“A Deep Learning Framework for Spatio-Temporal Wind Forecasting and Offshore Wind Potential Assessment in the Gulf of Tonkin”
Nguyễn Văn Thái, Van Truong VU, Hien Nguyen THI;
GIS-IDEAS Oral Presentation
Offshore wind energy is a strategic component of sustainable energy development, where reliable wind forecasting is essential for long-term planning, grid integration, and investment decision-making. The Gulf of Tonkin exhibits substantial offshore wind potential but is characterized by complex wind variability driven by monsoonal circulation and frequent tropical cyclones. Such conditions introduce significant uncertainty for wind resource assessment, particularly in offshore regions where in-situ observations are sparse and expensive. While numerical weather prediction models are effective for short-term forecasting, they are computationally intensive and less suitable for long-term planning-oriented applications. This study proposes a spatio-temporal wind forecasting framework based on long-term atmospheric reanalysis data, designed to support offshore wind resource assessment rather than short-term operational weather prediction. A hybrid deep learning model integrating Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks is developed to capture both spatial wind structures and long-term temporal dependencies from multi-decadal wind fields. The CNN component extracts regional spatial patterns from gridded wind data, while the LSTM component models temporal evolution and seasonal variability, enabling data-driven estimation of future wind behavior. The model is trained and validated using a rolling forecasting strategy to ensure temporal generalization and to mitigate data leakage. Performance is evaluated using standard statistical metrics such as RMSE, MAE, and R², demonstrating improved predictive stability compared with standalone CNN, LSTM, and traditional machine learning approaches. Forecasted wind fields are used to compute wind speed and Wind Power Density (WPD), which are subsequently integrated into a Geographic Information System (GIS) for spatial analysis and offshore wind potential assessment. Spatial distribution maps of mean wind speed, WPD, and interannual variability are generated to identify high-potential zones and quantify uncertainty levels. By shifting from purely historical analysis to a forecasting-based assessment, the proposed approach reduces uncertainty in site selection and provides practical support for offshore wind planning and risk-aware decision-making. Experimental results confirm that the CNN–LSTM framework effectively captures monsoonal dynamics and extreme wind events, offering a scalable and computationally efficient solution for long-term offshore wind energy assessment in data-scarce coastal regions.
Keywords: Offshore wind forecasting; ERA5 Reanalysis; Spatiotemporal Deep Learning; CNN-LSTM; Gulf of Tonkin; GIS Zoning; Wind Power Density.
“A GIS-Based Framework for Evaluating Spatiotemporal Traffic Density in Hanoi”
Nguyen Thi Minh Hien;
GIS-IDEAS Oral Presentation
Traffic congestion in Hanoi has become a significant obstacle to the city's strategic goals of sustainable urban development. As the urban area undergoes rapid expansion, pressure on existing infrastructure necessitates advanced analytical methods to monitor and mitigate mobility constraints. This study is designed to systematically analyze the current state of traffic, specifically to clarify the complex spatio-temporal distribution patterns of urban traffic density. Unlike traditional studies relying on local sensors and hardware dependence, this research leverages the power of Big Data analytics to provide a more comprehensive and scalable view of urban mobility. The methodological framework begins with the collection of high-precision empirical data. The input data included four levels of traffic density collected from Google Maps, classifying congestion into distinct intensity states, along with real-time movement metrics obtained via the Distance Matrix API. By utilizing these diverse data streams, the study overcame the limitations of traditional measurement devices, such as inductive loops or static cameras, which are often costly to maintain and geographically limited. These datasets were then subjected to spatial analysis techniques, specifically density estimation to visualize vehicle concentrations, hotspot identification to detect statistically significant congestion clusters, and surface interpolation to model continuous traffic conditions across the entire urban area. The output was density maps depicting the development of traffic bottlenecks with high accuracy. These visualizations allow for a better understanding of how congestion changes spatially and temporally, revealing the daily rhythm of the city's transport network. Beyond mere observation, this research establishes a solid scientific foundation for the next generation of Intelligent Transportation Systems (ITS). By providing a quantitative basis for assessing external environmental impacts—such as carbon emissions associated with vehicle stop time—this research serves as a crucial catalyst for infrastructure optimization. Furthermore, the integration of real-time API data lays the groundwork for AI-enhanced urban forecasting, enabling planners to predict congestion before it occurs. These advances are essential to fostering smart, sustainable mobility ecosystems, improving quality of life, and ensuring the long-term sustainability of the modern urban environment.
“A uniform sampling and band combination method for land cover mapping”
Trinh Thanh;
GIS-IDEAS Oral Presentation
Land cover classification is an important field of object recognition in remote sensing imagery. Machine learning models are widely used for high-resolution land-cover mapping, but their performance strongly depends on input data. In other words, the quantity and quality of samples determine the accuracy of the models and the reliability of land cover maps. In this paper, we first present the band combination method for identifying object types in imagery. Then, we conduct a study on the uniform distribution of object types in a study area. A Sentinel-2 imagery is employed to produce a land-cover map for Van Yen area, Lao Cai province, covering 1,400 km².
Different bands of the imagery were combined to identify five types of land cover: rivers and water surfaces; barren lands; populated areas and roads; sparse vegetation; and dense vegetation. To study the uniform distribution of sample collection within the study area, we divided the area into 3x3 km grids. For each grid, we collected samples of all types to ensure even distribution and coverage within the study area. We obtained 1726 samples for 5 types of objects (equivalent to 1.1 km²).
