Landslides are an important category of geo-hazards which endanger millions of people annually, especially during rainy season. Conventional approaches to monitoring landslides are progressively being trumped by modern methods such as remote sensing and deep learning models. This technology not only supports for landslide identification but also classification. Automatic classification for landslide types has not attended dur to various challenges related to obtaining large, rich data and the overfitting. This study focuses on new process to detect and classify different landslide types using high-resolution remote-sensing images and deep learning models. This study classifies five major types of landslides: (1) Debris flow, (2) Transitional landslides, (3) Rotational landslides, (4) Shallow landslides, and (5) Human-induced landslides. Two architectures including U‑Net and BiSeNet were applied in classifying five types of landslides using WorldView‑2 images with various optimization techniques. These models demonstrate high accuracy for differentiating landslides from non-landslides as well as for classifying landslides into different categories where average accuracy is higher than 95

Due to the danger of shallow landslides, mountain people annually experience house destruction, which results in the complete loss of their belongings. In the era of technology 4.0 nowadays, the implementation of warning technology systems based on artificial intelligence should become operational to reduce this hazard. The research study aims to develop and apply two AI models: (1) a deep learning model for detecting shallow-landslide traces and (2) a machine learning model for warning of shallow landslides. The first model utilized high-resolution remote sensing images (Worldview-2) from Google Earth Pro to detect shallow landslides, while the second model employed various terrain features, land cover, weather conditions, and remote sensing indices such as NDVI, NDBI and BSI from Sentinel-2 images to assess shallow-landslide susceptibility specifically in the Da River watershed area of northwest Vietnam. As a result, a DeepLab-v3 model (the first one) for shallow-landslide detection achieves 97% accuracy. The main locations for shallow landslide in the Northwestern part of Vietnam are on steep surfaces since the destruction of vegetation reveals vulnerable areas and water sources located near the streams. The Random Forest model (the second one) can warn of shallow-landslide susceptibility using 11 verified input variables with an accuracy of 99%. After testing with average daily rainfall in 12 months, the susceptibility evaluation maps show how factors increase significantly during April through July and then remain stable between September and the rainy season expiration. The combination of heavy rainfall across May through July results in the maximum natural hazard for shallow landslides. Based on the integration of two AI models, the monitoring and warning of shallow landslides can become more precise, which will aid executives during their disaster response operations in the future.

Abstract. Current forecasting models for landslides and debris flows mostly look at environmental or socio-economic factors on their own. They rarely combine both into a single probabilistic framework that might give warning in complicated and uncertain situations. This constraint is especially clear in Vietnam, where intense subtropical rain, steep and extensively dissected mountainous terrain, and quick changes in land use and infrastructure are the main causes of landslides and debris flows. This research introduces a novel approach using a Bayesian Belief Network (BBN) to enhance landslide-risk prediction through the integrated analysis of environmental and socioeconomic data. The developed BBN model incorporates inputs from diverse sources, including Geographic Information Systems (GIS), remote sensing, and field survey observations. Structural Equation Modeling was employed to align the BBN with established relationships between landslides and influencing factors. The analysis explored different scenarios by combining rainfall intensity with land-use patterns and assessing the protective role of embankments. Results indicate that precipitation exceeding 130 mm over a period longer than three days markedly increases the likelihood of landslides and debris flows, particularly in agricultural regions. Gabion embankments were found to be highly effective in mitigating risks to both human safety and built environments.

Debris flows are one of the most common and hazardous types of natural disasters in mountainous regions of Vietnam, particularly in the Than Uyen region, Lai Chau Province. This region is characterized by steep terrain, high rainfall, degraded vegetation cover, and weak soil structure, all of which contribute to the frequent triggering of debris flows. This study aims to assess debris flow susceptibility using Linear Regression and the Random Forest (RF) model, integrating remote sensing data, GIS, and field surveys. A total of ten input variables were selected to train the models using 422 sample points. The RF model demonstrated superior performance, achieving an accuracy of 86% and an AUC of 0.91, compared to 74% accuracy and an R² of 0.65 from the regression model. Given its higher reliability and practical relevance, the RF model was used to generate a debris flow susceptibility map for the entire region. Field validation conducted in Khoen On, Ta Mung, and Muong Cang communes confirmed a strong spatial agreement between high-susceptibility zones and actual debris flow events. The study recommends implementing early warning systems, continuous monitoring, and resettlement planning in high- susceptibility areas. It also emphasizes integrating the model outputs into spatial planning frameworks to enhance climate change adaptation and disaster risk reduction in mountainous areas.

