Performance of shallow and deep learning models in classifying mechanized pit drilling events based on GNSS input data

Decision-making on spatial and temporal machine allocation is typically supported by dedicated time-and-motion studies, and long-term monitoring of forest operations is a promising way to improve our understanding of the performance of the machines used. Nevertheless, the level of resources and expertise required to conduct long-term studies often prevents such efforts aimed at collecting, processing and making use of the data. Most of the currently used machines come equipped with Global Navigation Satellite System (GNSS) dataloggers, which provide georeferenced, time-stamped position records, enabling digital data transfer and processing. Coupling deep learning with event-based labelled GNSS datasets may be a suitable tool for gathering information and automating the prediction effort on very large datasets, thereby offering access to long-term data for operational planning and machine allocation. To the best of our knowledge, this work pioneers the use of sequential deep learning on GNSS data for six-class event recognition in mechanized pit drilling operations. Our primary original contributions include: formulating a novel operational problem, curating one of the largest annotated GNSS datasets for this task, and providing a rigorous performance comparison of Temporal Convolutional Network (TCN), Long Short-Term Memory Network (LSTM), and Gated Recurrent Unit network (GRU) models using comparable hyperparameter-tuning and data-splitting protocols. The goal of this study was to evaluate the performance of sequential deep learning architectures in predicting operational events by considering six event classes that are relevant for mechanized pit-drilling operations using as a benchmark the performance of five classical shallow models. The study was based on a large annotated GNSS dataset containing >127,000 observations collected from two machines at a rate of 1 Hz. The eight models were each evaluated using a randomized hyperparameter search across 100 trials. The optimization process considered lookback windows ranging from 10 to 50 time steps and batch sizes of 16, 32, and 64 for the sequential deep learning architectures. Among all models, Random Forest achieved the highest test macro-averaged F1 score (0.7355), outperforming all three sequential deep learning architectures. Among the deep learning models, GRU yielded the highest test macro-averaged F1 (0.6976), followed by LSTM (macro-F1 = 0.6918) and TCN (macro-F1 = 0.6254). These findings indicate that, within the operational context of this study, well-tuned shallow classifiers - particularly tree-based models - match or exceed sequential deep learning architectures for automated activity recognition in pit drilling operations using GNSS-derived data. The code and data are publicly available at https://github.com/lmsasu/drill_events_classification.