Two methods, Rule-based and Random Forest (RF), were applied to build land-cover classification models. 70% of the objects were used for training, and the results showed that RF outperformed Rule-based in terms of accuracy. To further investigate data distribution, we conducted further experiments. We first selected a single 3x3km grid with the highest number of objects of different types to build an RF model, and used the remaining objects for testing. RF yielded 92% accuracy. Next, we selected all objects from any three grids for training; the RF result was 96%. Finally, we used 10% of the objects in each grid to train the model, achieving 98% accuracy.
RF results show that the object types labeled using the band-combination method are highly accurate. Moreover, the results show that dividing the study area into grids for object collection is crucial for RF performance. We established a land cover map with only 10% of the objects for training. Furthermore, we compared the established land cover map with official statistical data and found that vegetation types differ by less than 1% from the actual statistical area. These findings confirm that the proposed uniform sampling strategy is highly effective for producing an accurate and reliable land cover map.
“An integrated FAHP, machine learning, and GIS approach for hospital site selection: A case study of five wards in Tuyen Quang province, Vietnam”
Bui Ngoc Tu;
GIS-IDEAS Oral Presentation
Hospital site selection is a strategic planning task that directly affects the effectiveness, quality, and equity of health-service delivery, and it is therefore critical to hospital infrastructure development. This study presents an integrated approach combining fuzzy AHP (FAHP), machine learning, and GIS to identify suitable hospital sites in a five-ward study area in Tuyen Quang Province, Vietnam. The workflow captures expert judgment through FAHP and complements it with data-driven learning from machine-learning models to enhance the robustness of multi-criteria suitability assessment. Twelve criteria across economic, social, and environmental dimensions were compiled as thematic GIS layers, including distance to existing hospital, distance to industrial zone, distance to residential area, distance to main road, distance to local road, slope, distance to landfill, distance to cemeterie, distance to historical and cultural site, distance to surface water, distance to power substation, and land use type. The study area was discretized into a 300 × 300 m grid, yielding 2,101 centroid samples, from which predictor values were extracted to construct the training and validation dataset. Three algorithms were evaluated, including Random Forest (RF), Gradient Boosting (GB), and eXtreme Gradient Boosting (XGBoost). The predictive results (R² = 0.86 for RF, 0.95 for GB, and 0.96 for XGBoost) indicate that XGBoost achieved the best predictive performance. Feature importance analysis further suggests that distance to industrial zone has the strongest influence on hospital site suitability, followed by distance to main road, slope and distance to existing hospital. The resulting suitability map delineates five suitability classes and supports the identification and prioritization of candidate locations for hospital planning. Specifically, 33.44 km² (13.83%) of the study area was classified as very highly suitable, 39.77 km² (16.44%) as highly suitable, 73.78 km² (30.51%) as moderately suitable, and the remaining 94.88 km² (39.22%) was classified as low to very low suitability. Overall, the proposed approach provides a robust and practical decision-support tool for hospital siting. It offers a useful reference for local authorities in Tuyen Quang to guide sustainable hospital infrastructure planning and can be adapted to other areas with similar planning contexts.
Keywords: Hospital site selection, Machine learning, GIS, FAHP.
“Application of GeoInformatics for Monitoring Agricultural Drought in Gia Lai Province, Vietnam Using MODIS Data and Google Earth Engine”
Ngoc_Anh_Nguyen;
GIS-IDEAS Poster Presentation
Agricultural drought is one of the most severe environmental challenges affecting agricultural productivity and livelihood sustainability in Vietnam’s Central Highlands, a region characterized by complex topography, extensive rainfed agriculture, and a pronounced monsoonal climate with a rainy season from May to October and a dry season from November to April. In recent decades, increasing climate variability and prolonged dry periods have intensified drought impacts, highlighting the need for effective spatio-temporal monitoring approaches in data-scarce regions. This study applies a GeoInformatics-based framework to analyze the dynamics of agricultural drought in Gia Lai Province (prior to administrative reorganization) during the period 2005–2024.
The analysis is based on MODIS MOD09A1 surface reflectance data, which provide 8-day composite observations at 500 m spatial resolution and ensure long-term consistency for environmental monitoring. The dataset was processed using the Google Earth Engine (GEE) cloud-computing platform, enabling efficient handling of large multi-temporal satellite archives. After quality screening and cloud masking, a suite of vegetation- and moisture-related indices was derived, including the Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Moisture Stress Index (MSI), Normalized Difference Drought Index (NDDI), and Vegetation Condition Index (VCI). Monthly mean values were calculated to characterize seasonal drought variability, while long-term mean composites were generated to examine spatial patterns of drought conditions across the province.
The results reveal a strong and consistent seasonal drought signal. Vegetation vigor and surface moisture conditions decline during the dry season (November–April), with the most severe drought conditions occurring in the late dry months from February to April, as reflected by low NDVI and NDWI values and elevated MSI and NDDI values. In contrast, substantial recovery is observed during the rainy season from May to October, with improved vegetation conditions and moisture availability indicated by higher NDVI, NDWI, and VCI values. Interannual variability is evident throughout the study period, with several years exhibiting persistent drought conditions consistently captured across multiple indices. Spatial analysis highlights recurrent drought-prone areas, particularly in zones dominated by rainfed agriculture and limited irrigation infrastructure.
This study demonstrates that the integration of satellite remote sensing, GIS-based spatial analysis, and cloud-based computing constitutes an effective GeoInformatics approach for agricultural drought monitoring. The proposed framework provides valuable information to support drought assessment, agricultural planning, and climate change adaptation strategies in monsoon-influenced regions of Vietnam.
