Spatiotemporal Dependence Research Articles

Long-term wind power forecasting with series decomposition and spatio-temporal graph neural network

ABSTRACT The stable operation of the power grid requires accurate predictions of wind power generation and the stabilization of its fluctuations through the integration of other energy sources. An increasing number of deep learning methods are now being employed in the field. However, due to the instability of wind power data, existing methods struggle to uncover deep spatiotemporal dependencies. We propose a novel method named SD-STGNN (Series Decomposition and Spatio-Temporal Graph Neural Network). SD-STGNN first decomposes unstable wind power data into seasonal and trend components. For the seasonal data reflecting short-term fluctuation patterns, we employ a Gated Temporal Convolutional Network and Graph Convolutional Network to capture spatiotemporal relationships. Additionally, for trend data reflecting long-term fluctuation patterns, we introduce a Temporal-Feature Enhancement module, utilizing Multi-Layer Perceptrons to extract deep information along both temporal and feature dimensions. Extensive experiments were conducted on the SDWPF public dataset. Compared to existing state-of-the-art baseline methods, our proposed SD-STGNN model achieves a notable average reduction in Mean Absolute Error by approximately 6.26%, in Root Mean Squared Error by 7.55%, and in Mean Absolute Percentage Error by 2.65%. Additionally, there is an average improvement of about 4.74% in the coefficient of determination.

Short-Term Origin-Destination Demand Prediction Based on Spatiotemporal Encoder-Decoder Network with a Residual Feature Extractor

Online ride-hailing services play a crucial role in daily transportation, However, challenges persist in certain regions with limited access, and drivers encounter difficulties in receiving orders. Accurate prediction of short-term origin-destination (OD) demand is crucial for addressing these issues. This study leverages recent advancements in artificial intelligence and big data to introduce a spatiotemporal encoder-decoder network with a residual feature extractor (RF-STED) for short-term OD demand prediction in online ride-hailing services. The RF-STED model, built on deep learning models such as graph convolutional networks and convolutional long short-term memory (Conv-LSTM), includes spatiotemporal networks, encoding layers, and a residual feature extractor. The spatiotemporal network has two branches: branch one processes multi-pattern OD data using a multi-pattern temporal feature extraction module, utilizing a multi-channel Conv-LSTM to capture temporal correlations. Branch two utilizes a multi-spatial feature extraction module to convert OD pair associations into a spatial topology, extracting multi-spatial correlations. The encoding layer captures spatiotemporal dependencies, while the residual feature extractor decodes compressed vectors back into an OD graph for forecasting future demand. Experiments with a Manhattan taxi dataset in the U.S. show the RF-STED model outperforms 10 baseline models and four ablation models. The results emphasize the model’s strength and robustness in short-term OD flow prediction.

Scientific logic and spatio-temporal dependence in analyzing extreme-precipitation frequency: negligible or neglected?

Abstract. Statistics is often misused in hydro-climatology, thus causing research to get stuck on unscientific concepts that hinder scientific advances. In particular, neglecting the scientific rationale of statistical inference results in logical and operational fallacies that prevent the discernment of facts, assumptions, and models, thus leading to systematic misinterpretations of the output of data analysis. This study discusses how epistemological principles are not just philosophical concepts but also have very practical effects. To this aim, we focus on the iterated underestimation and misinterpretation of the role of spatio-temporal dependence in statistical analysis of hydro-climatic processes by analyzing the occurrence process of extreme precipitation (P) derived from 100-year daily time series recorded at 1106 worldwide gauges of the Global Historical Climatology Network. The analysis contrasts a model-based approach that is compliant with the well-devised but often neglected logic of statistical inference and a widespread but theoretically problematic test-based approach relying on statistical hypothesis tests applied to unrepeatable hydro-climatic records. The model-based approach highlights the actual impact of spatio-temporal dependence and a finite sample size on statistical inference, resulting in over-dispersed marginal distributions and biased estimates of dependence properties, such as autocorrelation and power spectrum density. These issues also affect the outcome and interpretation of statistical tests for trend detection. Overall, the model-based approach results in a theoretically coherent modeling framework where stationary stochastic processes incorporating the empirical spatio-temporal correlation and its effects provide a faithful description of the occurrence process of extreme P at various spatio-temporal scales. On the other hand, the test-based approach leads to theoretically unsubstantiated results and interpretations, along with logically contradictory conclusions such as the simultaneous equi-dispersion and over-dispersion of extreme P. Therefore, accounting for the effect of dependence in the analysis of the frequency of extreme P has a huge impact that cannot be ignored, and, more importantly, any data analysis can be scientifically meaningful only if it considers the epistemological principles of statistical inference such as the asymmetry between confirmatory and disconfirmatory empiricism, the inverse-probability problem affecting statistical tests, and the difference between assumptions and models.

