Combination Of Long Short-term Memory Research Articles

Anomaly-based Network Intrusion Detection System using Deep Intelligent Technique

Background and objectives: Computer systems and network infrastructures are still exposed to many security risks and cyber-attack vulnerabilities despite advancements of information security. Traditional signature-based intrusion detection systems and security solutions by matching rule-based mechanism and prior knowledge are insufficient of fully protecting computer networks against novel attacks. For this purpose, Anomaly-based Network Intrusion Detection System (A-NIDS) as cyber security tool is considered for identifying and detecting anomalous behavior in the flow-based network traffic alongside with firewalls and other security measures. The main objective of the research is to improve the detection rate and reduce false-positive rates of the classifier using anomaly-based technique. Methods: an intelligent technique using deep learning algorithm and mutual information feature selection (MIFS) method to select optimal features on the benchmark datasets. Proposed method accurately capable of classifying normal and anomalous states of the data packets in a comprehensive way by combination of Long-Short term memory (LSTM) algorithm and MIFS method. Results: The model achieved encouraging results in terms of accuracy 99.79%, 0.002 false-positive rate with minimum time compared to other models recorded only 81.75s on CSE-CIC-IDS2018 dataset. At the end of the study, comparative studies are conducted to verify the effectiveness of proposed method on three realistic and latest intrusion detection datasets, named CSE_CIC-IDS2018, CIC-IDS2017, and NF-CSE-CIC-IDS2018 dataset. Conclusions: Proposed model in a combination of LSTM NN and Feature selection method (MIFS) increased detection rate and reduced false-positive alarms, also the model able to detect low frequent attacks while other existing models are suffering from.

Deep Learning Methods for Arabic Autoencoder Speech Recognition System for Electro-Larynx Device

Recent advances in speech recognition have achieved remarkable performance comparable with human transcribers’ abilities. But this significant performance is not the same for all the spoken languages. The Arabic language is one of them. Arabic speech recognition is bounded to the lack of suitable datasets. Artificial intelligence algorithms have shown promising capabilities for Arabic speech recognition. Arabic is the official language of 22 countries, and it has been estimated that 400 million people speak the Arabic language worldwide. Speech disabilities have been one of the expanding problems in the last decades, even in kids. Some devices can be used to generate speech for those people. One of these devices is the Servox Digital Electro-Larynx (EL). In this research, we developed an autoencoder with a combination of long short-term memory (LSTM) and gated recurrent units (GRU) models to recognize recorded signals from Servox Digital EL Electro-Larynx. The proposed framework consisted of three steps: denoising, feature extraction, and Arabic speech recognition. The experimental results show 95.31% accuracy for Arabic speech recognition with the proposed model. In this research, we evaluated different combinations of LSTM and GRU for constructing the best autoencoder. A rigorous evaluation process indicates better performance with the use of GRU in both encoder and decoder structures. The proposed model achieved a 4.69% word error rate (WER). Experimental results confirm that the proposed model can be used for developing a real-time app to recognize common Arabic spoken words.

Real-Time Economic Dispatch of CHP Systems with Battery Energy Storage for Behind-the-Meter Applications

The use of combined heat and power (CHP) systems has recently increased due to their high combined efficiency and low emissions. Using CHP systems in behind-the-meter applications, however, can introduce some challenges. Firstly, the CHP system must operate in load-following mode to prevent power export to the grid. Secondly, if the load drops below a predefined threshold, the engine will operate at a lower temperature and hence lower efficiency, as the fuel is only half-burnt, creating significant emissions. The aforementioned issues may be solved by combining CHP with a battery energy storage system (BESS); however, the dispatch of CHP and BESS must be optimised. Offline optimisation methods based on load prediction will not prevent power export to the grid due to prediction errors. Therefore, this paper proposes a real-time Energy Management System (EMS) using a combination of Long Short-Term Memory (LSTM) neural networks, Mixed Integer Linear Programming (MILP), and Receding Horizon (RH) control strategy. The RH control strategy is suggested to reduce the impact of prediction errors and enable real-time implementation of the EMS exploiting actual generation and demand data on the day. Simulation results show that the proposed method can prevent power export to the grid and reduce the operational cost by 8.75% compared to the offline method.