“Assessing Environmental Vulnerability to Landslides Using GIS and Unsupervised Clustering: A Case Study of the Non Nuoc Cao Bang Global Geopark, Vietnam”
Phan Hoa;
GIS-IDEAS Oral Presentation
Landslides are a major environmental hazard in mountainous regions, frequently causing extensive degradation of land cover, soil stability, hydrological systems, and ecological integrity, thereby posing significant challenges to environmental conservation and sustainable land management. These challenges are particularly pronounced in geopark areas, where geological heritage, biodiversity, and landscape values are closely interconnected and highly sensitive to disturbance. This study develops a GIS-based framework for assessing landslide-induced environmental damage using unsupervised clustering algorithms, with the Non Nuoc Cao Bang Global Geopark in northern Vietnam selected as a representative case study. A comprehensive indicator system was constructed to characterize environmental damage from multiple dimensions, including anthropogenic disturbance, vegetation and land cover degradation, soil surface conditions, hydrological impacts, and ecological sensitivity. The indicators were derived from multispectral remote sensing imagery, digital elevation model–based terrain analysis, and GIS-based spatial proximity assessments, allowing both natural and human-related factors influencing environmental vulnerability to be integrated within a unified analytical framework. Prior to analysis, all indicators were normalized to ensure comparability across different units and value ranges. Unsupervised clustering techniques, specifically K-means and hierarchical clustering, were then applied to classify the study area into zones exhibiting similar levels of environmental damage, enabling spatial patterns to emerge directly from the data without subjective weighting. The clustering results reveal distinct spatial patterns of environmental degradation across the geopark, clearly differentiating areas experiencing severe landslide-induced damage from zones characterized by relatively lower levels of environmental disturbance. These spatial patterns show strong correspondence with documented landslide occurrences and the underlying geomorphological characteristics of the region, indicating that the proposed approach effectively captures the spatial heterogeneity of landslide impacts on environmental systems. In addition, the results provide insights into the interactions between terrain conditions, land cover changes, hydrological responses, and human activities in shaping environmental damage patterns within geopark environments. Overall, the proposed framework offers an objective, reproducible, and scalable method for assessing landslide-induced environmental damage using widely available geospatial datasets, and the findings provide valuable scientific support for landslide risk zoning, environmental monitoring, and sustainable management in geopark and mountainous regions.
“Assessing Factor Effects on Land-Use Plan Implementation Using Explainable AI (XAI) in Son La Province, Vietnam”
Bui Ngoc Tu;
GIS-IDEAS Oral Presentation
Bui Ngoc Tu1, Tran Quoc Binh1
1University of Science, Vietnam National University, Thanh Xuan, Hanoi, Vietnam
Email: buingoctu@hus.edu.vn
ABSTRACT
Assessing the factors that influence land-use plan implementation outcomes is essential for improving the effectiveness of land administration and implementation management. However, it remains challenging when the impacts of multiple factor groups must be quantified for individual planned parcels on maps. With the growing availability of multisource datasets and increasing demand for spatially detailed assessments, machine-learning models combined with explainable artificial intelligence (XAI) provide a suitable approach to support objective, transparent, and verifiable evaluation.
This study proposes and tests a machine-learning workflow to assess both the relative importance and the directional influence of factors on local land-use plan implementation outcomes. The dataset was compiled from multiple sources, including thematic maps, technical reports/documents, and expert knowledge, yielding 1,406 parcel-level samples and 28 explanatory features. The features were grouped into: (i) land-category and planned-parcel characteristics, (ii) spatial proximity variables, (iii) topographic conditions, (iv) socio-economic factors, and (v) legal factors. This multisource integration strengthens quantitative representation and enhances objectivity, while still incorporating domain expertise for context-specific factors.
Five models were evaluated: Random Forest (RF), Extremely Randomized Trees (ET), Gradient Boosting (GB), XGBoost, and LightGBM. The results indicate that RF achieved superior overall performance and produced a more plausible distribution of feature importance, better reflecting practical land-use plan implementation conditions. To interpret the RF model, SHAP was employed to provide explainable insights that go beyond ranking feature importance by also characterizing the direction of influence of each factor on implementation outcomes. SHAP results identify the most influential determinants as land category, distance to residential areas, distance to schools, planned-parcel area, the Y-coordinate (capturing a north–south spatial gradient), and the intensity of inspection and compliance monitoring. The SHAP-based interpretation further reveals whether these factors tend to increase or decrease the likelihood of successful plan implementation, thereby supporting spatially explicit explanations of underlying mechanisms rather than relying solely on importance rankings.
Overall, this study advances a novel parcel-level approach for evaluating determinants of land-use plan implementation outcomes, shifting from qualitative judgment to data-driven, transparent, and verifiable assessment. The findings can support local governments in identifying key drivers, prioritizing resources, refining implementation measures, and strengthening management instruments—particularly in spatially sensitive areas—thereby improving objectivity, reliability, and the overall effectiveness of land-use plan implementation in practice.
Keywords: Land-use plan implementation, Explainable artificial intelligence , Son La Province
“Assessing Flood Inundation and Agricultural Damage from Extreme Typhoon-Induced Flooding Using SAR Data”
Pham Thi Thanh Hoa;
GIS-IDEAS Poster Presentation
Vietnam is highly vulnerable to extreme flooding due to situated in an active cyclonic region and the increasing impacts of climate change. In September 2024, Super Typhoon Yagi caused widespread flooding in the Red River Delta, including the capital city of Hanoi. This event is considered an extreme natural disaster in history, which resulted in severe socio-economic losses and exposed limitations in conventional flood monitoring approaches. A critical challenge during the immediate response phase was the lack of synoptic, near-real-time visual data.