Spatio-Temporal Parallel Transformer based model for Traffic Prediction

Traffic forecasting problems involve jointly modeling the non-linear spatio-temporal dependencies at different scales. While Graph Neural Network models have been effectively used to capture the non-linear spatial dependencies, capturing the dynamic spatial dependencies between the locations remains a major challenge. The errors in capturing such dependencies propagate in modeling the temporal dependencies between the locations, thereby severely affecting the performance of long-term predictions. While transformer-based mechanisms have been recently proposed for capturing the dynamic spatial dependencies, these methods are susceptible to fluctuations in data brought on by unforeseen events like traffic congestion and accidents. To mitigate these issues we propose an improvised Spatio-temporal parallel transformer (STPT) based model for traffic prediction that uses multiple adjacency graphs passed through a pair of coupled graph transformer-convolution network units, operating in parallel, to generate more noise-resilient embeddings. We conduct extensive experiments on 4 real-world traffic datasets and compare the performance of STPT with several state-of-the-art baselines, in terms of measures like RMSE, MAE, and MAPE. We find that using STPT improves the performance by around \(10-34\%\) as compared to the baselines. We also investigate the applicability of the model on other spatio-temporal data in other domains. We use a covid-19 dataset to predict the number of future occurrences in different regions from a given set of historical occurrences. The results demonstrate the superiority of our model for such datasets.

DPP: A Novel Disease Progression Prediction Method for Ginkgo Leaf Disease Based on Image Sequences

Ginkgo leaf disease poses a grave threat to Ginkgo biloba. The current management of Ginkgo leaf disease lacks precision guidance and intelligent technologies. To provide precision guidance for disease management and to evaluate the effectiveness of the implemented measures, the present study proposes a novel disease progression prediction (DPP) method for Ginkgo leaf blight with a multi-level feature translation architecture and enhanced spatiotemporal attention module (eSTA). The proposed DPP method is capable of capturing key spatiotemporal dependencies of disease symptoms at various feature levels. Experiments demonstrated that the DPP method achieves state-of-the-art prediction performance in disease progression prediction. Compared to the top-performing spatiotemporal predictive learning method (SimVP + TAU), our method significantly reduced the mean absolute error (MAE) by 19.95% and the mean square error (MSE) by 25.35%. Moreover, it achieved a higher structure similarity index measure (SSIM) of 0.970 and superior peak signal-to-noise ratio (PSNR) of 37.746 dB. The proposed method can accurately forecast the progression of Ginkgo leaf blight to a large extent, which is expected to provide valuable insights for precision and intelligent disease management. Additionally, this study presents a novel perspective for the extensive research on plant disease prediction.

Convolutional Non-Homogeneous Poisson Process and its Application to Wildfire Ignition Risk Quantification for Power Delivery Networks

To quantify wildfire ignition risks on power delivery networks, the current practice predominantly relies on the empirically calculated fire danger indices, which may not well capture the effects of dynamically changing environmental factors. This article proposes a spatio-temporal point process model, known as the Convolutional Non-homogeneous Poisson Process (cNHPP), and applies the model to quantify wildfire ignition risks for power delivery networks. The proposed model captures both the current (i.e., instantaneous) and cumulative (i.e., historical) effects of key environmental processes (i.e., covariates) on wildfire risks, as well as the spatio-temporal dependency among different segments of the power delivery network. The computation and interpretation of the intensity function are thoroughly investigated. We apply the proposed approach to estimate wildfire ignition risks on major transmission lines in California, using historical fire data, meteorological and vegetation data obtained from the National Oceanic and Atmospheric Administration and National Aeronautics and Space Administration. A comprehensive comparison study is performed to show the applicability and predictive capability of the proposed approach.