Debris flow inventory is an essential task for scientists and managers to mitigate danger to humans, especially in mountainous areas. However, rapid land use and cover change, as well as technological limitations, make it a challenging task. Monitoring debris-flow efforts, especially in hilly places with limited transportation and technology, may improve management to minimize damage caused by this hazard. This work assesses U-shaped deep learning architectures, focusing on the roles of image size, optimization procedures, and data quality in debris flow trace identification using U-Net and U2-Net. While new debris flows can be detected through machine learning modeling, the U-Net model, combined with the Adam optimizer and an input size of 64×64, has been proven to be efficient, accurate, and stable. Small debris traces that can be used for planning debris thickness maps were easily identified in Worldview-2 and UAV images but not in the medium-resolution remote sensing data. When applied to Bat Xat district, Vietnam, the models identified that the distribution of debris flows is not uniform and depends on natural factors, such as rainfall and human-interpolated factors, including the construction of structures. The study also establishes the need to continually assess and incorporate big data for enhanced debris flow hazard assessment and mitigation. Further developments should focus on the effective use of multi-spectral and large-scale topographic data to strengthen disaster risk identification and provide recommendations for disaster risk reduction.

The occurrence of natural disasters, especially with landslides, threatens mountainous districts and has serious consequences on local tourism development. Future disaster management must develop efficient innovative tools to control the rising frequency and intensity of landslides due to the impacts of economic development and climate change. Minimizing the risk and effects of these occurrences relies on the establishment of an optimal early warning system. This study focuses on the integration of artificial intelligence approaches to identify landslides and evaluate their susceptibility, with an emphasis on early warning systems on tourist routes in Da Bac district. As the first tool in the system, advanced deep learning models using satellite data at high resolution assist in identifying landslides. As a result, a developed DeepLab-v3 model demonstrated high performance by reaching 0.213 dice coefficient and 96.8% accuracy for landslide detection without restrictions from specific input resolution sizes. As the second tool, various machine learning tools, such as Random Forest and Support Vector Machine, utilize the identified landslide locations from the first tool to assess and map their susceptibility based on environmental and human-made factors. Accordingly, the study proposed an early warning system for landslide disaster management using real-time ecological factors and historical data. The proposed integrated system helps tourists and local communities take preventive actions that reduce landslide impacts, thus achieving safety goals in tourism activities, particularly in the Da Bac district of Hoa Binh province, Vietnam. It enhances strategies to minimize risk, increases the ability to predict landslide-prone tourist areas, and aids in implementing sustainable tourism in the future.

As a slope hazard that seriously affects mountainous people, landslides always have inheritance and repetition. Due to the rapid change in land use/cover and the poor performance of remote sensing technologies, the landslide-monitoring task becomes more difficult. Ongoing monitoring endeavors have the potential to enhance the precision and effectiveness of models. Therefore, this study looks at several semantic segmentation models and their backbone architecture and input size in order to find landslides using Worldview-2 images. An indicator system to detect landslide traces on the field and very high-resolution images was proposed in detail. An examination of the training and assessment of U-Net, DeepLab-v3, and PSPN indicates that DeepLab-v3 exhibits consistently high accuracy (from 96 to 99