Remote sensing offers effective benefits in disaster studies with its image acquisition repeat cycle and wide coverage capabilities. However, passive optical remote sensing satellites, such as Landsat 8/9 and Sentinel-2, rely on solar illumination and cloud-free atmospheric conditions, which severely limit their effectiveness during extreme weather events. This limitation underscores the indispensable role of Synthetic Aperture Radar (SAR) - systems transmit their own signals and record the backscatter, allowing them to penetrate clouds, and to operate day or night. This study conducts a flood assessment in Hanoi, using archived Sentinel-1 C-band Synthetic Aperture Radar (SAR) data, which enables reliable flood detection under persistent cloud cover and extreme weather conditions. Google Earth Engine was used to process data, from correcting for noise with Sar image to identifying changes in time before and after floods, thereby determining the flooded area.
The main objectives of this study are to: (1) delineate the spatial extent of inundation after the Yagi flood event; (2) estimate flooded areas and (3) estimate damage to agriculture using the European Space Agency (ESA) WorldCover 10 m data. The results demonstrate the effectiveness of SAR-based flood mapping for rapid inundation assessments. This study provides critical spatial evidence to support flood risk zoning and informs flood mitigation and land-use planning strategies in the Red River Delta during the development period.
“Assessment of Rice Growth at Different Phenological Stages Based on UAV Multispectral and High-Resolution UAV RGB Imagery”
Le Van Canh;
GIS-IDEAS Oral Presentation
The Rice growth at different stages play a critical role in determining yield production. Timely and spatially detailed monitoring of rice growth conditions is therefore essential for supporting precision agricultural management and site-specific decision-making. This study investigates the effectiveness of integrating unmanned aerial vehicle (UAV) multispectral imagery and high-resolution UAV RGB imagery to assess rice growth conditions across key developmental stages, including tillering, heading, and milk ripening, within an experimental rice cultivation area. UAV multispectral and RGB data were acquired at the three growth stages. Multispectral imagery was used to derive the Normalized Difference Vegetation Index (NDVI), which serves as an indicator of vegetation vigor, biomass accumulation, and photosynthetic activity. The spatial distribution of NDVI was analyzed at both field and each plot levels to evaluate variations in rice growth conditions across different plots. Beside that, high-resolution UAV RGB orthomosaic images were employed to visually interpret and identify the underlying causes of spatial heterogeneity observed in NDVI values. The results indicate clear variations in NDVI among the examined growth stages. The average NDVI reached 0.84 during the tillering stage, decreased slightly to 0.81 at the heading stage, and further declined to 0.76 during the milk ripening stage, reflecting a gradual reduction in chlorophyll content and photosynthetic activity as rice plants progressed toward maturity. Significant intra-plot variability in NDVI was also observed, indicating non uniform growth conditions within individual fields. The integration of high-resolution UAV RGB imagery proved effective in explaining these spatial NDVI variations by providing detailed visual information on crop structure and field conditions, such as uneven plant density, localized stress, or areas affected by pests or diseases. While multispectral NDVI analysis quantitatively captured overall growth status, RGB imagery complemented this information by enabling qualitative interpretation of specific causes of variability. Compared with the use of multispectral data alone, the integrated approach improves the reliability and interpretability of rice growth assessment. Overall, the combined use of UAV multispectral and RGB imagery provides a detailed, intuitive, and effective method for monitoring rice development at key growth stages, supporting precision crop management and contributing to improved rice yield and production sustainability.
“Assessment of Seasonal Land Surface Temperature Variations in Hanoi Using Remote Sensing Imagery”
Nguyen Trung Nghia;
GIS-IDEAS Oral Presentation
Parallel to its rapid urbanization, Hanoi is experiencing an intensified Urban Heat Island (UHI) effect, driven by anthropogenic heat from transportation and industrial activities that elevate Land Surface Temperature (LST). This research aims to identify regions adversely affected by thermal radiation and to provide a quantitative analysis of temperature differentials between areas with high versus low vegetation cover. This research leverages high-resolution (10m) Sentinel-2 satellite imagery to extract the Normalized Difference Vegetation Index (NDVI), integrated with MODIS thermal data selected for its superior temporal resolution on the Google Earth Engine cloud computing platform. Utilizing descriptive statistics and Hotspot Analysis, the study identified high-temperature nuclei (hotspots) with a confidence level exceeding 95%. These areas predominantly coincide with regions of high building density and extensive impervious surface cover. Conversely, coldspots were recorded in suburban districts characterized by significant water bodies, such as the Red River and West Lake, and abundant vegetation. Notably, the statistical results demonstrate a significant inverse correlation between the Normalized Difference Vegetation Index (NDVI) and Land Surface Temperature (LST). In areas with high NDVI values, average surface temperatures were observed to be 2°C to 3°C lower than in sparsely vegetated zones. Seasonal temperature variations further reveal that the Urban Heat Island (UHI) effect peaks during the summer, with the thermal gradient between urban and suburban areas reaching up to 5°C. In contrast, this differential diminishes during winter but remains stable within industrial manufacturing zones. These findings not only corroborate the pivotal role of vegetation cover in thermal regulation but also provide empirical scientific evidence regarding the impact of urbanization on the climate. This study serves as a critical baseline dataset, empowering policymakers to make informed decisions within the framework of sustainable planning. Optimizing green space density based on these data represents a strategic intervention to regulate urban temperatures, thereby bolstering the city’s adaptive capacity and enhancing its resilience against the increasingly complex and extreme manifestations of future climate change.
