Mainstream Deep Learning Research Articles

RS-Net: Hyperspectral Image Land Cover Classification Based on Spectral Imager Combined with Random Forest Algorithm

Recursive neural networks and transformers have recently become dominant in hyperspectral (HS) image classification due to their ability to capture long-range dependencies in spectral sequences. Despite the success of these sequential architectures, mainstream deep learning methods primarily handle two-dimensional structured data. However, challenges such as the curse of dimensionality, spectral variability, and confounding factors in hyperspectral remote sensing images limit their effectiveness, especially in remote sensing applications. To address this issue, this paper proposes a novel land cover classification algorithm that integrates random forests with a spectral transformer network structure (RS-Net). Firstly, this paper presents a combination of the Gramian Angular Field (GASF) and Gramian Angular Difference Field (GADF) algorithms, which effectively maps the multidimensional time series constructed for each pixel onto two-dimensional image features, enabling precise extraction and recognition in the backend network algorithms and improving the classification accuracy of land cover types. Secondly, to capture the relationships between features at different scales, this paper proposes a SpectralFormer network architecture using the Context and Structure Encoding (CASE) module to effectively learn dependencies between channels. This architecture enhances important features and suppresses unimportant ones, thereby addressing the semantic gap and improving the recognition capability of land cover features. Finally, the final prediction results are determined by a voting mechanism from the Random Forest algorithm, which synthesizes predictions from multiple decision trees to enhance classification stability and accuracy. To better compare the performance of RS-Net, this paper conducted extensive experiments on three benchmark HS datasets obtained from satellite and airborne imagers, comparing various classic neural network models. Surprisingly, the RS-Net algorithm achieves high performance and efficiency, offering a new and effective tool for land cover classification.

Small sample gearbox fault diagnosis based on improved deep forest in noisy environments

ABSTRACT Deep forest methods have gradually emerged as a well-liked substitute for conventional deep neural networks in diagnosing faults in mechanical systems. However, in practical industrial applications, limited training data and severe noise interference pose significant challenges to these models. Existing deep forest models have limitations in information extraction, making it difficult to handle the complexities of industrial environments. To better meet the practical needs of industrial applications, this paper proposes an improved deep forest model – sgic-Forest, specifically designed for fault diagnosis of small sample gearboxes in noisy environments. First, we developed a stacked multi-grained scanning module, which enhances the diversity of feature extraction by integrating the advantages of multiple base learners, thereby better addressing the complexities of industrial data. Secondly, we introduced an important feature selection module, which effectively filters out irrelevant information, significantly improving the model’s robustness in high-noise environments. Experiments on two gearbox datasets show that the proposed method outperforms the basic deep forest model and mainstream deep learning methods in terms of diagnostic accuracy under small sample and noisy conditions.

A New Accurate Aircraft Trajectory Prediction in Terminal Airspace Based on Spatio-Temporal Attention Mechanism

Trajectory prediction serves as a prerequisite for future trajectory-based operation, significantly reducing the uncertainty of aircraft movement information within airspace by scientifically forecasting the three-dimensional positions of aircraft over a certain period. As convergence points in the aviation network, airport terminal airspace exhibits the most complex traffic conditions in the entire air route network. It has stronger mutual influences and interactions among aircraft compared to the en-route phase. Current research typically uses the trajectory time series information of a single aircraft as input for subsequent predictions. However, it often lacks consideration of the close-range spatial interactions between multiple aircraft in the terminal airspace. This results in a gap in the study of aircraft trajectory prediction that couples spatiotemporal features. This paper aims to predict the four-dimensional trajectories of aircraft in terminal airspace, constructing a Spatio-Temporal Transformer (ST-Transformer) prediction model based on temporal and spatial attention mechanisms. Using radar aircraft trajectory data from the Guangzhou Baiyun Airport terminal airspace, the results indicate that the proposed ST-Transformer model has a smaller prediction error compared to mainstream deep learning prediction models. This demonstrates that the model can better integrate the temporal sequence correlation of trajectory features and the potential spatial interaction information among trajectories for accurate prediction.