The robust and accurate identification of different forms of manure stands as a pivotal imperative within the domain of agriculture. Near-infrared (NIR) Spectroscopy has emerged as an expeditious, efficient, non-destructive, and reliable approach to addressing this challenging task. NIR spectroscopy has the potential to serve as a valuable tool for the classification and identification of manure varieties. In order to enhance fertilizer identification performance, this study proposes a novel model called NIRsViT which classifies fertilizers by employing a combination of deep learning Vision Transformer model on NIR spectral data. The introduced model’s performance outperforms existing deep learning models, with an F1-Score of 86.42% and an accuracy rate of 95.19%. Additionally, the model’s classification performance has been significantly improved by proposed imbalanced data processing approaches, Focal Loss, and Upsample, with an F1-Score up to 93.91%, the improved F1-score proved that imbalanced data was considerably solved. The proposed method is a promising approach to handling imbalanced NIR spectral data and acts as a pioneering benchmark for subsequent models in manure identification through NIR spectroscopy. Future research gears toward improving the NIRsViT model’s temporal efficiency and computational load, while also testing the introduced imbalanced data handling approach for efficiency comparison across various models and larger datasets.

We consider unmanned aerial vehicle (UAV)-enabled wireless systems where downlink communications between a multi-antenna UAV and multiple users are assisted by a hybrid active-passive reconfigurable intelligent surface (RIS). We aim at a fairness design of two typical UAV-enabled networks, namely the static-UAV network where the UAV is deployed at a fixed location to serve all users at the same time, and the mobile-UAV network which employs the time division multiple access protocol. In both networks, our goal is to maximize the minimum rate among users through jointly optimizing the UAV's location/trajectory, transmit beamformer, and RIS coefficients. The resulting problems are highly nonconvex due to a strong coupling between the involved variables. We develop efficient algorithms based on block coordinate ascend and successive convex approximation to effectively solve these problems in an iterative manner. In particular, in the optimization of the mobile-UAV network, closed-form solutions to the transmit beamformer and RIS passive coefficients are derived. Numerical results show that a hybrid RIS equipped with only 4 active elements and a power budget of 0 dBm offers an improvement of 38%-63% in minimum rate, while that achieved by a passive RIS is only about 15%, with the same total number of elements.

In recent years, Magnetic Resonance Imaging (MRI) has emerged as a prevalent medical imaging technique, offering comprehensive anatomical and functional information. However, the MRI data acquisition process presents several challenges, including time-consuming procedures, prone motion artifacts, and hardware constraints. To address these limitations, this study proposes a novel method that leverages the power of generative adversarial networks (GANs) to generate multi-domain MRI images from a single input MRI image. Within this framework, two primary generator architectures, namely ResUnet and StarGANs generators, were incorporated. Furthermore, the networks were trained on multiple datasets, thereby augmenting the available data, and enabling the generation of images with diverse contrasts obtained from different datasets, given an input image from another dataset. Experimental evaluations conducted on the IXI and BraTS2020 datasets substantiate the efficacy of the proposed method compared to an existing method, as assessed through metrics such as Structural Similarity Index (SSIM), Peak Signal-to-Noise Ratio (PSNR) and Normalized Mean Absolute Error (NMAE). The synthesized images resulting from this method hold substantial potential as invaluable resources for medical professionals engaged in research, education, and clinical applications. Future research gears towards expanding experiments to larger datasets and encompassing the proposed approach to 3D images, enhancing medical diagnostics within practical applications.

Forest plant identification and classification play a pivotal role in various domains, encompassing biodiversity conservation, agricultural advancement, and beyond. Conventional plant identification methods often rely on expert botanists or manual identification approaches, which can be time-consuming and subjective. Deep learning models have emerged as a promising approach to automatically classify plants, offering high accuracy and efficiency. However, these models often rely on convolutional neural networks (CNNs) and their variants to extract features, which may fail to capture the complex relationships among plant characteristics. This paper proposes a novel feature extraction method using semi-supervised learning techniques combined with Masked Autoencoder architecture to enhance the feature extraction of plant data, applicable to problems with limited datasets. The proposed model, named MAE SGD, achieves an accuracy of nearly 94% on the QuangNamForestPlant - a dataset collected by our research team in Quang Nam province, Central Vietnam, comprising 24,314 images of 710 different forest plant species. Future research directions will focus on expanding the forest plant dataset and improving the recognition model to increase the model’s accuracy and overall performance in identifying forest vegetation.