“Automated Evacuation Route Generation for Citywide Tephra-Fall Evacuation: An Imminent Eruption Scenario in Kagoshima City, Japan”
Andreas Keler;
GIS-IDEAS Oral Presentation
Large-scale volcanic eruptions can produce widespread tephra (volcanic ash and pumice) fall that rapidly degrades urban lifelines, including transport, water, and telecommunications. In such conditions, Kagoshima City, with the active Sakurajima volcano, may require a coordinated, citywide evacuation to minimize casualties, particularly under east-wind conditions that threaten densely populated western districts. This study proposes a scalable workflow to operationalize Kagoshima City’s Massive Tephra Fall Response Plan by automatically generating evacuation movements for the full household population (>300,000 households), while integrating background traffic and public transport operations.
As a preprocessing step, we translate the guideline information into explicit spatial scenario inputs by defining evacuation zones and destinations and constructing wind-direction-dependent tephra hazard wedge polygons. These polygons are used to select applicable destinations and corridor constraints, and to generate alternative routes where guideline road-name sequences are incomplete or where disruptions require detours.
Using recent census-based household counts, we synthesize household-level origins and assign each household a passenger-car trip, consistent with local guidance. Evacuation routes and destinations outside city boundaries are generated on a routable OpenStreetMap-derived network, and public transport services are represented using GTFS schedules and stop locations. To translate aggregated census data into network demand, we compare two interpolation strategies for distributing households to road edges: (a) proportional allocation by edge length, and (b) distance-weighted allocation from district centroids. We further incorporate wind-direction-dependent ash-fall considerations to contrast decision-making under an imminent-eruption (pre-emptive) evacuation versus an already ongoing eruption, where disruption and compliance constraints may differ.
Finally, we conduct microscopic traffic-flow simulation with more than 400,000 simulated road users to quantify evacuation performance and identify congestion and conflict-prone locations, enabling a joint discussion of evacuation efficiency and road-traffic safety. The results provide an evidence-based evaluation of the current municipal plan and highlight the traffic control measures required for an organized citywide evacuation.
“Climate Chnage Impacts On Robusta Coffee Suitability In Central Highland, Vietnam: Insights From CMIP6 Climate Projections”
Nguyễn Nữ Kim Linh;
GIS-IDEAS Oral Presentation
Globally, coffee plays an important role in economic development and provides significant public health benefits for society. Viet Nam is one of the major coffee-exporting countries in the world and currently ranks as the second largest coffee producer, with the Central Highlands is the primary growing area where approximately 95% of Vietnam’s Robusta Coffee production is concentrated. However, coffee is one of the crops most vulnerable to climate change, since even small variations in precipitation and temperature, both in magnitude and spatial distribution, play an important role in determining suitable conditions for coffee growth. Under the impacts of climate change, the natural suitability area of Robusta Coffee is expected to change significantly. This study evaluates the impacts of climate change on the spatial distribution of Robusta Coffee under different future climate scenarios from climate projections based on Shared Socioeconomic Pathways (SSPs). Specifically, we used 4 precipition and temperature indices obtained from 3 climate models outputs under SSP2-4.5, SSP3-7.0 and SSP5-8.5 emission scenarios. These scenarios were simulated using General Circulation Models (GCMs) from the Coupled Model Intercomparison Project Phase 6 (CMIP6). Bias correction and spatial downscaling techniques were applied to precipitation and temperature data to obtain climate information suitable for the Central Highlands region. Key environmental factors influencing coffee suitability were weighted using the Analytic Hierarchy Process (AHP) method. The weighted factors were then integrated within a Geographic Information System (GIS) environment to assess and map the natural suitability areas of Robusta Coffee. The results show the spatial distribution of suitable areas for coffee cultivation under future climate scenarios compared to the historical period. Differences among SSP2-4.5, SSP3-7.0, and SSP5-8.5 are observed in both mid-century and late-century periods. These changes highlight the potential impacts of climate change on coffee production in the Central Highlands and provide useful scientific information for adaptation planning and sustainable coffee development in the region.
Keywords: AHP, Climate change, CMIP6, GIS, Robusta Coffee, SSPs.
“Comparative Evaluation of Urban Expansion Mapping Methods in Diriyah Using GHSL, NDBI, and Unsupervised Classification”
Muhannad Mohammed Alfehaid;
GIS-IDEAS Oral Presentation
Accurate monitoring of urban expansion in dryland environments remains a persistent challenge in geospatial science due to the strong spectral similarity between built-up surfaces and highly reflective bare soils. This challenge is particularly pronounced in rapidly developing heritage-sensitive regions, where reliable urban mapping is essential for sustainable planning and decision support. This study contributes to the transition from traditional geoinformatics toward predictive geodata science by empirically evaluating and integrating multiple geospatial approaches for urban expansion mapping in Diriyah, Saudi Arabia—an area experiencing accelerated transformation under the national Vision 2030 development agenda.
The research assesses three widely applied built-up mapping methods across three temporal snapshots (2015, 2020, and 2025): the Global Human Settlement Layer (GHSL), the Normalized Difference Built-up Index (NDBI), and unsupervised k-means classification derived from multispectral satellite imagery. These methods represent different paradigms of geospatial analysis, ranging from globally calibrated products to local spectral indices and data-driven clustering. To enhance interpretability and methodological insight, the study introduces a Hybrid Built-up Detection Model (HBDM), which synthesizes the outputs of GHSL, NDBI, and unsupervised classification into a continuous urban intensity surface. Rather than acting as an independent classifier, the hybrid model functions as a diagnostic and integrative layer that highlights spatial convergence and divergence among methods.
Results reveal substantial discrepancies in estimated built-up areas. GHSL provides conservative yet temporally consistent estimates that align with known patterns of urban growth and infrastructure development. In contrast, NDBI and unsupervised classification significantly overestimate built-up extents, largely due to confusion between impervious surfaces and bright bare soils typical of dryland settings. Accuracy assessment based on reference samples confirms GHSL as the most reliable standalone method, while the HBDM layer improves spatial interpretation by balancing global calibration with local spectral and spatial cues.