Forecasting for Chaotic Time Series Based on GRP-lstmGAN Model: Application to Temperature Series of Rotary Kiln

Rotary kiln temperature forecasting plays a significant part of the automatic control of the sintering process. However, accurate forecasts are difficult owing to the complex nonlinear characteristics of rotary kiln temperature time series. With the development of chaos theory, the prediction accuracy is improved by analyzing the essential characteristics of time series. However, the existing prediction methods of chaotic time series cannot fully consider the local and global characteristics of time series at the same time. Therefore, in this study, the global recurrence plot (GRP)-based generative adversarial network (GAN) and the long short-term memory (LSTM) combination method, named GRP-lstmGAN, are proposed, which can effectively display important information about time scales. First, the data is subjected to a series of pre-processing operations, including data smoothing. Then, transforming one-dimensional time series into two-dimensional images by GRP makes full use of the global and local information of time series. Finally, the combination of LSTM and improves GAN models for temperature time series prediction. The experimental results show that our model is better than comparison models.

A Garlic-Price-Prediction Approach Based on Combined LSTM and GARCH-Family Model

The frequent and sharp fluctuations in garlic prices seriously affect the sustainable development of the garlic industry. Accurate prediction of garlic prices can facilitate correct evaluation and scientific decision making by garlic practitioners, thereby avoiding market risks and promoting the healthy development of the garlic industry. To improve the prediction accuracy of garlic prices, this paper proposes a garlic-price-prediction method based on a combination of long short-term memory (LSTM) and multiple generalized autoregressive conditional heteroskedasticity (GARCH)-family models for the nonstationary and nonlinear characteristics of garlic-price series. Firstly, we obtain volatility characteristic information such as the volatility aggregation of garlic-price series by constructing GARCH-family models. Then, we leverage the LSTM model to learn the complex nonlinear relationships between the garlic-price series and the volatility characteristic information of the series, and predict the garlic price. We applied the proposed model to a real-world garlic dataset. The experimental results show that the prediction performance of the combined LSTM and GARCH-family model containing volatility characteristic information of garlic price is generally better than those of the separate models. The combined LSTM model incorporating GARCH and PGARCH models (LSTM-GP) had the best performance in predicting garlic price in terms of evaluation indexes, such as mean absolute error, root mean-square error, and mean absolute percentage error. The combined model of LSTM-GARCH provides the best results in garlic price prediction and can provide support for garlic price prediction.

Research on the Filtering and Classification Method of Interactive Music Education Resources Based on Neural Network.

This work intends to classify and integrate music genres and emotions to improve the quality of music education. This work proposes a web image education resource retrieval method based on semantic network and interactive image filtering for a music education environment. It makes a judgment on these music source data and then uses these extracted feature sequences as the emotions expressed in the model of the combination of Long Short-Term Memory (LSTM) and Attention Mechanism (AM), thus judging the emotion category of music. The emotion recognition accuracy has increased after improving LSTM-AM into the BiGR-AM model. The greater the difference between emotion genres is, the easier it is to analyze the feature sequence containing emotion features, and the higher the recognition accuracy is. The classification accuracy of the excited, relieved, relaxed, and sad emotions can reach 76.5%, 71.3%, 80.8%, and 73.4%, respectively. The proposed interactive filtering method based on a Convolutional Recurrent Neural Network can effectively classify and integrate music resources to improve the quality of music education.

Comparative analysis of BPNN, SVR, LSTM, Random Forest, and LSTM-SVR for conditional simulation of non-Gaussian measured fluctuating wind pressures

Non-Gaussian fluctuating wind pressure occurs in the separation zone of a structure, which will cause structural fatigue damage. The study of non-Gaussian wind pressure, accordingly, is essential, and the best way to study the non-Gaussian fluctuating wind pressure is the field measurement undoubtedly. However, an emergency such as sensor unavailability or sensor failure can lead to data missing. To this end, it is crucial to simulate the data at some unmeasured or missing points, and the multivariate conditional simulation can solve such problems. Although some scholars have conducted studies using SVM and BPNN, fewer studies use machine learning and deep learning to perform multivariate non-Gaussian conditional simulation. In addition to BPNN and SVM, there are some excellent prediction algorithms such as Long Short-Term Memory (LSTM), Random Forest (RF), and a combination of LSTM and SVR (LSTM-SVR). It should be noted that they can also simulate the time series with a small amount of data and obtain good predictions accuracy. Hence, different algorithms (BPNN, SVR, RF, LSTM, LSTM-SVR) are utilized to perform multivariate non-Gaussian wind pressure conditional simulation and comparative analysis using prediction measures (MAPE, RMSE, R). In this paper, the non-Gaussian characteristics of the wind pressure data measured in the field are analyzed first. Then different algorithms (BPNN, SVR, RF, LSTM, LSTM-SVR) are utilized to perform multivariate non-Gaussian wind pressure conditional simulation by spatial interpolation prediction, and their prediction performance measures (MAPE, RMSE, R) are compared. The comparison of prediction measures (MAPE, RMSE, R) reveals that the combined algorithm (LSTM-SVR) gets better accuracy than other algorithms. Finally, the multivariate non-Gaussian wind pressure series simulated by the combined algorithm (LSTM-SVR) are analyzed statistically and found to be consistent with the statistical properties (PSD, ACF, PDF) of the actual measured wind pressure data. The results indicate that the combined algorithm (LSTM-SVR) can utilize a small amount of data to realize the multivariate non-Gaussian conditional simulation by spatial interpolation prediction more accurately.