STU3Net: An Improved U‐Net With Swin Transformer Fusion for Thyroid Nodule Segmentation

ABSTRACTThyroid nodules are a common endocrine system disorder for which accurate ultrasound image segmentation is important for evaluation and diagnosis, as well as a critical step in computer‐aided diagnostic systems. However, the accuracy and consistency of segmentation remains a challenging task due to the presence of scattering noise, low contrast and resolution in ultrasound images. Therefore, we propose a deep learning‐based CAD (computer‐aided diagnosis) method, STU3Net in this paper, aiming at automatic segmentation of thyroid nodules. The method employs a modified Swin Transformer combined with a CNN encoder, which is capable of extracting morphological features and edge details of thyroid nodules in ultrasound images. In decoding through the features for image reconstruction, we introduce a modified three‐layer U‐Net network with cross‐layer connectivity to further enhance image reduction. This cross‐layer connectivity enhances the network's capture and representation of the contained image feature information by creating skip connections between different layers and merging the detailed information of the shallow network with the abstract information of the deeper network. Through comparison experiments with current mainstream deep learning methods on the TN3K and BUSI datasets, we validate the superiority of the STU3Net method in thyroid nodule segmentation performance. The experimental results show that STU3Net outperforms most of the mainstream models on the TN3K dataset, with Dice and IoU reaching 0.8368 and 0.7416, respectively, which are significantly better than other methods. The method demonstrates excellent performance on these datasets and provides radiologists with an effective auxiliary tool to accurately detect thyroid nodules in ultrasound images.

A local enhanced mamba network for hyperspectral image classification

Deep learning has significantly advanced hyperspectral image (HSI) classification, primarily due to its robust nonlinear feature extraction capabilities. The vision transformer has achieved notable performance but is limited by the quadratic computational burden of its self-attention mechanism. Recently, a network based on state space model named Mamba, has attracted considerable attention for its linear complexity and commendable performance. Nevertheless, Mamba was originally designed for one-dimensional causal sequence modeling, and its effectiveness in inherent non-causal HSI classification remains to be fully validated. To address this issue, we propose a novel Local Enhanced Mamba (LE-Mamba) network for hyperspectral image classification, which mainly comprises a Local Enhanced Spatial SSM (LES-S6), a Central Region Spectral SSM (CRS-S6), and a Multi-Scale Convolutional Gated Unit (MSCGU). The LES-S6 improves non-causal local feature extraction by incorporating a multi-directional local spatial scanning mechanism. Additionally, the CRS-S6 employs a bidirectional scanning mechanism in the spectral dimension to capture fine spectral details and integrate them with spatial information. The MSCGU utilizes a convolutional gating mechanism to aggregate features from diverse scanning routes and extract high-level semantic information. The overall accuracies of LE-Mamba on Indian Pines, WHU-Hi-HanChuan, WHU-Hi-LongKou, and Pavia University datasets are 99.16 %, 98.16 %, 99.57 %, and 99.63 %, respectively. Extensive experimental results on these four public datasets demonstrate that the LE-Mamba outperforms eight mainstream deep learning models in classification performance.

Deep Learning Models for PV Power Forecasting: Review

Accurate forecasting of photovoltaic (PV) power is essential for grid scheduling and energy management. In recent years, deep learning technology has made significant progress in time-series forecasting, offering new solutions for PV power forecasting. This study provides a systematic review of deep learning models for PV power forecasting, concentrating on comparisons of the features, advantages, and limitations of different model architectures. First, we analyze the commonly used datasets for PV power forecasting. Additionally, we provide an overview of mainstream deep learning model architectures, including multilayer perceptron (MLP), recurrent neural networks (RNN), convolutional neural networks (CNN), and graph neural networks (GNN), and explain their fundamental principles and technical features. Moreover, we systematically organize the research progress of deep learning models based on different architectures for PV power forecasting. This study indicates that different deep learning model architectures have their own advantages in PV power forecasting. MLP models have strong nonlinear fitting capabilities, RNN models can capture long-term dependencies, CNN models can automatically extract local features, and GNN models have unique advantages for modeling spatiotemporal characteristics. This manuscript provides a comprehensive research survey for PV power forecasting using deep learning models, helping researchers and practitioners to gain a deeper understanding of the current applications, challenges, and opportunities of deep learning technology in this area.

