Quartz is an important mineral in many metal deposits and can provide important indications about the deposit's origin through its chemical composition. However, traditional low-dimensional analysis methods are ineffective in utilizing quartz's chemical composition to reveal the deposit's origin type. In this study, 1140 quartz samples from eight geological environments were collected, and three machine learning (ML) models—random forest, eXtremely Greedy tree Boosting (XGBoost), and light gradient boosting machine (LightGBM) were used to classify quartz deposits. The application of the Shapley Additive Explanation (SHAP) algorithm and Spearman correlation analysis is utilized to interpret the predictive results of the model and analyze feature correlations, aiming to enhance the credibility of the classification results and discover underlying patterns. Finally, a visualization method based on XGBoost and t-SNE was proposed. By calculating SHAP values, the key geochemical indicators that differentiate each type of quartz deposit were determined. Furthermore, the impact of varying concentrations of different trace elements on the identification of quartz deposits was analyzed. This study demonstrated the effectiveness of using machine-learning algorithms based on trace elements to classify quartz and provided new insights into the relationships between trace elements and quartz genesis, as well as the effects of different trace element combinations and concentrations on quartz identification.
Traffic flow prediction is a challenging spatiotemporal prediction task due to its spatiotemporal dynamics and uncertainty. In recent years, graph convolutional neural networks (GCNs) have been applied to traffic flow prediction due to their exceptional capabilities in processing network information. However, current GCN-based methods fail to capture long-range spatial dependencies between regions with similar functionalities, and the modeling of spatial dependencies within adjacent time steps is neglected. To solve these problems, this work proposes a new dynamic graph convolutional recurrent network (CDGCRN) with the spatiotemporal category information embedding. An embedding layer with spatiotemporal category information is introduced to enhance the model’s capability in capturing long-range spatial dependencies. This layer aids the model in rapidly identifying traffic patterns with distinct features. Morever, Dynamic spatial information fusion graph convolutional recurrent unit is proposed to capture spatial dependencies within all time steps and adjacent time steps. It incorporates a learnable graph structure with a time window for capturing spatial dependency relationships within adjacent time steps and a globally learnable graph structure for learning spatial dependency relationships across all time steps. Experimental results on the four real-world data sets demonstrate that CDGCRN is not only computationally efficient but also outperforms 17 state-of-the-art baseline methods.
22 January 2024 in IEEE Geoscience and Remote Sensing Letters
Sinkholes are natural depressions or cavities on the Earth’s surface, and accurate detection of sinkholes can prevent them from posing significant risks to human life, infrastructure, and the environment. Due to their varied and complex surface manifestations, sinkholes are relatively rare and pose challenges in precise detection when compared to other geological features. To achieve precise sinkhole detection, this paper introduces a network with multi-scale feature fusion and enhancement capabilities (AMFENet). AMFNet is the first to propose considering the irregular shapes, varying sizes, relatively small target areas, and uneven distribution characteristics of sinkholes. To capture the contextual global information of sinkholes, AMFENet utilizes the Swin Transformer as its encoder. To address the scarcity and uneven distribution of sinkholes, a Selective Amplification Block (SAB) is introduced, enhancing the significance of sinkhole features. To handle complex and diverse sinkhole features, the Semantic Integration Block (SIB) is introduced. It integrates both global and local semantic information and adaptively enhances crucial semantic features from both spatial and channel dimensions. According to extensive experiments conducted on the sinkhole DEM dataset in Kentucky, AMFNet outperforms other semantic segmentation methods.
28 October 2025 in IEEE Transactions on Neural Networks and Learning Systems
The spatiotemporal dynamics of traffic forecasting make it a challenging task. In recent years, by adapting to the topology of traffic networks where road segments serve as nodes, graph convolutional networks (GCNs) have been able to capture spatiotemporal dependencies, thereby improving traffic forecasting performance. However, there are two shortcomings of GCN-based methods: 1) existing methods treat the delays between nodes in the traffic network as equally important and fail to extract critical information effectively, leading to information redundancy, the introduction of irrelevant noise, and increased computational costs and 2) most methods overlook the issue that spatiotemporal correlations between nodes are inconsistent across different timescales. This article designs a new dynamic delay-aware multiscal spatiotemporal graph convolutional network (DDAMGCN) for traffic forecasting. Specifically, a dynamic delay-aware module is designed to identify key nodes and model the important delays from key nodes, so that the model focuses on key information and reduces computational cost. Additionally, a novel multiscale spatiotemporal graph convolution module is designed to achieve fine-grained modeling of the spatiotemporal correlation of different nodes at different timescales. Experiments on eight real traffic datasets verify the superiority of the proposed method compared to several state-of-the-art baselines.
As urbanization accelerates, the management of urban solid waste poses increasingly intricate challenges. Traditional urban metrics, such as GDP and per capita consumption rates, have become inadequate for accurately reflecting the realities of waste generation; moreover, the linear correlation between these metrics and waste production is progressively diminishing. Consequently, this study introduces a novel methodology leveraging nighttime light remote sensing data to enhance the precision of urban solid waste production forecasts. By processing remote sensing data to mitigate noise and integrating it with conventional urban datasets, an innovative index system and predictive model were developed. Using Beijing as a case study, the gradient boosting regression algorithm yielded a prediction accuracy of 92%. Furthermore, in light of the substantial costs associated with waste recovery route planning and site selection for treatment facilities, this research further devised a location and distribution framework for waste treatment centers based on high-precision predictions of waste production while employing multi-objective evolutionary algorithms (MOEAs) alongside the non-dominated sorting genetic algorithm II (NSGA-II) for optimization. Distinct from prior studies, this study suggests that service point waste quantities are not fixed values but rather adhere to a normal distribution within specified ranges and thus provides a more realistic simulation of fluctuations in waste production while enhancing both the robustness and predictive accuracy of the model. In conclusion, by incorporating nighttime light remote sensing data along with advanced machine learning techniques, this study markedly improves forecasting accuracy for waste production while offering effective optimization strategies for site selection and recovery route planning—thereby establishing a robust data foundation aimed at refining urban solid waste management systems.