5G Cognitive Radio Networks Using Reliable Hybrid Deep Learning Based on Spectrum Sensing

Spectrum sensing is critical in allowing the cognitive radio network, which will be used in the next generation of wireless communication systems. Several approaches, including cyclostationary process, energy detectors, and matching filters, have been suggested over the course of several decades. These strategies, on the other hand, have a number of disadvantages. Energy detectors have poor performance when the signal-to-noise ratio (SNR) is changing, cyclostationary detectors are very complicated, and matching filters need previous knowledge of the main user (PU) signals. Additionally, these strategies rely on thresholds under particular signal-noise model assumptions in addition to the thresholds, and as a result, the detection effectiveness of these techniques is wholly dependent on the accuracy of the sensor. In this way, one of the most sought-after difficulties among wireless researchers continues to be the development of a reliable and intelligent spectrum sensing technology. In contrast, multilayer learning models are not ideal for dealing with time-series data because of the large computational cost and high rate of misclassification associated with them. For this reason, the authors propose a hybrid combination of long short-term memory (LSTM) and extreme learning machines (ELM) to learn temporal features from spectral data and to exploit other environmental activity statistics such as energy, distance, and duty cycle duration for the improvement of sensing performance. The suggested system has been tested on a Raspberry Pi Model B+ and the GNU-radio experimental testbed, among other platforms.

An Automated Toxicity Classification on Social Media Using LSTM and Word Embedding.

The automated identification of toxicity in texts is a crucial area in text analysis since the social media world is replete with unfiltered content that ranges from mildly abusive to downright hateful. Researchers have found an unintended bias and unfairness caused by training datasets, which caused an inaccurate classification of toxic words in context. In this paper, several approaches for locating toxicity in texts are assessed and presented aiming to enhance the overall quality of text classification. General unsupervised methods were used depending on the state-of-art models and external embeddings to improve the accuracy while relieving bias and enhancing F1-score. Suggested approaches used a combination of long short-term memory (LSTM) deep learning model with Glove word embeddings and LSTM with word embeddings generated by the Bidirectional Encoder Representations from Transformers (BERT), respectively. These models were trained and tested on large secondary qualitative data containing a large number of comments classified as toxic or not. Results found that acceptable accuracy of 94% and an F1-score of 0.89 were achieved using LSTM with BERT word embeddings in the binary classification of comments (toxic and nontoxic). A combination of LSTM and BERT performed better than both LSTM unaccompanied and LSTM with Glove word embedding. This paper tries to solve the problem of classifying comments with high accuracy by pertaining models with larger corpora of text (high-quality word embedding) rather than the training data solely.

Prediction of Shanghai Stock Index Based on Investor Sentiment and CNN-LSTM Model

In view of the breakthrough progress of the depth learning method in image and other fields, this paper attempts to introduce the depth learning method into stock price forecasting to provide investors with reasonable investment suggestions. This paper proposes a stock prediction hybrid model named ISI-CNN-LSTM considering investor sentiment based on the combination of long short-term memory (LSTM) and convolutional neural network (CNN). The model adopts an end-to-end network structure, using LSTM to extract the temporal features in the data and CNN to mine the deep features in the data can effectively improve the prediction ability of the model by increasing investor sentiment in the network structure. The empirical part makes a comparative experimental analysis based on Shanghai stock index in China. By comparing the experimental prediction results and evaluation indicators, it verifies the prediction effectiveness and feasibility of ISI-CNN-LSTM network model.