The study demonstrates how integrating multiple geospatial data sources and analytical techniques can bridge descriptive urban mapping and predictive geodata analysis. From a practical perspective, the findings support urban planning, heritage conservation, and evidence-based decision-making in dryland regions undergoing rapid transformation. Methodologically, the research advances applied geospatial science by presenting a scalable hybrid framework that aligns with emerging GeoAI and predictive modelling paradigms, offering a foundation for future high-resolution, data-driven urban monitoring in similar environments.
“Comparison of Sentinel Imagery Performance Using Continuous Wavelet Transform-based CNN model for Rice-Crop System Mapping in Suphan Buri, Thailand”
Kritchayan Intarat, Nutcha Tuphimai, Phatsachon Kuankhamnuan, and Woraman Jangsawang;
GIS-IDEAS Oral Presentation
Rice cropping system mapping is crucial for agricultural monitoring and precision agriculture. Such systems are mainly found in irrigated areas with two or more crops. The reliable of information on cropping intensity and temporal dynamics is required to support cultivation activities. Suphan Buri is a major rice-growing area in Thailand's central plain. It consists of an efficient irrigation system that benefits continuous rice cultivation. Information in this area is essential for rice management. However, cloud-prone areas are a significant issue in the tropical region and affect rice data acquisition. It leads to data gaps and uncertainty when using satellite data alone. Herein, this study integrates multi-temporal remotely sensed data with the continuous wavelet transform (CWT) and a convolutional neural network (CNN) to classify rice crop systems. Such a network improves the extraction of temporal features from satellite time-series data. Sentinel-1 and -2 (S1 and S2) are then incorporated into the experiment to compare the performance of radar- and optical-based inputs under the same classification framework. The VH polarization from S1 and the enhanced vegetation index (EVI) time series data from 2023 to 2025 are acquired. The Savitzky-Golay (SG) filter is then used to transform the 1D signal into 2D scalograms using the Morlet wavelet. The images are then fed into the CNN model and classified into four classes e.g. single-crop (SC), double-crop (DC), two-and-a-half-crop (HC), and triple-crop (TC). The results from S1 and S2 are then tested and evaluated, including accuracy, F1-score, and Kappa coefficient (K). In comparison, the CWT-based CNN model integrated with S2 imagery achieves superior performance compared to S1, reaching an accuracy of 0.892, an F1-score of 0.890, and a K of 0.855. These results reveal that the developed framework significantly improves the classification of rice cropping systems, highlighting its value as an efficient tool for rice cropping system mapping and agricultural monitoring in the cloud-prone region like Thailand.
“Integrating GIS in Climate Change Education: A Case Study at Hanoi University of Natural Resources and Environment”
Ngoc_Anh_Nguyen, Bùi Thị Phương Thùy;
GIS-IDEAS Poster Presentation
The rapid development of geospatial technologies has significantly transformed approaches to climate change research, shifting from traditional descriptive methods toward data-driven and predictive geospatial science approaches. In this context, the integration of Geographic Information Systems (GIS) into university teaching is increasingly recognized as a critical solution for fostering spatial thinking and enhancing students’ analytical capacity. Developing these competencies is particularly essential for students in climate change–related disciplines, where the ability to analyze spatial patterns, trends, and interactions plays a central role in scientific research and evidence-based decision-making.
This study was conducted to evaluate the effectiveness of integrating GIS into university teaching in improving spatial analysis skills for climate change research among undergraduate students majoring in Climate Change and Sustainable Development at Hanoi University of Natural Resources and Environment (HUNRE). The research adopted a quasi-experimental design involving 30 students from the Faculty of Climate Change and Sustainable Development. Students participated in two different learning modalities: (i) a course integrating GIS applications into climate change analysis, and (ii) a course delivered using traditional teaching methods without GIS integration.
The GIS-integrated instructional content focused on key competencies, including spatial data visualization, thematic map interpretation, and the assessment of spatial differentiation of climatic conditions. These activities were designed to enhance students’ ability to connect theoretical knowledge with spatial evidence and to support analytical reasoning in climate change research contexts.
Research data were collected through a structured questionnaire survey aimed at assessing students’ perceptions of learning effectiveness, spatial thinking development, data analysis skills, and readiness to apply GIS in climate change research. Quantitative data were analyzed using SPSS software, employing descriptive statistics, reliability testing of measurement scales, and comparative statistical analyses.
The findings indicate that the integration of GIS into university teaching has a positive and significant impact on students’ ability to read, interpret, and analyze spatial data, as well as on their confidence in applying GIS tools in academic research. The study confirms the important role of GIS-based teaching approaches and provides empirical evidence to support curriculum innovation in higher education toward strengthening the application of geospatial technologies in climate change education.