A method for disturbance identification in power quality based on cross-attention fusion of temporal and spatial features

The rapid growth of energy storage scale in power systems has posed higher demands on the monitoring and classification of power quality disturbances (PQDs). Therefore, a method based on cross-attention fusion of temporal and spatial features is proposed to improve the accuracy and robustness of fine classification of PQDs in this study.Firstly, utilizing the improved Sample Convolution and Interaction Model (SCINet) with its recursive downsampling-convolution-interaction architecture, and adding convolutional pooling layers to extract the temporal features of PQDs.Then, the improved VGG model is employed to extract the spatial features of PQDs. Finally, by introducing a cross-attention mechanism to capture the correlation and interaction between the temporal and spatial features, the representation ability of the features is enhanced to achieve disturbance recognition. This study is based on IEEE standards and mathematical models, and generates 25 types of PDQ data through Python simulation for training and testing.The proposed method achieves an average classification accuracy of 95.01 % in a high-noise environment of 20 dB, as well as compared with other mainstream deep learning models and verified through ablation study. Furthermore, the classification test of 10 types of PQDs sampled on a hardware platform demonstrates an average accuracy of 99.66 %, further validating the reliability of the proposed method.

A rendering‐based lightweight network for segmentation of high‐resolution crack images

AbstractHigh‐resolution (HR) crack images provide detailed structural assessments crucial for maintenance planning. However, the discrete nature of feature extraction in mainstream deep learning algorithms and computational limitations hinder refined segmentation. This study introduces a rendering‐based lightweight crack segmentation network (RLCSN) designed to efficiently predict refined masks for HR crack images. The RLCSN combines a deep semantic feature extraction architecture—merging Transformer with a super‐resolution boundary‐guided branch—to reduce environmental noise and preserve crack edge details. It also incorporates customized point‐wise refined rendering for training and inference, focusing computational resources on critical areas, and an efficient sparse training method to ensure efficient inference on commercial mobile computing platforms. Each RLCSN's components are validated through ablation studies and field tests, demonstrating its capability to enable unmanned aerial vehicle‐based inspections to detect cracks as narrow as 0.15 mm from a distance of 3 m, thereby enhancing inspection safety and efficiency.

RefineDet with Improved Attention Block for Chest X-ray Image Lesion Detection

In recent years, artificial intelligence technology has been widely used to assist radiologists in diagnosing and analyzing medical images. The use of artificial intelligence technology can well assist doctors in the localization of lesions. However, the mainstream target detection models at this stage are difficult to be practically applied in medical systems because of factors such as the use of large backbone networks and high input resolution, which leads to low model accuracy and high consumption of computational resources. In this paper, we propose a fast detection speed and high accuracy of the lung lesion detection network IAB- RefineDet. By improving the channel and spatial attention mechanisms and introducing the improved attention module into RefineDet, the lesion detection accuracy is dramatically improved without significantly increasing the number of parameters. We conduct extensive experiments on VinDr-CXR, the world's largest publicly available chest radiograph detection dataset, and comparative experiments with existing mainstream target detection models. The experimental results show that IAB-RefineDet achieves a mAP of 16.23%, and the lesion detection performance is significantly better than mainstream deep learning models.