Traffic Congestion Prediction Using Graph Convolution Networks

Traffic congestion is a critical issue in urban transportation systems, leading to increased travel times, fuel consumption, and air pollution. Furthermore, accurate prediction of traffic conditions is essential for effective traffic management and route planning. However, traditional approaches often fail to capture the complex spatiotemporal dependencies inherent in road networks. This study compares the performance of Graph Convolutional Networks (GCNs) for traffic congestion prediction using Amsterdam sensor location datasets from 13 locations (01-10-2023 to 31-10-2023) and 18 locations (01-01-2024 to 26-01-2024). The GCN model achieves an accuracy above 0.5, with a peak accuracy of 0.6 for the 18-location dataset and 0.55 for the 13-location dataset. Precision ranges from 0.5 to 0.8, while recall oscillates between 0.5 and 0.6. Also, the F1-score reaches 0.6 for the 18-location dataset and remains above 0.4 for the 13-location dataset. The results demonstrate the GCN's effectiveness in capturing spatial dependencies and achieving high-performance metrics, with better performance observed for larger datasets. Moreover, the findings contribute to the development of intelligent schemes for GCNs and the Internet of Vehicles in Intelligent Transportation Systems (ITS), advancing traffic congestion prediction capabilities.

Dynamic Spatio-Temporal Adaptive Graph Convolutional Recurrent Networks for Vacant Parking Space Prediction

The prediction of vacant parking spaces (VPSs) can reduce the time drivers spend searching for parking, thus alleviating traffic congestion. However, previous studies have mostly focused on modeling the temporal features of VPSs using historical data, neglecting the complex and extensive spatial characteristics of different parking lots within the transportation network. This is mainly due to the lack of direct physical connections between parking lots, making it challenging to quantify the spatio-temporal features among them. To address this issue, we propose a dynamic spatio-temporal adaptive graph convolutional recursive network (DSTAGCRN) for VPS prediction. Specifically, DSTAGCRN divides VPS data into seasonal and periodic trend components and combines daily and weekly information with node embeddings using the dynamic parameter-learning module (DPLM) to generate dynamic graphs. Then, by integrating gated recurrent units (GRUs) with the parameter-learning graph convolutional recursive module (PLGCRM) of DPLM, we infer the spatio-temporal dependencies for each time step. Furthermore, we introduce a multihead attention mechanism to effectively capture and fuse the spatio-temporal dependencies and dynamic changes in the VPS data, thereby enhancing the prediction performance. Finally, we evaluate the proposed DSTAGCRN on three real parking datasets. Extensive experiments and analyses demonstrate that the DSTAGCRN model proposed in this study not only improves the prediction accuracy but can also better extract the dynamic spatio-temporal characteristics of available parking space data in multiple parking lots.

Vehicle Behavior Discovery and Three-Dimensional Object Detection and Tracking Based on Spatio-Temporal Dependency Knowledge and Artificial Fish Swarm Algorithm.

In complex traffic environments, 3D target tracking and detection are often occluded by various stationary and moving objects. When the target is occluded, its apparent characteristics change, resulting in a decrease in the accuracy of tracking and detection. In order to solve this problem, we propose to learn the vehicle behavior from the driving data, predict and calibrate the vehicle trajectory, and finally use the artificial fish swarm algorithm to optimize the tracking results. The experiments show that compared with the CenterTrack method, the proposed method improves the key indicators of MOTA (Multi-Object Tracking Accuracy) in 3D object detection and tracking on the nuScenes dataset, and the frame rate is 26 fps.

Prediction on Demand for Regional Online Car-Hailing Travel Based on Self-Attention Memory and ConvLSTM

High precision in forecasting travel demand for online car-hailing is crucial for traffic management to schedule vehicles, hence reducing energy consumption and achieving sustainable development. Netflix demand forecasting relies on the capture of spatiotemporal correlations. To extract the spatiotemporal information more fully, this study designs and develops a novel spatiotemporal prediction model with multidimensional inputs (MSACL) by embedding a self-attention memory (SAM) module into a convolutional long short-term memory neural network (ConvLSTM). The SAM module can extract features with long-range spatiotemporal dependencies. The experimental data are derived from the Chengdu City online car-hailing trajectory data set and the external factors data set. Comparative experiments demonstrate that the proposed model has higher accuracy. The proposed model outperforms the Sa-ConvLSTM model and has the highest prediction accuracy, shows a reduction in the mean absolute error (MAE) by 1.72, a reduction in the mean squared error (MSE) by 0.43, and an increase in the R-squared (R2) by 4%. In addition, ablation experiments illustrate the effectiveness of each component, where the external factor inputs have the least impact on the model accuracy, but the removal of the SAM module results in the most significant decrease in model accuracy.

Spatiotemporal dynamics of epidemiology diseases: mobility based risk and short-term prediction modeling of COVID-19.