Intensity modulation/direct detection (IM/DD) remains to be the preferred optical transmission scheme for short-range applications for its simplicity of application, inexpensiveness, and small footprint. However, the impairments of low-cost device and fiber chromatic dispersion lead to the limitation of system performance when the data rate rises to 100 Gbps or higher. In this paper, we demonstrated that an equalizer using neural networks can effectively improve the transmission performance of high-speed IM/DD systems. An optimization of a long short-term memory (LSTM) structure in terms of network depth and distribution of neurons in hidden layers leads to an enhancement of the overall performance of the 50 Gbaud PAM4 communications. Furthermore, the results for a system using a LSTM-based equalizer give the better outcome than the traditional feed-forward equalizer (FFE) or artificial neural network (ANN)-based equalizer.

Labeling and classifying a large number of products is one of the key challenges that e-commerce managers face. Building an automatic model that can accurately classify products helps to optimize the consumer search experience and ensure that they can easily find the products that meet their needs. In this study, we propose an improved Multimodal Deep Learning Model, based on the attention mechanism. This model has the ability to significantly improve accuracy over both traditional Unimodal Deep Learning and Multimodal Deep Learning models. The accuracy of our proposed model reaches 91.18

A cell-free (CF) massive multiple-input-multiple-output (mMIMO) system can provide uniform spectral efficiency (SE) with simple signal processing. On the other hand, a recently introduced technology called hybrid relay-reflecting intelligent surface (HR-RIS) can customize the physical propagation environment by simultaneously reflecting and amplifying radio waves in preferred directions. Thus, it is natural that incorporating HR-RIS into CF mMIMO can be a symbiotic convergence of these two technologies for future wireless communications. This motivates us to consider an HR-RIS-aided CF mMIMO system to utilize their combined benefits. We first model the uplink/downlink channels and derive the minimum-mean-square-error estimate of the effective channels. We then present a comprehensive analysis of SE performance of the considered system. Specifically, we derive closed-form expressions for the uplink and downlink SE. The results reveal important observations on the performance gains achieved by HR-RISs compared to conventional systems. The presented analytical results are also valid for conventional CF mMIMO systems and those aided by passive reconfigurable intelligent surfaces. Such results play an important role in designing new transmission strategies and optimizing HR-RIS-aided CF mMIMO systems. Finally, we provide extensive numerical results to verify the analytical derivations and the effectiveness of the proposed system design under various settings.

The advanced payload technology has opened up a new way to design future NGSO satellite systems exploiting the full flexibility in radio resource and beam coverage management. Conventional spatial multiplexing techniques, which require the CSI, however, cannot be efficiently applied in NGSO due to long round-trip time(RTT). In this paper, we tackle the long RTT in the precoding design by proposing a joint channel prediction and dynamic radio resource management framework. Our aim is to optimize the bandwidth and transmit power in every spot beam based on the predicted channel gains to maximize the system capacity. Since the satellite’s orbit is time-varying but predictable, Kalman filter-based channel estimation method is employed. Given the predicted channels, a joint bandwidth allocation and precoding design is formulated. The effectiveness of the proposed framework is demonstrated via practical satellite channel models using the STK software and 3GPP codebook- and non-codebook-based precoding designs.