Tracking social media during the COVID-19 pandemic: The case study of lockdown in New York State

Facing the COVID-19 pandemic, governments have implemented a wide range of policies to contain the spread of the virus. During the pandemic, large amounts of COVID-19-related tweets emerge every day. Real-time processing of daily tweets may offer insights for monitoring public opinion about intervention measures implemented. In this work, lockdown policy in New York State has been set as a target of public opinion research. This task includes two stages, stance detection and opinion monitoring. For the stance detection stage, we explored several combinations of different text representations and classification algorithms, finding that the combination of Long Short-Term Memory (LSTM) with Global Vectors for Word Representation (GloVe) outperforms others. Due to the shortage of labeled data, we adopted the data distillation method for the training data augmentation. The augmentation of the training data allows to improve the performance of the model with a very small amount of manually-labeled data. After applying the distillation method, the accuracy of the model has been significantly improved. Utilizing the enhanced model, automatically classified tweets are analyzed over time to monitor the public opinion. By exploring the tweets in New York from January 22nd until September 30th, 2020, we show the correlation of public opinion with COVID-19 cases and mortality data, and the effect of government responses on the opinion shift. These results demonstrate the capability of the presented method to effectively and efficiently monitor public opinion during a pandemic.

A federated learning system with enhanced feature extraction for human activity recognition

With the rapid growth of mobile devices, wearable sensor-based human activity recognition (HAR) has become one of the hottest topics in the Internet of Things. However, it is challenging for traditional approaches to achieving high recognition accuracy while protecting users’ privacy and sensitive information. To this end, we design a federated learning system for HAR (HARFLS). Based on the FederatedAveraging method, HARFLS enables each user to handle its activity recognition task safely and collectively. However, the recognition accuracy largely depends on the system’s feature extraction ability. To capture sufficient features from HAR data, we design a perceptive extraction network (PEN) as the feature extractor for each user. PEN is mainly composed of a feature network and a relation network. The feature network, based on a convolutional block, is responsible for discovering local features from the HAR data while the relation network, a combination of long short-term memory (LSTM) and attention mechanism, focuses on mining global relationships hidden in the data. Four widely used datasets, i.e., WISDM, UCI_HAR 2012, OPPORTUNITY, and PAMAP2, are used for performance evaluation. Experimental results demonstrate that PEN outperforms 14 existing HAR algorithms on these datasets in terms of the F1-score; HARFLS with PEN obtains better recognition results on the WISDM and PAMAP2 datasets, compared with 11 existing federated learning systems with various feature extraction structures.

Short-Term Net Load Forecasting with Singular Spectrum Analysis and LSTM Neural Networks

Short-term electricity load forecasting is key to the safe, reliable, and economical operation of power systems. An important challenge that arises with high-frequency load series, e.g., hourly load, is how to deal with the complex seasonal patterns that are present. Standard approaches suggest either removing seasonality prior to modeling or applying time series decomposition. This work proposes a hybrid approach that combines Singular Spectrum Analysis (SSA)-based decomposition and Artificial Neural Networks (ANNs) for day-ahead hourly load forecasting. First, the trajectory matrix of the time series is constructed and decomposed into trend, oscillating, and noise components. Next, the extracted components are employed as exogenous regressors in a global forecasting model, comprising either a Multilayer Perceptron (MLP) or a Long Short-Term Memory (LSTM) predictive layer. The model is further extended to include exogenous features, e.g., weather forecasts, transformed via parallel dense layers. The predictive performance is evaluated on two real-world datasets, controlling for the effect of exogenous features on predictive accuracy. The results showcase that the decomposition step improves the relative performance for ANN models, with the combination of LSTM and SAA providing the best overall performance.

Prediction of Head Movement in 360-Degree Videos Using Attention Model.

In this paper, we propose a prediction algorithm, the combination of Long Short-Term Memory (LSTM) and attention model, based on machine learning models to predict the vision coordinates when watching 360-degree videos in a Virtual Reality (VR) or Augmented Reality (AR) system. Predicting the vision coordinates while video streaming is important when the network condition is degraded. However, the traditional prediction models such as Moving Average (MA) and Autoregression Moving Average (ARMA) are linear so they cannot consider the nonlinear relationship. Therefore, machine learning models based on deep learning are recently used for nonlinear predictions. We use the Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) neural network methods, originated in Recurrent Neural Networks (RNN), and predict the head position in the 360-degree videos. Therefore, we adopt the attention model to LSTM to make more accurate results. We also compare the performance of the proposed model with the other machine learning models such as Multi-Layer Perceptron (MLP) and RNN using the root mean squared error (RMSE) of predicted and real coordinates. We demonstrate that our model can predict the vision coordinates more accurately than the other models in various videos.