“Integrating Multi-Sensor Satellite Imagery and Machine Learning to Assess the Impact of Urban Expansion on Air Quality and Surface Urban Heat Islands”
Nguyen Manh Hung;
GIS-IDEAS Oral Presentation
Nguyen Manh Hung1,2,, Vu Anh Tuan1, Nguyen Hong Quang3, Dao Duy Toan4, Le Mai Son1, Le Thi Thu Hang1
1 Vietnam National Space Center, Vietnam Academy of Science and Technology, Hanoi, Vietnam
2 Graduate University of Science and Technology, Vietnam Academy of Science and Technology, Hanoi, Vietnam
3 Technology Multimedia Technology Department, Multimedia Faculty, Posts and Telecommunications Institute of Technology
4 Faculty of Bridge and Road, Hanoi University of Civil Engineering, Hanoi, Vietnam
E-mail:nmhung@vnsc.org.vn
Rapid urbanization in emerging megacities is intensifying dual environmental crises: localized microclimate shifts and air quality degradation. This study establishes an integrated analytical framework using multi-sensor satellite data and machine learning to decode the synergistic nexus between urban sprawl, surface urban heat islands (SUHI), and atmospheric pollution in Hanoi, Vietnam. Leveraging the computational power of Google Earth Engine (GEE), we synthesized a multi-temporal dataset including Landsat (for LST, SUHI, and LULC), MODIS (for long-term Aerosol Optical Depth - AOD trends), and Sentinel-5P TROPOMI (for NO2 and SO2 concentrations). A Random Forest (RF) algorithm was implemented for land cover classification, incorporating spectral bands alongside NDVI, NDWI, NDBI, and Digital Elevation Model (DEM) data to mitigate topographic effects and enhance classification robustness. The model achieved superior performance with an overall accuracy (OA) exceeding 98% and a Kappa coefficient above 0.95. Spatiotemporal results reveal that the uncontrolled expansion of impervious surfaces, primarily in western and southern Hanoi, acts as the primary driver for thermal and pollution "hotspots." The conversion of agricultural land and water bodies into high-density built-up areas has led to a significant increase in Land Surface Temperature (LST), with peak SUHI intensities observed within industrial clusters and along radial transport axes. Statistical analysis confirms a strong positive correlation between built-up density, SUHI intensity, and pollutant accumulation. These findings highlight the severe consequences of green space loss and "heat-trapping" effects within dense urban structures, where stagnant atmospheric conditions exacerbate pollutant concentrations. Furthermore, the integration of Sentinel-5P data identified clear spatial overlaps between thermal anomalies and NO2 emission plumes, particularly in the vicinity of coal-fired power plants and key transport corridors. These empirical findings provide a robust scientific basis for the "Urban-Heat-Air" nexus. The study advocates for the strategic implementation of nature-based solutions, such as green infrastructure and ventilation corridors, to enhance ecological resilience and mitigate adverse urban microclimate impacts in Hanoi amidst global climate change.
Keywords: Urban expansion, Air Quality, SHUI, Machine Learning, Multi-Sensor Satellite.
“Quantifying the contribution of topographic and spectral indices to LULC classification accuracy: a Random Forest-based feature selection study”
Nguyen Manh Hung;
GIS-IDEAS Oral Presentation
Quantifying the contribution of topographic and spectral indices to LULC classification accuracy: a Random Forest-based feature selection study
Vu Anh Tuan1, Nguyen Manh Hung1,2,, Ngo Duc Anh1,2, Phạm Công Giang1, Nguyen Thi Phuong Hao1
1 Vietnam National Space Center, Vietnam Academy of Science and Technology, Hanoi, Vietnam
2 Graduate University of Science and Technology, Vietnam Academy of Science and Technology, Hanoi, Vietnam
E-mail: nmhung@vnsc.org.vn
Abstract:
Accurate Land Use and Land Cover (LULC) mapping is essential for sustainable urban planning and environmental monitoring. However, spectral similarity between diverse land cover types, such as barren land and impervious surfaces, or urban shadows and water bodies, often leads to significant classification inaccuracies when relying solely on raw optical bands. This study proposes a robust analytical framework to quantify the individual and collective contributions of spectral indices and topographic variables to LULC classification performance using the Random Forest (RF) algorithm. Leveraging the Google Earth Engine (GEE) cloud-computing platform, we synthesized a multi-temporal dataset for Hanoi, Vietnam, integrating Landsat-8 optical imagery with a 30-meter Digital Elevation Model (DEM). The methodology employed a systematic feature selection approach by evaluating three distinct scenarios: (1) raw spectral bands; (2) integration of spectral indices (NDVI, NDWI, NDBI); and (3) a comprehensive fusion of spectral, index, and topographic data (elevation and slope). To quantify variable contribution, the Random Forest model’s internal importance mechanism, based on Mean Decrease Gini and Mean Decrease Impurity, was utilized to rank the significance of each input feature. Our results demonstrate that the synergy of spectral indices and topographic data significantly enhances classification robustness. The optimized RF model achieved a superior Overall Accuracy (OA) of 98.3% and a Kappa coefficient of 0.96. Feature importance analysis revealed that while raw spectral bands provide the foundational data, the NDBI and Elevation variables emerged as the most influential contributors to model precision. Specifically, NDBI was crucial for accurately delineating high-density built-up areas, while the inclusion of DEM and slope effectively mitigated topographic-induced spectral confusion in the city’s peripheral hilly regions. This research underscores that a multi-feature approach, prioritizing variables that minimize spectral overlap, is essential for high-precision LULC mapping in complex, rapidly urbanizing landscapes. These findings provide a scientifically rigorous template for selecting optimal feature combinations, improving the reliability of Earth observation products using accessible optical and terrain data without the need for active sensor integration.
Keywords: LULC classification, Random Forest, Feature importance, Spectral indices, DEM.
