Abstract:We introduce a novel Graph Attention Autoencoder (GAE) with spatial regularization to address the challenge of scalable anomaly detection in spatiotemporal rainfall data across India from 1990 to 2015. Our model leverages a Graph Attention Network (GAT) to capture spatial dependencies and temporal dynamics in the data, further enhanced by a spatial regularization term ensuring geographic coherence. We construct two graph datasets employing rainfall, pressure, and temperature attributes from the Indian Meteorological Department and ERA5 Reanalysis on Single Levels, respectively. Our network operates on graph representations of the data, where nodes represent geographic locations, and edges, inferred through event synchronization, denote significant co-occurrences of rainfall events. Through extensive experiments, we demonstrate that our GAE effectively identifies anomalous rainfall patterns across the Indian landscape. Our work paves the way for sophisticated spatiotemporal anomaly detection methodologies in climate science, contributing to better climate change preparedness and response strategies.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the anomaly detection in spatio - temporal rainfall data in the Indian region from 1990 to 2015. Specifically, the authors propose a new Graph Attention Autoencoder (GAE) and introduce spatial regularization to capture the long - range spatial dependence and geographical consistency in spatio - temporal rainfall data. The following are the key issues and methods involved in the paper: ### 1. **Research Background and Problem Description** - **Limitations of Traditional Methods**: Traditional autoencoders are computationally inefficient and prone to overfitting when dealing with high - dimensional and spatially complex meteorological data, especially in capturing long - range dependencies. - **Advantages of Graph Autoencoders**: Graph Autoencoders (GAEs) can effectively perform anomaly detection, accurate prediction, and denoising by encoding climate variables (such as precipitation, temperature, etc.) into scalable graph structures, while retaining important long - distance correlations that are often ignored in traditional methods. ### 2. **Proposed Model** - **Model Name**: Spatially Regularized Graph Attention Autoencoder (SRGAttAE) - **Core Objective**: Effectively detect rainfall anomaly patterns in spatially correlated data in the Indian subcontinent. - **Model Architecture**: - **Encoder**: It contains two Graph Attention Network (GAT) layers, which compress the input graph into a 4 - dimensional latent space. - **Decoder**: It also contains two GAT layers, which reconstruct the original graph from the latent representation to minimize the reconstruction error. - **Spatial Regularization**: Ensure that the distance between the latent representations of geographically neighboring nodes is minimized, enhancing geographical consistency. ### 3. **Key Formulas** - **Graph Attention Mechanism**: \[ \alpha_{ij}=\frac{\exp(\text{LeakyReLU}(a^{T}[Wx_{i}\parallel Wx_{j}]))}{\sum_{k\in N(i)}\exp(\text{LeakyReLU}(a^{T}[Wx_{i}\parallel Wx_{k}]))} \] where \(x_{i}\) and \(x_{j}\) are the feature vectors of nodes \(i\) and \(j\), \(W\) is the shared linear transformation matrix, \(a\) is the learning weight vector of the attention mechanism, and \(N(i)\) represents the set of neighbors of node \(i\). - **Spatial Consistency Regularization (SCR)**: \[ L_{\text{SCR}}=\frac{1}{N}\sum_{i = 1}^{N}\sum_{j = 1}^{N}w_{ij}\|z_{i}-z_{j}\|^{2} \] where \(w_{ij}\) is a weight based on the Gaussian kernel, \(\lambda\) is a balancing hyperparameter, and \(\|\cdot\|\) represents the Euclidean distance. - **Total Loss Function**: \[ L_{\text{total}}=L_{\text{recon}}+\lambda\cdot L_{\text{SCR}} \] where \(L_{\text{recon}}=\|X - \hat{X}\|_{F}^{2}\), and \(\|\cdot\|_{F}\) represents the Frobenius norm. ### 4. **Experimental Results** - **Training Data**: Use spatio - temporal rainfall data from 1990 to 2015 for training. - **Performance Evaluation**: The model performs well in reconstructing general rainfall patterns but is lacking in extreme values. By setting a 95 - percentile threshold based on the daily reconstruction error, rainfall anomalies can be effectively identified. - **Computational Efficiency**: Although the model is complex, the training phase only takes about 6 minutes, and the inference operation can complete the data processing of 25 years within one minute.

Spatially Regularized Graph Attention Autoencoder Framework for Detecting Rainfall Extremes

Predicting Air Quality Using a Multi-Scale Spatiotemporal Graph Attention Network

Spatial-Temporal Graph Discriminant AutoEncoder for Traffic Congestion Forecasting

Auto-AD: Autonomous Hyperspectral Anomaly Detection Network Based on Fully Convolutional Autoencoder

Accurate Anomaly Detection Leveraging Knowledge-enhanced GAT

AddAG-AE: Anomaly Detection in Dynamic Attributed Graph Based on Graph Attention Network and LSTM Autoencoder

GD-CAF: Graph Dual-stream Convolutional Attention Fusion for Precipitation Nowcasting

Hybrid Anomaly Detection via Multihead Dynamic Graph Attention Networks for Multivariate Time Series

Hyperspectral Anomaly Detection with Auto-Encoder and Independent Target

MST-GAT: A Multimodal Spatial-Temporal Graph Attention Network for Time Series Anomaly Detection

Graph Regularized Autoencoder and its Application in Unsupervised Anomaly Detection

GAD-NR: Graph Anomaly Detection via Neighborhood Reconstruction

Detection of Spatiotemporally Coherent Rainfall Anomalies Using Markov Random Fields

Graph Masked Autoencoder for Spatio-Temporal Graph Learning

Graph Convolutional Adversarial Networks for Spatiotemporal Anomaly Detection

AddGraph: Anomaly Detection in Dynamic Graph Using Attention-based Temporal GCN

GA-SmaAt-GNet: Generative Adversarial Small Attention GNet for Extreme Precipitation Nowcasting

Dictionary trained attention constrained low rank and sparse autoencoder for hyperspectral anomaly detection

Graph neural network-based anomaly detection for river network systems

Pluvial flood susceptibility mapping for data-scarce urban areas using graph attention network and basic flood conditioning factors