Nowadays, epidemiological modeling is applied to a wide range of diseases, communicable and non-communicable, namely AIDS, Ebola, influenza, Dengue, Malaria, Zika. More recently, in the context of the last pandemic declared by the World Health Organization (WHO), several studies applied these models to SARS-CoV-2. Despite the increasing number of researches using spatial analysis, some constraints persist that prevent more complex modeling such as capturing local epidemiological dynamics or capturing the real patterns and dynamics. For example, the unavailability of: (i) epidemiological information such as the frequency with which it is made available; (ii) sociodemographic and environmental factors (e.g., population density and population mobility) at a finer scale which influence the evolution patterns of infectious diseases; or (iii) the number of cases information that is also very dependent on the degree of testing performed, often with severe territorial disparities and influenced by context factors. Moreover, the delay in case reporting and the lack of quality control in epidemiological information is responsible for biases in the data that lead to many results obtained being subject to the ecological fallacy, making it difficult to identify causal relationships. Other important methodological limitations are the control of spatiotemporal dependence, management of non-linearity, ergodicy, among others, which can impute inconsistencies to the results. In addition to these issues, social contact, is still difficult to quantify in order to be incorporated into modeling processes. This study aims to explore a modeling framework that can overcome some of these modeling methodological limitations to allow more accurate modeling of epidemiological diseases. Based on Geographic Information Systems (GIS) and spatial analysis, our model is developed to identify group of municipalities where population density (vulnerability) has a stronger relationship with incidence (hazard) and commuting movements (exposure). Specifically, our framework shows how to operate a model over data with no clear trend or seasonal pattern which is suitable for a short-term predicting (i.e., forecasting) of cases based on few determinants. Our tested models provide a good alternative for when explanatory data is few and the time component is not available, once they have shown a good fit and good short-term forecast ability.

Estimating Daily Snow Density Through a Spatiotemporal Random Forest Model

AbstractSnow density is of paramount importance in water resource management, snow avalanche warning, and climate change research. However, the lack of competent methods for long‐term and vast‐scale snow density mapping persists due to the intricate spatiotemporal dependencies inherent in snow density, resulting in scarce and inaccurate snow density products. To address this challenge, a spatiotemporal random forest (STRF) model is constructed by leveraging in‐situ measurements, multisource remote sensing, and reanalysis data. It tackles the spatiotemporal dependencies in snow density arising from its inherent heterogeneity and the relations involving snow density and nonlinearly connected meteorological, terrain, vegetation, and snow‐related factors. The effectiveness of the model is substantiated through rigorous validation methods, including random, temporal, and spatial block cross‐validations as well as independent validation, apparently surpassing ERA5‐Land snow density. The estimated snow density is also demonstrated to be able to improve existing snow water equivalent data set using fixed snow density. Utilizing the proposed model, a data set of daily 25‐km snow density from 1980 to 2018 is constructed for stable snow cover areas in China, which holds significant potential for research and applications in the realm of snow hydrology.

Short-term wind power forecasting based on multi-scale receptive field-mixer and conditional mixture copula

Integrating renewable wind power (WP) into the grid exacerbates variability and challenges reliability. Establishing an effective forecasting system is crucial for risk avoidance, but achieving effective, interpretable, and accurate predictions remains an obstacle due to the high volatility and uncertainty of WP. This study proposes an innovative and trustworthy wind power forecasting system combined with point and interval predictions. Specifically, the adaptive multi-scale convolutional receptive field (AMSCF) combined with gated recurrent unit (GRU) for WP point prediction is proposed to capture the spatiotemporal dependencies of WP generation. This AMSCF method uses different size receptive fields, fuses them using channel dimension concatenation, operates pooling layer and adaptive fusion through field attention, which extracts hierarchical features for WP forecasting. Then conditional mixture Copula (Con-mCopula) function integrated with multiobjective optimizer is established for WP interval prediction, which reduces the WP interval prediction accuracy error as some information neglected. The wind power cluster dataset from Germany was selected for experiments. Based on the results, for both datasets, the designed prediction system can achieve better point and interval prediction performance resulting in a reduction of mean absolute error (MAE) by approximately 40 % compared to other existing components in the market. This study deconstructs the “Black Box” in deep network for energy forecasting through interpretability analysis, which accelerates the development of AI technology in WP forecasting system and provides decision-makers with trustworthy WP forecasting assistance.

