In recent years, sequence prediction, particularly in natural language processing tasks, has made significant progress due to advanced neural network architectures like Transformer and enhanced computing power. However, challenges persist in modeling and analyzing certain types of sequence data, such as human daily activities and competitive ball games. These segmented sequence data are characterized by short length, varying local dependencies, and coarse-grained unit states. These characteristics limit the effectiveness of conventional probabilistic graphical models and attention-based or recurrent neural networks in modeling and analyzing segmented sequence data. To address this gap, we introduce a novel generative model for segmented sequences, employing an ensemble of multiple variable-order Markov models (VOMMs) to flexibly represent state transition dependencies. Our approach integrates probabilistic graphical models with neural networks, surpassing the representation capabilities of single high-order or variable-order Markov models. Compared to end-to-end deep learning models, our method offers improved interpretability and reduces overfitting in short segments. We demonstrate the efficacy of our proposed method in two tasks: predicting tennis shot types and forecasting daily action sequences. These applications highlight the broad applicability of our segmented sequence modeling approach across diverse domains.

High-resolution (HR) digital bathymetric models (DBMs) are essential for marine science studies. Currently, there are very few areas where detailed seafloor surveys have been conducted, which is insufficient to provide insight into the geological and oceanographic processes of our planet. In recent years, HR DBMs have been successfully reconstructed from low-resolution (LR) DBMs using deep learning. However, HR DBM reconstruction has high uncertainties due to the complexity of seafloor topography. Meanwhile, deep learning-based methods do not provide uncertainty in the HR DBM reconstruction results. Here, based on a deep convolutional neural network with skip connection and network in network (DCSCN) we propose a modified model, DCSCN-DBM, to reconstruct HR DBMs from LR DBMs and evaluate reconstruction uncertainties using the dropout technique. DCSCN-DBM is a stacked convolutional layer in the original DCSCN replaced by several residual blocks, which effectively improves the performance of the network. However, the structure of network in network (NIN) in them is not favorable for reconstruction. Furthermore, dropout with appropriate probabilities at the right positions is also found to be not only beneficial or non-influential for deep learning-based HR DBM reconstruction but also to provide reliable reconstruction uncertainty. In regions with large errors, the DCSCN-DBM without NIN can provide more reliable uncertainty and accurate reconstruction. These results can offer HR DBMs with confidence level reconstructions for other areas where DBMs are needed to better address downstream tasks, such as habitat mapping, climate modeling, hazard prediction, resource assessment and many other field applications.

Quantitative motion analysis to evaluate the performance of athletes has been actively studied recently. Although various methods based on wearable inertial sensors have been developed for simple and repetitive movements recognition, the understanding of continuous complex movements of in-field sports is still challenging. In this paper, we propose a new motion segmentation and recognition method based on foot swing trajectory state to achieve robust and efficient recognition of motion of interest (MOI) in the lower limbs from continuous and complex movements. In order to segment complex movements in the lower limbs, a series of foot motion states are defined based on foot-ground contact status and foot trajectory during swing. The lower body motion state sequence combining the states of both feet is matched to a prior knowledge of MOI cycle sequences obtained in advance, so as to obtain a motion type candidate set. In this case, the continuous movement is segmented based on the prescreened motion types to realize adaptive time window for feature extraction. Finally, according to the prescreened motion type candidate set, corresponding trained neural network binary classifiers are used to make the classification based on the calculated kinematic features. The proposed method is verified through experiments of football movements consisting of walking, dribbling and stepover. As the result, the motion type recognition accuracy is 95%.