Abstract Accurate localisation in wireless networks is essential, especially in environments with low signal-to-noise ratios. Reconfigurable intelligent surfaces (RIS) have recently emerged as an effective approach to enhance localisation by controlling signal reflections toward target terminals. This study evaluates a hybrid localisation framework combining classical machine learning (ML) and deep learning (DL) models, applied to both engineered features and raw time-series data. Time-domain features—including mean, standard deviation, minimum, and maximum over short temporal windows—were extracted and reduced via PCA for ML models (Decision Tree, Random Forest, SVM, KNN, Logistic Regression, MLP, XGBoost), while DL models (DNN, CNN1D, RNN, LSTM) directly processed sequences of received signal strength and channel phase measurements. Experiments with 1 m inter-antenna spacing demonstrate that, with RIS activated, ML models leveraging feature extraction achieve high performance, with MLP and SVM (RBF) reaching F1-scores of 0.877 and 0.875, respectively. DL models on raw signals also perform competitively, with CNN1D achieving a macro F1 of 0.768. Without RIS, performance decreases substantially: ML models reach F1-scores below 0.57, while DL models achieve up to 0.668. These results indicate that feature extraction strongly benefits classical ML models, DL can effectively exploit raw sequences, and RIS activation significantly enhances localisation accuracy. The study demonstrates that combining feature engineering, deep learning, and RIS-assisted channels provides a robust framework for precise localisation in challenging wireless scenarios, outperforming conventional approaches across multiple metrics.

Training and test dataset samples are provided.