land cover classification in the Lower Tapajós region of the Brazilian Amazon using L-band Synthetic Aperture Radar (SAR). We classified SAR images using the Random Forest classifier, trained on labeled pixels from a similar image from another date. Here, we tested results from different normalization techniques, such as mean and variance equalization and linear regression-based normalization. These analyses were conducted using two legends with different levels of detail. For the most detailed legend level (10 classes), classification scenarios using transferred training samples generated very different classifications with similar Global Accuracy, which were lower than the one calculated for the scenario with samples collected within the same image being classified. For the least detailed legend level (4 classes), accuracy values were similar for all classification scenarios, although the agreement between pairs of results remained low.