Fine-tuning vision foundation model for crack segmentation in civil infrastructures

Large-scale foundation models have become the mainstream deep learning method, while in civil engineering, the scale of AI models is strictly limited. In this work, a vision foundation model is introduced for crack segmentation. Two parameter-efficient fine-tuning methods, adapter and low-rank adaptation, are adopted to fine-tune the foundation model in semantic segmentation: the Segment Anything Model (SAM). The fine-tuned CrackSAM shows excellent performance on different scenes and materials. To test the zero-shot performance of the proposed method, two unique datasets related to road and exterior wall cracks are collected, annotated and open-sourced, for a total of 810 images. Comparative experiments are conducted with twelve mature semantic segmentation models. On datasets with artificial noise and previously unseen datasets, the performance of CrackSAM far exceeds that of all state-of-the-art models. CrackSAM exhibits remarkable superiority, particularly under challenging conditions such as dim lighting, shadows, road markings, construction joints, and other interference factors. These cross-scenario results demonstrate the outstanding zero-shot capability of foundation models and provide new ideas for developing vision models in civil engineering.

Business chatbots with deep learning technologies: state-of-the-art, taxonomies, and future research directions

With the support of advanced hardware and software technology, Artificial Intelligence (AI) techniques, especially the increasing number of deep learning algorithms, have spawned the popularization of online intelligent services and accelerated the contemporary development and applications of chatbot systems. The promise of providing 24/7 uninterrupted business services and minimizing workforce costs has made business chatbots a hot topic due to the impact of the pandemic. It has attracted considerable attention from academic researchers and business practitioners. However, a thorough technical review of advanced chatbot technologies and their relevance and applications to various business domains is rare in the literature. The main contribution of this review article is the critical analysis of various chatbot development approaches and the underlying deep learning computational methods in the context of some business applications. We first conceptualize current business chatbot architectures and illustrate the technical characteristics of two common structures. Next, we explore the mainstream deep learning technologies in chatbot design from the perspective of computational methods and usages. Then, we propose a new framework to classify chatbot construction architectures and differentiate the traditional retrieval-based and generation-based chatbots in terms of the modern pipeline and end-to-end structures. Finally, we highlight future research directions for business chatbots to enable researchers to devote their efforts to the most promising research topics and commercial scenarios and for practitioners to benefit from realizing the trend in business chatbot development and applications.

Wildlife Real-Time Detection in Complex Forest Scenes Based on YOLOv5s Deep Learning Network

With the progressively deteriorating global ecological environment and the gradual escalation of human activities, the survival of wildlife has been severely impacted. Hence, a rapid, precise, and reliable method for detecting wildlife holds immense significance in safeguarding their existence and monitoring their status. However, due to the rare and concealed nature of wildlife activities, the existing wildlife detection methods face limitations in efficiently extracting features during real-time monitoring in complex forest environments. These models exhibit drawbacks such as slow speed and low accuracy. Therefore, we propose a novel real-time monitoring model called WL-YOLO, which is designed for lightweight wildlife detection in complex forest environments. This model is built upon the deep learning model YOLOv5s. In WL-YOLO, we introduce a novel and lightweight feature extraction module. This module is comprised of a deeply separable convolutional neural network integrated with compression and excitation modules in the backbone network. This design is aimed at reducing the number of model parameters and computational requirements, while simultaneously enhancing the feature representation of the network. Additionally, we introduced a CBAM attention mechanism to enhance the extraction of local key features, resulting in improved performance of WL-YOLO in the natural environment where wildlife has high concealment and complexity. This model achieved a mean accuracy (mAP) value of 97.25%, an F1-score value of 95.65%, and an accuracy value of 95.14%. These results demonstrated that this model outperforms the current mainstream deep learning models. Additionally, compared to the YOLOv5m base model, WL-YOLO reduces the number of parameters by 44.73% and shortens the detection time by 58%. This study offers technical support for detecting and protecting wildlife in intricate environments by introducing a highly efficient and advanced wildlife detection model.