Landslides endanger lives and public infrastructure in mountainous areas. Monitoring landslide traces in real-time is difficult for scientists, sometimes costly and risky because of the harsh terrain and instability. Nowadays, modern technology may be able to identify landslide-prone locations and inform locals for hours or days when the weather worsens. This study aims to propose indicators to detect landslide traces on the fields and remote sensing images; build deep learning (DL) models to identify landslides from Sentinel-2 images automatically; and apply DL-trained models to detect this natural hazard in some particular areas of Vietnam. Nine DL models were trained based on three U-shaped architectures, including U-Net, U2-Net, and U-Net3+, and three options of input sizes. The multi-temporal Sentinel-2 images were chosen as input data for training all models. As a result, the U-Net, using an input image size of 32 × 32 and a performance of 97 % with a loss function of 0.01, can detect typical landslide traces in Vietnam. Meanwhile, the U-Net (64 × 64) can detect more considerable landslide traces. Based on multi-temporal remote sensing data, a different case study in Vietnam was chosen to see landslide traces over time based on the trained U-Net (32 × 32) model. The trained model allows mountain managers to track landslide occurrences during wet seasons. Thus, landslide incidents distant from residential areas may be discovered early to warn of flash floods.

In the current digital era, text documents become valuable for businesses to reach potential customers and curtail advertising costs. However, extracting and classifying beneficial information from texts can prove challenging and time-consuming, particularly in complex languages like Vietnamese. This study aims to classify the sentiment of Vietnamese comments on e-commerce websites into negative and positive classes. To enhance the performance of sentiment classification, the study fine-tuned traditional models of Convolutional Neural Networks and Recurrent Neural Networks (RNN). Then, this research proposed a combination of RNN and attention mechanisms at the word and word-and-sentence levels of the input document. The results showed an impressive accuracy of 93.72% and an F1 score of 93.7% on the RNN model with a word-and-sentence-level attention mechanism. This research outcome contributes to the field of text classification and could be applied in opinion mining, customer feedback analysis, and natural language processing. Future work aims to enhance sentiment analysis accuracy and expand the models’ scope to encompass more languages.

The length of global coastline is about 356 thousand kilometers with various dynamic natural and anthropogenic. Although the number of studies on coastal landscape categorization has been increasing, it is still difficult to distinguish precisely them because the used methods commonly are traditional qualitative ones. With the leverage of remote sensing data and GIS tools, it helps categorize and identify a variety of features on land and water based on multi-source data. The aim of study is using different natural - social profile data obtained from ALOS, NOAA, and multi-temporal Landsat satellite images as input data of the convolutional-neural-network (CvNet) models for coastal landscape classification. Studies used 900 cut-line samples which represent coastal landscapes in Vietnam for training and optimizing CvNet models. As a result, nine coastal landscapes were identified including: deltas, alluvial, mature and young sand dunes, cliff, lagoon, tectonic, karst, and transitional landscapes. Three CvNet models using three different optimizer types classified the landscapes of other 1150 cut-lines in Vietnam with the accuracies about 98% and low loss function value. Excepting dalmatian, karst and delta coastal landscapes, five others distribute heterogeneous along the coasts in Vietnam. Therefore, the evaluation of additional natural components is necessary and CvNet model have ability to update new landscape types in variety of tropical nation as a step toward coastal landscape classification at both national and global scales.

The natural ecosystem incorporates thousands of plant species and distinguishing them is normally manual, complicated, and time-consuming. Since the task requires a large amount of expertise, identifying forest plant species relies on the work of a team of botanical experts. The emergence of Machine Learning, especially Deep Learning, has opened up a new approach to plant classification. However, the application of plant classification based on deep learning models remains limited. This paper proposed a model, named PlantKViT, combining Vision Transformer architecture and the KNN algorithm to identify forest plants. The proposed model provides high efficiency and convenience for adding new plant species. The study was experimented with using Resnet-152, ConvNeXt networks, and the PlantKViT model to classify forest plants. The training and evaluation were implemented on the dataset of DanangForestPlant, containing 10,527 images and 489 species of forest plants. The accuracy of the proposed PlantKViT model reached 93%, significantly improved compared to the ConvNeXt model at 89% and the Resnet-152 model at only 76%. The authors also successfully developed a website and 2 applications called 'plant id' and 'Danangplant' on the iOS and Android platforms respectively. The PlantKViT model shows the potential in forest plant identification not only in the conducted dataset but also worldwide. Future work should gear toward extending the dataset and enhance the accuracy and performance of forest plant identification.