“Seasonal Variation Assessment of Land Surface Temperature Based on Land Cover Types in Gia Lai Province, Vietnam”
Eaint Chit Phone;
GIS-IDEAS Oral Presentation
Seasonal variation in land surface temperature (LST) is a key indicator of land-atmosphere interactions, reflecting the combined influence of climate conditions and land cover characteristics. This study examines seasonal land surface temperature variations across land cover types in a tropical highland region of Vietnam using Landsat 8 OLI/TIRS data acquired during the dry and wet seasons of 2024. The findings reveal pronounced land cover–driven thermal contrasts, demonstrating significant spatial and seasonal land surface temperature variation across Gia Lai Province that is primarily determined by land cover characteristics rather than regional climate alone. The greater spatial temperature variability is observed during the dry season, indicating differing land cover responses to climatic conditions. In both seasons, artificial surfaces consistently exhibit higher LST, while natural surfaces provide more effective thermal regulation. The results indicate that LST exhibits substantial seasonal variability that differs markedly among land cover types. The greatest seasonal LST variations are observed in agricultural and bare soil areas. Agricultural land, which occupies the largest proportion of the study area, shows markedly higher LST during the dry season than during the wet season, primarily due to reduced soil moisture, crop seasons, decreased vegetation cover, and increased surface exposure. In contrast, natural forest areas dominated by dense vegetation consistently exhibit lower and more stable LST during the wet season than in the dry season, reflecting the enhanced cooling effects of a healthy vegetation canopy through shading and evapotranspiration. Meanwhile, land cover types with relatively stable surface properties, such as built-up areas and water bodies, show minimal seasonal thermal variation, highlighting their strong thermal regulation capacity. These findings underscore the dominant role of land cover in controlling surface thermal dynamics and emphasize the importance of preserving natural surfaces, particularly vegetation and water bodies, to mitigate heat stress and enhance climate resilience in tropical highland environments.
“Spatial Heterogeneity and Determinants of the MODIS AOD–PM2.5 Relationship in a Tropical Megacity: A Case Study of Hanoi, Vietnam”
Lê Thị Thu Hằng;
GIS-IDEAS Oral Presentation
Satellite-derived Aerosol Optical Depth (AOD) is crucial for estimating ground-level fine particulate matter (PM₂.₅) over large areas, yet the AOD–PM₂.₅ relationship exhibits substantial spatial heterogeneity in complex urban environments. This study quantifies such heterogeneity and identifies its key drivers in Hanoi, Vietnam—a rapidly urbanizing tropical megacity influenced by monsoon climate. Using spatiotemporally matched Level-2 MODIS (MOD04_3K) AOD at 550 nm with hourly PM₂.₅ measurements from 11 ground stations (2019–2025), and validated against AERONET data (R² = 0.62, RMSE = 0.25, MAE = 0.19), site-specific linear regression analyses revealed marked variability in correlation strength (R²: 0.34–0.64) and sensitivity (slopes: 0.0007–0.0114 per µg/m³). Spatial interpolation of these parameters highlighted stronger linkages in the dense urban core (e.g., R² > 0.56 at central stations like Bán Đảo Linh Đàm) and weaker correlations in suburban and vegetated areas (e.g., R² = 0.34 at Công viên Nhân Chính). Multivariate analysis incorporating ERA5 meteorological variables (planetary boundary layer height [PBLH], relative humidity [RH]) and Landsat-derived built-up index (IBI) and normalized difference vegetation index (NDVI) demonstrated that built-up intensity is the primary driver of steeper slopes and stronger correlations, whereas higher vegetation cover weakens the relationship. Error metrics, such as RMSE and MAE calculated for site-specific regressions, further confirm this spatial heterogeneity by showing greater estimation accuracy in urban cores compared to greener peripheries, underscoring variations not only in trends (R²) but also in predictive precision across different urban zones. These findings emphasize the intrinsically location-specific nature of the AOD–PM₂.₅ linkage within a megacity. We conclude that single city-wide conversion models are suboptimal; future satellite-based PM₂.₅ retrievals in tropical urban environments should adopt zone-specific frameworks or integrate local environmental covariate to account for this intra-urban heterogeneity.
“STUDY OF GRAVITY VARIATIONS OVER VIETNAM AND SURROUNDING REGIONS USING TIME-VARIABLE GRAVITY MODELS”
Ngo Thi Men Thuong;
GIS-IDEAS Poster Presentation
Gravity variations directly reflect internal dynamic processes of the Earth as well as the redistribution and transport of mass on and near the surface, and therefore play an important role in studies of geodynamics, geodesy, climate change, and resource management. In the field of geodesy, the determination of temporal gravity variations is essential for maintaining national reference systems, updating Geoid models, and correcting errors in high-precision measurements. In Vietnam, research on gravity variations remains limited due to resource constraints for data acquisition and the assessment of time-dependent gravity signals. The advent of satellite missions such as GRACE and GRACE-FO has opened a new era for observing gravity variations from space with high temporal resolution.
This paper presents the results of a study and analysis of gravity field variations over Vietnam and the East Sea using time-variable gravity field models. The study employs time-variable gravity models derived from GRACE/GRACE-FO satellite gravity data provided by the International Centre for Global Earth Models (ICGEM). These Earth’s time-variable gravity field models are represented as a sequence of monthly spherical harmonic coefficients. The monthly gravity solutions enable the identification of short-term fluctuations as well as the analysis of long-term trends in gravity field variations.
The methodology focuses on the processing of multi-temporal spherical harmonic coefficients, the computation of gravity field differences, and trend analysis using geospatial data analysis techniques. Model-derived gravity variations are compared with ground-based gravity observations at selected sites, based on repeated measurements conducted during the period 2016–2020. The correlation analysis between international satellite-based data and locally acquired ground gravity measurements plays an important role in proposing strategies for integrating and exploiting multiple data sources, thereby improving the effectiveness of gravity variation studies under the current resource conditions in Vietnam.
The results indicate that the gravity field over Vietnam and surrounding regions exhibits clear temporal variations, characterized by pronounced long-term trends and distinct spatial contrasts between continental and marine areas. The findings provide valuable data and a methodological foundation for future studies on the prediction of mass redistribution and lithospheric dynamics, contributing to the modernization of the spatial data infrastructure and national geodetic reference systems.