Learning rate burst for superior SGDM and AdamW integration

Current mainstream deep learning optimization algorithms can be classified into two categories: non-adaptive optimization algorithms, such as Stochastic Gradient Descent with Momentum (SGDM), and adaptive optimization algorithms, like Adaptive Moment Estimation with Weight Decay (AdamW). Adaptive optimization algorithms for many deep neural network models typically enable faster initial training, whereas non-adaptive optimization algorithms often yield better final convergence. Our proposed Adaptive Learning Rate Burst (Adaburst) algorithm seeks to combine the strengths of both categories. The update mechanism of Adaburst incorporates elements from AdamW and SGDM, ensuring a seamless transition between the two. Adaburst modifies the learning rate of the SGDM algorithm based on a cosine learning rate schedule, particularly when the algorithm encounters an update bottleneck, which is called learning rate burst. This approach helps the model to escape current local optima more effectively. The results of the Adaburst experiment underscore its enhanced performance in image classification and generation tasks when compared with alternative approaches, characterized by expedited convergence and elevated accuracy. Notably, on the MNIST, CIFAR-10, and CIFAR-100 datasets, Adaburst attained accuracies that matched or exceeded those achieved by SGDM. Furthermore, in training diffusion models on the DeepFashion dataset, Adaburst achieved convergence in fewer epochs than a meticulously calibrated AdamW optimizer while avoiding abrupt blurring or other training instabilities. Adaburst augmented the final training set accuracy on the MNIST, CIFAR-10, and CIFAR-100 datasets by 0.02%, 0.41%, and 4.18%, respectively. In addition, the generative model trained on the DeepFashion dataset demonstrated a 4.62-point improvement in the Frechet Inception Distance (FID) score, a metric for assessing generative model quality. Consequently, this evidence suggests that Adaburst introduces an innovative optimization algorithm that simultaneously updates AdamW and SGDM and incorporates a learning rate burst mechanism. This mechanism significantly enhances deep neural networks’ training speed and convergence accuracy.

METEO-DLNet: Quantitative Precipitation Nowcasting Net Based on Meteorological Features and Deep Learning

Precipitation prediction plays a crucial role in people’s daily lives, work, and social development. Especially in the context of global climate variability, where extreme precipitation causes significant losses to the property of people worldwide, it is urgently necessary to use deep learning algorithms based on radar echo extrapolation for short-term precipitation forecasting. However, there are inadequately addressed issues with radar echo extrapolation methods based on deep learning, particularly when considering the inherent meteorological characteristics of precipitation on spatial and temporal scales. Additionally, traditional forecasting methods face challenges in handling local images that deviate from the overall trend. To address these problems, we propose the METEO-DLNet short-term precipitation prediction network based on meteorological features and deep learning. Experimental results demonstrate that the Meteo-LSTM of METEO-DLNet, utilizing spatial attention and differential attention, adequately learns the influence of meteorological features on spatial and temporal scales. The fusion mechanism, combining self-attention and gating mechanisms, resolves the divergence between local images and the overall trend. Quantitative and qualitative experiments show that METEO-DLNet outperforms current mainstream deep learning precipitation prediction models in natural spatiotemporal sequence problems.

MFCANN: A feature diversification framework based on local and global attention for human activity recognition

Human activity recognition (HAR) is a crucial detection technique extensively employed in various contexts demanding accurate identification of human actions. Currently, mainstream HAR approaches frequently rely on sequential data generated by wearable sensors, with a primary focus on extracting feature representations that align with various human activities. Nevertheless, complex and diverse human activities encompass various kinds of features and substantial variations in the features emphasized across different activities, making it challenging for current artificial intelligence methods to achieve accurate recognition. To address this challenge, this paper introduces a network framework called the Multi-Feature Combining Attention Neural Network (MFCANN). This framework incorporates diverse feature extraction, along with both local and global feature attention mechanisms. It is composed of stacked Multi-Feature Combining Attention Blocks (MFCAB) designed by us. In contrast to traditional convolutional methods, MFCAB parallelly stacks various convolutional components, enabling the extraction of a broader range of features from human activity data. Additionally, we propose the Intra-Module Attention Block (Intra-MAB) and the Inter-Module Attention Block (Inter-MAB), which simultaneously focus on local fine-grained features within module feature maps and global distinguishing features across module feature maps. This aims to achieve more targeted feature learning, reinforcing the network’s ability to distinguish between different human activities. Abundant experimental results demonstrate that the proposed MFCANN in this paper outperforms current mainstream deep learning algorithms in HAR tasks, achieving recognition accuracies of 0.9813, 0.9324, and 0.9930 on the UCI-HAR, USC-HAD, and RealWorld datasets respectively. We release our code at https://github.com/RangerHeart/MFCANN.