An integrated deep learning model for motor intention recognition of multi-class EEG Signals in upper limb amputees

Background and objectiveRecognition of motor intention based on electroencephalogram (EEG) signals has attracted considerable research interest in the field of pattern recognition due to its notable application of non-muscular communication and control for those with severe motor disabilities. In analysis of EEG data, achieving a higher classification performance is dependent on the appropriate representation of EEG features which is mostly characterized by one unique frequency before applying a learning model. Neglecting other frequencies of EEG signals could deteriorate the recognition performance of the model because each frequency has its unique advantages. Motivated by this idea, we propose to obtain distinguishable features with different frequencies by introducing an integrated deep learning model to accurately classify multiple classes of upper limb movement intentions. MethodsThe proposed model is a combination of long short-term memory (LSTM) and stacked autoencoder (SAE). To validate the method, four high-level amputees were recruited to perform five motor intention tasks. The acquired EEG signals were first preprocessed before exploring the consequence of input representation on the performance of LSTM-SAE by feeding four frequency bands related to the tasks into the model. The learning model was further improved by t-distributed stochastic neighbor embedding (t-SNE) to eliminate feature redundancy, and to enhance the motor intention recognition. ResultsThe experimental results of the classification performance showed that the proposed model achieves an average performance of 99.01% for accuracy, 99.10% for precision, 99.09% for recall, 99.09% for f1_score, 99.77% for specificity, and 99.0% for Cohen's kappa, across multi-subject and multi-class scenarios. Further evaluation with 2-dimensional t-SNE revealed that the signal decomposition has a distinct multi-class separability in the feature space. ConclusionThis study demonstrated the predominance of the proposed model in its ability to accurately classify upper limb movements from multiple classes of EEG signals, and its potential application in the development of a more intuitive and naturalistic prosthetic control.

Fast‐moving coin recognition using deep learning

Deep learning has made essential contributions to the development of visual object detection and recognition. Identifying fast-moving objects from the viewpoint of computer vision remains as a challenging problem. The best solution in deep learning that can well represent the characteristics of object motion is related to recurrent neural network (RNN), the best model for fast-moving object recognition is Long Short-Term Memory (LSTM) in RNN. Therefore, the combination of LSTM and CNN fully utilizes spatial and temporal features of moving objects. In this paper, our goal is to identify fast-moving coins from digital videos by using deep learning methods, especially based on the mixture of LSTM and CNN. By using the proposed method, we gain the attainment with high accuracy of fast-moving coin recognition which is superior to our human visual system.

Hotspots for Emerging Epidemics: Multi-Task and Transfer Learning over Mobility Networks

The sudden emergence of epidemics, such as COVID-19, entails economic and social challenges requiring immediate attention from policy makers. An essential building block in implementing mitigation policies (e.g., lockdowns, testing, and vaccination) is the identification of potential hotspots, defined as locations that contribute significantly to the spatial diffusion of infections. During the initial stages of an epidemic, information related to the pathways of spatial diffusion of infection is not fully observable, making the detection of hotspots difficult. This work proposes a data-driven framework to identify hotspots using advanced analytical methodologies, specifically, a combination of interpretable long short-term memory (LSTM) model, multi-task learning, and transfer learning. Our methodology considers mobility within- and across-locations, which is the primary driving factor for the diffusion of infection over a network of connected locations. Additionally, to augment the signals of infection diffusion and the emergence of hotspots, we use transfer learning from past influenza transmission data, which follow a similar transmission mechanism as COVID-19. To illustrate the practical importance of our framework in deciding on lockdown policies, we compare the hotspots-based policy with a pure infection load-based policy and the state-wide lockdown policy used in practice. We show that the hotspots-based lockdown policy can achieve up to 21% improvement in reducing new infections as compared to an infection-based lockdown policy. In addition, we illustrate that locking down only top few hotspot counties can achieve almost similar performance as a state-wide lockdown policy used in practice. Finally, we demonstrate that the inclusion of transfer learning improves hotspot prediction accuracy by 53.4%. We also compare our model performance with the commonly used compartmental epidemiological model and demonstrate the superior prediction performance. Our paper addresses a practical problem with hotspot identification framework, which policy makers can use to improve mitigation decisions related to the control of epidemics.