AEMNet: Unsupervised Video Anomaly Detection Method Based on Attention-Enhanced Memory Networks

Video anomaly detection has always been a challenging task in computer vision due to data imbalance and susceptibility to scene variations such as lighting and occlusions. In response to this challenge, this paper proposes an unsupervised video anomaly detection method based on an attention-enhanced memory network. The method utilizes a dual-stream network structure of autoencoders, enhancing the model’s learning ability for important features in appearance and motion by introducing coordinate attention mechanisms and variance attention mechanisms, emphasizing significant characteristics of static objects and rapidly moving regions. By adding memory modules to both the appearance and motion branches, the network structure’s memory information is reinforced, enabling it to capture long-term spatiotemporal dependencies in videos and thereby improving the accuracy of anomaly detection. Furthermore, by optimizing the network structure’s activation functions to handle negative inputs, it enhances its nonlinear modeling capabilities, enabling better adaptation to complex environments, including variations in lighting and occlusions, further improving the effectiveness of anomaly detection. The paper conducts comparative experiments and ablation studies using three public available datasets and various models. The results demonstrate that compared to baseline models, the AUC performance is improved by 3.9%, 4.7%, and 1.7% on UCSD Ped2, CHUK Avenue, and ShanghaiTech datasets, respectively. When compared with the other models, the average AUC performance is improved by 4.3%, 5.4%, and 6.2%, with an average improvement of 8.75% in the ERR metric, validating the effectiveness and adaptability of the proposed method. The code can be obtained at the following URL: https://github.com/AcademicWhite/AEMNet .

Sentinel Shield: Leveraging ConvLSTM and Elephant Herd Optimization for Advanced Network Intrusion Detection

Given the escalating intricacy of network environments and the rising level of sophistication in cyber threats, there is an urgent requirement for resilient and effective network intrusion detection systems (NIDS). This document presents an innovative NIDS approach that utilizes Convolutional Long Short-Term Memory (ConvLSTM) networks and Elephant Herd Optimization (EHO) to achieve precise and timely intrusion detection. Our proposed model combines the strengths of ConvLSTM, which can effectively capture spatiotemporal dependencies in network traffic data, and EHO, which allow the model to focus on relevant information while filtering out noise. To achieve this, we first preprocess network traffic data into sequential form and use ConvLSTM layers to learn both spatial and temporal features. Subsequently, we introduce Elephant Herd Optimization that dynamically assigns different weights to different parts of the input data, emphasizing the regions most likely to contain malicious activity. To evaluate the effectiveness of our approach, we conducted extensive experiments on publicly available network intrusion CICIDS2017 Dataset. The experimental results demonstrate the efficacy of the proposed approach (Accuracy = 99.98%), underscoring its potential to revolutionize modern network intrusion detection and proactively safeguard digital assets.

Predicting air quality using a multi-scale spatiotemporal graph attention network

As urbanization accelerates, air quality has become a pressing concern. Accurate air quality prediction is essential for informed governmental decision-making and for protecting public health. Variations in air quality are influenced by complex multi-scale spatiotemporal processes. Existing research primarily relies on capturing single spatiotemporal features of air quality to predict changes. Meanwhile, when constructing spatiotemporal dynamic graphs, the inherent characteristics of the input data and the comprehensive effects of both global and local influences are not fully considered. To address these problems, we propose a graph-attention-based approach, named Multi-scale Spatiotemporal Graph Attention Network (MSTGAN). MSTGAN addresses the intricate spatiotemporal patterns of air quality across various scales through three key components: (1) a multistation transformer to model the temporal patterns of air quality at individual monitoring stations; (2) a bilinear spatiotemporal attention mechanism to capture the spatiotemporal dynamic global dependencies among all stations in a region; and (3) a set of spatiotemporal dependence graph-coupled Chebyshev graph convolution gate recurrent units to extract and aggregate the local spatiotemporal features of interrelated stations. Experiments conducted on three real-world datasets demonstrated that MSTGAN achieved significant improvements of 4.2%, 3.9%, and 7.8% in the mean absolute error, root mean square error, and R2 evaluation metrics, respectively, compared to seven state-of-the-art time-series forecasting methods. This code is publicly available at https://github.com/HPSCIL/MSTGAN-airquality-prediction.