CrackTinyNet: A novel deep learning model specifically designed for superior performance in tiny road surface crack detection

AbstractWith the rapid advancement of highway construction, the maintenance of highway infrastructure has become particularly vital. During highway maintenance, the effective detection of tiny road surface cracks helps to extend the lifespan of roads and enhance traffic efficiency and safety. To elevate the performance of existing road detection models, the CrackTinyNet (CrTNet) algorithm is specifically proposed for detecting tiny road surface cracks. This algorithm utilizes the novel BiFormer general visual transformer, designed expressly for tiny objects, and optimizes the loss function to a normalized Wasserstein distance loss function. It replaces traditional downsampling with Space‐to‐Depth Conv to prevent the excessive loss of tiny object information in the network structure. To highlight the model's advantage in detecting tiny road cracks, ablation experiments and comparison trials were conducted with mainstream deep learning models for crack detection. The results of the ablation experiments show that, compared to the baseline, CrTNet improved the Mean Average Precision (MAP) by 0.22. When compared to other network models suitable for road detection, these results exhibited an improvement of over 8.9%. In conclusion, the CrTNet proposed in this study enables a more accurate detection of tiny road cracks, playing a significant role in the advancement of intelligent traffic management.

Visual Prompting Based Incremental Learning for Semantic Segmentation of Multiplex Immuno-Flourescence Microscopy Imagery.

Deep learning approaches are state-of-the-art for semantic segmentation of medical images, but unlike many deep learning applications, medical segmentation is characterized by small amounts of annotated training data. Thus, while mainstream deep learning approaches focus on performance in domains with large training sets, researchers in the medical imaging field must apply new methods in creative ways to meet the more constrained requirements of medical datasets. We propose a framework for incrementally fine-tuning a multi-class segmentation of a high-resolution multiplex (multi-channel) immuno-flourescence image of a rat brain section, using a minimal amount of labelling from a human expert. Our framework begins with a modified Swin-UNet architecture that treats each biomarker in the multiplex image separately and learns an initial "global" segmentation (pre-training). This is followed by incremental learning and refinement of each class using a very limited amount of additional labeled data provided by a human expert for each region and its surroundings. This incremental learning utilizes the multi-class weights as an initialization and uses the additional labels to steer the network and optimize it for each region in the image. In this way, an expert can identify errors in the multi-class segmentation and rapidly correct them by supplying the model with additional annotations hand-picked from the region. In addition to increasing the speed of annotation and reducing the amount of labelling, we show that our proposed method outperforms a traditional multi-class segmentation by a large margin.

A Pavement Crack Detection and Evaluation Framework for a UAV Inspection System Based on Deep Learning

Existing studies often lack a systematic solution for an Unmanned Aerial Vehicles (UAV) inspection system, which hinders their widespread application in crack detection. To enhance its substantial practicality, this study proposes a formal and systematic framework for UAV inspection systems, specifically designed for automatic crack detection and pavement distress evaluation. The framework integrates UAV data acquisition, deep-learning-based crack identification, and road damage assessment in a comprehensive and orderly manner. Firstly, a flight control strategy is presented, and road crack data are collected using DJI Mini 2 UAV imagery, establishing high-quality UAV crack image datasets with ground truth information. Secondly, a validation and comparison study is conducted to enhance the automatic crack detection capability and provide an appropriate deployment scheme for UAV inspection systems. This study develops automatic crack detection models based on mainstream deep learning algorithms (namely, Faster-RCNN, YOLOv5s, YOLOv7-tiny, and YOLOv8s) in urban road scenarios. The results demonstrate that the Faster-RCNN algorithm achieves the highest accuracy and is suitable for the online data collection of UAV and offline inspection at work stations. Meanwhile, the YOLO models, while slightly lower in accuracy, are the fastest algorithms and are suitable for the lightweight deployment of UAV with online collection and real-time inspection. Quantitative measurement methods for road cracks are presented to assess road damage, which will enhance the application of UAV inspection systems and provide factual evidence for the maintenance decisions made by road authorities.

