With the growing demand for location-based services (LBS) from personal mobile devices, indoor localization is facing increasingly strict requirements in terms of accuracy, scalability, and robustness. Current methods don’t fully consider sparse RSS location fingerprints caused by reasons such as wall obstructions, and the high similarity of RSS location fingerprints caused by overlapping WiFi access point coverage, resulting in reduced indoor localization accuracy. Therefore, we propose an Encoder-Decoder driven Adaptive Multiscale-CNN based indoor Localization model (EDAMLoc) to solve the above issues, which is designed for multi-building and multi-floor environments. It utilizes feature enhancement to process the features available in sparse RSS location fingerprints and adaptively fuses cross-scale RSS location fingerprint features to better distinguish similar RSS location fingerprints. Furthermore, EDAMLoc can flexibly adapt its output branches based on the requirements of different localization tasks. Experimental results show that EDAMLoc achieves 100% Building Accuracy and 95.30% Floor Accuracy on the UJIIndoorLoc and Tampere datasets, respectively, and that its localization accuracy is better than that of mainstream indoor localization models.