De-identifying free text of Japanese electronic health records

BackgroundRecently, more electronic data sources are becoming available in the healthcare domain. Electronic health records (EHRs), with their vast amounts of potentially available data, can greatly improve healthcare. Although EHR de-identification is necessary to protect personal information, automatic de-identification of Japanese language EHRs has not been studied sufficiently. This study was conducted to raise de-identification performance for Japanese EHRs through classic machine learning, deep learning, and rule-based methods, depending on the dataset.ResultsUsing three datasets, we implemented de-identification systems for Japanese EHRs and compared the de-identification performances found for rule-based, Conditional Random Fields (CRF), and Long-Short Term Memory (LSTM)-based methods. Gold standard tags for de-identification are annotated manually for age, hospital, person, sex, and time. We used different combinations of our datasets to train and evaluate our three methods. Our best F1-scores were 84.23, 68.19, and 81.67 points, respectively, for evaluations of the MedNLP dataset, a dummy EHR dataset that was virtually written by a medical doctor, and a Pathology Report dataset. Our LSTM-based method was the best performing, except for the MedNLP dataset. The rule-based method was best for the MedNLP dataset. The LSTM-based method achieved a good score of 83.07 points for this MedNLP dataset, which differs by 1.16 points from the best score obtained using the rule-based method. Results suggest that LSTM adapted well to different characteristics of our datasets. Our LSTM-based method performed better than our CRF-based method, yielding a 7.41 point F1-score, when applied to our Pathology Report dataset. This report is the first of study applying this LSTM-based method to any de-identification task of a Japanese EHR.ConclusionsOur LSTM-based machine learning method was able to extract named entities to be de-identified with better performance, in general, than that of our rule-based methods. However, machine learning methods are inadequate for processing expressions with low occurrence. Our future work will specifically examine the combination of LSTM and rule-based methods to achieve better performance.Our currently achieved level of performance is sufficiently higher than that of publicly available Japanese de-identification tools. Therefore, our system will be applied to actual de-identification tasks in hospitals.

End-to-end multivariate time series classification via hybrid deep learning architectures

Deep learning has revolutionized many areas, including time series data mining. Multivariate time series classification (MTSC) remained to be a well-known problem in the time series data mining community, due to its availability in various practical applications such as healthcare, finance, geoscience, and bioinformatics. Recently, multivariate long short-term memory with fully convolutional network (MLSTM-FCN) and multivariate attention long short-term memory with fully convolutional network (MALSTM-FCN) have shown superior results over various state-of-the-art methods. So, in this paper, we explore the usage of recurrent neural network (RNN), and its variants, such as bidirectional recurrent neural network (BiRNN), bidirectional long short-term memory (BiLSTM), gated recurrent unit (GRU), and bidirectional gated recurrent unit (BiGRU). We augment these RNN variants separately by replacing long short-term memory (LSTM) in MLSTM-FCN, which is the combination of LSTM, squeeze-and-excitation (SE) block, and fully convolutional network (FCN). Moreover, we integrate the SE block within FCN to leverage its high performance for the MTSC task. The resulting algorithms do not require heavy pre-processing or feature crafting. Thus, they could be easily deployed on real-time systems. We conduct a comprehensive evaluation with a large number of standard datasets and demonstrate that our approaches achieve notable results over the current best MTSC approach.

The Use of LSTM-Based RNN and SVM Models to Detect Ludian Coseismic Landslides in Time Series Images

With the increase of temporal frequency of satellite images, time series analysis based on Artificial Neural Network has become a tendency to detect land cover changes in images. We briefly introduces a case study that uses Long-Short Term Memory (LSTM), a specific Recurrent Neural Network (RNN) for time series modelling and forecasting, and Support Vector Machine (SVM) to detect landslides triggered by Ms.6.5 earthquake in Ludian, China in 2014. The study uses 72 available images with 16/30m spatial resolution from Landsat-7/8, GaoFen-1 and HJ-1A/1B satellites. Firstly, per-pixel LSTM models are trained by Normalized Difference Vegetation Index (NDVI) images before the earthquake. Secondly, the trained LSTM models are used to predict NDVI images after the earthquake. Then, anomalies or changes are detected by comparing predicted and observed NDVI images. Finally, anomalies related to landslides are separated from other changes with a SVM model which was trained by multi-spectral images in the study area. Experiment demonstrates that the recall rate and precision rate of landslides detection are 82.09% and 76.21%, respectively. The study shows a potential that the combination of LSTM and SVM models can be used to detect landslides in Landsat-like time series images with 30m resolution.