- Research Article
- 10.1016/j.compeleceng.2026.111068
Cross-model distillation with few-shot semantic segmentation for agricultural counterfeit detection
- May 01, 2026
- Computers and Electrical Engineering
- Dat Tran-Anh + 3 more +3
Publications from 2021 to 2026
Showing 10 of 404 papers
Cross-model distillation with few-shot semantic segmentation for agricultural counterfeit detection
Transformer guided Multimodal VQA model for Fruit recognitions
Air Quality Alerts, Health Impacts, and Adaptation Implications Under Varying Climate Policy.
Without emission reductions, climate change may increase ozone and PM2.5 air pollution in the United States; however, we do not know how this will affect air quality alerts that prompt people to stay indoors. Here, we use an integrated modeling framework to find distributions of daily Air Quality Index (AQI) during the smog season at the start, middle, and end-of-century. Considering natural variability, climate change may cause air quality alerts to double (increase by a factor of 2 ± 0.2) by 2100. Days when both ozone and PM2.5 exceed alert thresholds quadruple (4.3 ± 1.2). More than 100,000,000 (±45,000,000) people experience mean air pollution deemed "Unhealthy for Sensitive Groups", a growth of 7 (±3) times compared to 2000. If people follow alerts by staying inside, they reduce exposure to outdoor-generated pollutants. Their health benefits are similar whether the alert is caused by ozone or PM2.5. Senior (age 65+) populations receive much higher benefits per day by adapting (e.g., 95CI for PM2.5: $4.60 to $147) as young adults (age 18-35; 95CI: $0.15 to $4.22)─more than 45 times higher on average. This disproportionate impact requires targeted messaging and guidance, especially as climate-related risks rise.
Read moreThong Dong/Tenang Pedagogy: A Dialogue From Global South Scholars in Nurturing Teacher Well‐Being Beyond Classrooms
ABSTRACT Teacher well‐being (TWB) is not just an individual responsibility, but a shared ecological responsibility where individuals live in unity. However, under the increasing pressures and workloads, well‐being seems to be a luxury and self‐care probably is the last option for Giáo đa đoan kiệt quệ (burnout‐multifaceted teacher). The term thong dong/tenang emerge naturally from our conversations about education and is grounded in our teaching practices. Despite our teaching and research backgrounds (i.e., Vietnam, Indonesia and the U.S.), we share the same concern about the teacher's burnout across the globe. In this turbulent time, we are pondering how we can regenerate the thong dong/tenang pedagogy toward our daily journey as multifaceted teachers working in the Global South. Using memory‐rewriting and dialogic autoethnography, our narratives braided together as we moved back and forth to explore what the thong dong/tenang pedagogy means and looks like so that we can work toward healing our own selves in the teaching and research journeys.
Read moreDevelopment and validation of finite element model of milling thin-walled part
Adaptive agro-ecological zoning for climate-resilient land use in the Mekong delta
ABSTRACT The Vietnamese Mekong Delta is increasingly exposed to dry-season freshwater scarcity and salinity intrusion, challenging static and rice-centric land-use planning frameworks. This study aims to develop an Adaptive Agro-Ecological Zoning (AAEZ) framework that translates hydrosalinity dynamics into operational land-use guidance under climate variability. The framework integrates percentile-based salinity scenarios (P50 and P85), dry-season discharge at Kratie, functional accessibility of water infrastructure (canals, sluices, pumping stations), and land suitability into a rule-based GIS zoning algorithm. Three hydrosalinity typologies are defined using salinity magnitude and exposure duration: freshwater (<1‰ for >300 days), transitional brackish (1–4‰ for 60–150 days), and high-risk saline (>4‰ for ≥90 days). Application of the AAEZ framework reclassifies the Delta into freshwater rice zones (about 38%), shrimp–rice transition zones ( about 42%), and brackish aquaculture zones (about 17%), with approximately 3% identified as priority transition areas where current land use conflicts with hydrological constraints. Salinity simulations using MIKE 11/21 were validated against observed records (RMSE < 0.6‰). District-level zoning guidance is provided for key provinces including Ca Mau, Tra Vinh, Bac Lieu, and Long An. The AAEZ framework offers a practical decision-support tool for integrating water-driven zoning into provincial land-use planning and climate adaptation strategies.
Read moreOn the end effect and end condition methods of empirical mode decomposition with cubic spline
First Agriculture Land Use Map in Vietnam Using an Adaptive Weighted Combined Loss Function for UNET++
Accurate and timely agricultural mapping is essential for supporting sustainable agricultural development, resource management, and food security. Despite its importance, Vietnam lacks detailed and consistent large-scale agricultural maps. In this study, we produced the first national-scale agricultural map of Vietnam for 2024 using a UNet++ deep learning architecture that integrates multi-temporal Sentinel-1 and Sentinel-2 imagery with Global-30 DEM data. The resulting product includes 15 land-cover categories, eight of which represent the most popular agricultural types in Vietnam. We further evaluate the model’s transferability by applying the 2024 trained model to generate a corresponding map for 2020. The approach achieves overall classification accuracies of 83.01%±1.37% (2020) and 80.09%±0.76% (2024). To address class imbalance within the training dataset, we introduced an adaptive weight combined loss function that automatically adjusts the weight of dice loss and cross-entropy loss within a combined loss function during the model training process.
Read morePostbuckling of magnetic rings
A Spectral Clustering-Based Approach for Balancing Data in TF-Target Gene Interaction Prediction Using Heterogeneous Network Embedding.
Identifying transcription factors (TFs) interacting with target genes plays a crucial role in regulating gene expression, providing deep insights into the molecular mechanisms involved in biological processes and diseases, while also opening new opportunities for research and treatment, particularly in precision medicine. However, challenges arise from the high cost of experiments and biological complexity. Recent machine learning models show promise for predicting TF target gene interactions but struggle with data imbalance, as the number of known interactions is much smaller than unknown ones, reducing prediction performance. To address this challenge, we propose an approach that integrates sample selection from potential unknown interactions with spectral clustering to balance the dataset. We construct an adjacency matrix from known interactions, apply spectral clustering to partition the data, and select unknown interactions farthest from the cluster center for diversity. A deep learning method combining random walk sampling and skip-gram embeddings is used to learn representations of biological networks. Experimental results, evaluated via five-fold cross-validation, show an average AUC of 0.9575 ± 0.0044, outperforming existing methods. These findings validate our approach's efficacy in improving prediction accuracy and resolving data imbalance, offering a robust framework for discovering novel TF-target gene interactions in molecular biology and biomedical research.
Read more