Outdoor Localization for LoRaWans Using Semi-Supervised Transfer Learning with Grid Segmentation

The existing outdoor localization schemes are either not accurate enough or too costly. To improve the existing outdoor localization schemes, a novel outdoor localization scheme using a semi-supervised transfer learning for LoRaWANs is proposed in this paper. Usually, it is hard to collect the complete dataset of labeled samples in an outdoor environment for localization using deep learning and hence the semi-supervised transfer learning is adopted. We have proposed a novel grid segmentation scheme to generate a number of virtual labeled samples by figuring out the relationship of labeled and unlabeled samples. With the relationship of labeled-unlabeled-samples, we may repeatedly fine-tune our target model by adding more new virtual labeled samples so as to further enhance the localization accuracy. The experimental results justify the accuracy of the proposed scheme.