TSPM-Net: A probabilistic spatio-temporal approach for scanpath prediction

Predicting the path followed by the viewer’s eyes when observing an image (a scanpath) is a challenging problem, particularly due to the inter- and intra-observer variability and the spatio-temporal dependencies of the visual attention process. Most existing approaches have focused on progressively optimizing the prediction of a gaze point given the previous ones. In this work we propose instead a probabilistic approach, which we call tSPM-Net. We build our method to account for observers’ variability by resorting to Bayesian deep learning and a probabilistic approach. Besides, we optimize our model to jointly consider both spatial and temporal dimensions of scanpaths using a novel spatio-temporal loss function based on a combination of Kullback–Leibler divergence and dynamic time warping. Our tSPM-Net yields results that outperform those of current state-of-the-art approaches, and are closer to the human baseline, suggesting that our model is able to generate scanpaths whose behavior closely resembles those of the real ones.

HSeq2Seq: Hierarchical graph neural network for accurate mobile traffic forecasting

Hierarchical graph neural networks (HGNNs) provide a feasible method for modeling complex spatiotemporal dependencies during mobile traffic forecasting. However, most existing studies have adopted spatial node clustering methods to construct a coarsened graph that overlooks the temporal correlation within the original station graph. Furthermore, existing state-of-the-art methods fail to fully exploit the cross-regional feature impacts on stations, which limits their ability to model nonlocal spatial dependency.To overcome these limitations, we proposed a hierarchical sequence-to-sequence (HSeq2Seq) approach that combines HGNNs with a sequence-to-sequence architecture (Seq2Seq) for mobile traffic forecasting. First, a spatiotemporal node clustering method was designed to construct the hierarchical structure. Second, a convolution encoder was employed to extract the local spatiotemporal features from a hierarchical perspective, and a novel attention-based feature fusion module was built to capture the nonlocal spatiotemporal features by identifying both the intra- and cross-regional feature impacts on the stations. Finally, a recurrent decoder is introduced to reweight the spatiotemporal features and recursively produce the final prediction.Extensive experiments on two real-world datasets demonstrate that our model outperforms state-of-the-art methods, while the ablation study also verifies the effectiveness of the spatiotemporal node clustering and the attention-based feature fusion module.

HSFE: A hierarchical spatial-temporal feature enhanced framework for traffic flow forecasting

Currently, spatio-temporal fusion strategy is a key direction in traffic flow prediction. Current work employs a higher degree of spatio-temporal self-attention in order to capture spatio-temporal dependencies. However, the features to which this approach is applied are more targeted, increasing the computational complexity of the process and making it more difficult to capture long-range dependencies. This paper proposes a new framework for predicting traffic flow that enhances spatio-temporal features at multiple levels. The framework includes a periodic embedding module that captures temporal periodicity and encodes input data into more representative feature vectors for model training. Also, a component for fusing parallel channel attention has been designed to adaptively weigh the aggregation of features from global, local, and aggregated channels. This enhances the attention given to important feature information in the model. In addition, a multilevel sequential feature fusion enhancer has been designed that ensures feature processing at different levels. Experimental results on four public transportation datasets demonstrate that the innovative approach enhances the MAE metrics by an average of 2.50%, respectively, over all metrics in the baseline models. Notably, it reduces training time by approximately 50%. This paper also discusses ablation experiments to evaluate the performance of each module.

A spatiotemporal graph transformer approach for Alzheimer’s disease diagnosis with rs-fMRI

Alzheimer’s disease (AD) is a neurodegenerative disease accompanied by cognitive impairment. Early diagnosis is crucial for the timely treatment and intervention of AD. Resting-state functional magnetic resonance imaging (rs-fMRI) records the temporal dynamics and spatial dependency in the brain, which have been utilized for automatically diagnosis of AD in the community. Existing approaches of AD diagnosis using rs-fMRI only assess functional connectivity, ignoring the spatiotemporal dependency mining of rs-fMRI. In addition, it is difficult to increase diagnosis accuracy due to the shortage of rs-fMRI sample and the poor anti-noise ability of model. To deal with these problems, this paper proposes a novel approach for the automatic diagnosis of AD, namely spatiotemporal graph transformer network (STGTN). The proposed STGTN can effectively extract spatiotemporal features of rs-fMRI. Furthermore, to solve the sample-limited problem and to improve the anti-noise ability of the proposed model, an adversarial training strategy is adopted for the proposed STGTN to generate adversarial examples (AEs) and augment training samples with AEs. Experimental results indicate that the proposed model achieves the classification accuracy of 92.58%, and 85.27% with the adversarial training strategy for AD vs. normal control (NC), early mild cognitive impairment (eMCI) vs. late mild cognitive impairment (lMCI) respectively, outperforming the state-of-the-art methods. Besides, the spatial attention coefficients reflected from the designed model reveal the importance of brain connections under different classification tasks.