Deep learning based bi-polar sentiment classification of movie reviews in Hindi

With evolving technology, Hindi web content is growing considerably and catching popularity as the larger audience feels more connected and heard using their native language. A volcanic growth in online movie reviews (MRs) in Hindi has been observed lately; manually analyzing them is impossible. Hence, the research problem of automatic organization and classification of Hindi reviews is apparent as this can help viewers decide whether a movie is worth watching or not. This work focuses on developing a deep learning-based system for bi-polar sentiment classification of MRs for resource deficient language – Hindi. To this end, a primary Hindi movie review (MR) corpus is made and manually annotated with binary polarity class labels - positive or negative. The corpus is preprocessed using the preprocessing steps, and Random Word Embeddings (WEs) are utilized for feature extraction. This paper proposes an ensemble CNN_BiGRU, which is an integration of 1D CNN with BiGRU for the bipolar classification of Hindi MRs. To prove our ensemble’s efficacy; other widely used mainstream deep learning models (DLMs) such as Convolutional Neural Network (CNN), Recurrent Neural Network (RNN) based models – Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM), are also applied and compared with proposed ensemble using average classification accuracy. Empirical results show the effectiveness of the proposed ensemble in achieving a reasonably good average accuracy of 89.366%. The results indicate that proposed CNN_BiGRU compares favorably to the state-of-the-art DLMs applied and hence gives an effective solution for sentence-level bi-polar classification of MRs in a resource deficient scenario.

Short-term prediction of tropical cyclone track and intensity via four mainstream deep learning techniques

Tropical cyclone (TC) is a highly destructive natural disaster, whose impact is closely correlated to its track and intensity. Thus, it is crucial to obtain the information on TC track and intensity both timely and accurately for early warning and mitigation of related disasters. While there are numerous prediction methods, deep learning (DL) technology has demonstrated exceptional potential in recent years. However, studies on TC prediction via DL techniques are still insufficient, and in particular there is a lack of efforts toward the dependence of prediction results upon some key parameters involved in the prediction strategy. To this end, this paper present a comprehensive study on TC prediction via four mainstream DL techniques, i.e., Long Short-Term Memory (LSTM) network, Convolutional LSTM (CovnLSTM) network, Temporal Convolutional Network (TCN), and the recently developed Transformer model. The performance of different DL techniques is examined, with a highlight on the effects of three setting parameters involved in the prediction process on the model performance, i.e., length of input records Nin, interval between input records ΔT, and lead time of the prediction Nout. Results reveal that TCN performs best from the perspectives of prediction accuracy, running efficiency and stability, followed by ConvLSTM; whilst LSTM and Transformer fail to provide competitive predictions, although Transformer showcases notable training efficiency. In reference to the prediction strategy, increasing ΔT and Nout can consistently enlarge prediction errors across all models, while the model performance can be optimized by setting the value of Nin in a range of 3–6. It is interesting to find that the overall performance of all models can be improved noticeably by using records with a finer ΔT in the scenarios with the same lead time but with varied Nout. Meanwhile, TCN can be further improved by extending Nout from Nout = m to Nout = n (n > m) and using the m-th step prediction in the scenario with Nout = n as the output in the scenario with Nout = m. The presented results provide useful insights for reasonably utilizing DL models to deal with prediction issues, which are applicable not only for TCs but also for other objectives that are widely concerned in wind engineering, e.g., wind speed/pressure and structural responses.