Wirba Pountianus Berinyuy
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Wirba Pountianus Berinyuy has a total of 17 co-authors. Here are the top co-authors by number of shared publications:
Ele Pierre
ENSP, Electrical and Telecommunications Engineering, University of Yaounde 1
4 shared publications- Joseph Mvogo Ngono
Laboratory of Applied Computer Science, Department of Computer Science, Faculty of Science, University of Douala, Douala, Cameroon
4 shared publications Mvogo Ngono Joseph
Laboratory of Applied Computer Science, Department of Computer Science, Faculty of Science, University of Douala
3 shared publicationsNoumsi Woguia Auguste Vigny
Laboratory of Applied Computer Science, Department of Computer Science, Faculty of Science, University of Douala
3 shared publications- Noumsi Woguia Auguste Vigny
Laboratory of Applied Computer Science, Department of Computer Science, Faculty of Science, University of Douala, Douala, Cameroon
3 shared publications Emile Tatinyuy Verdzekov
University of Douala
2 shared publications
Latest Publications
- Open Access
- 30 September 2026
- Published by Frontiers Media SA in Frontiers in Earth Science
- Vol. 14, 1895667
Scaling spatiotemporal transformers for regional food security: a tri-stream latent-dynamic approach to pre-harvest yield and price forecasting
Introduction: Pre-harvest yield forecasting and price stability prediction are critical for food security planning in climate-vulnerable regions, yet existing approaches struggle to integrate heterogeneous data streams spanning satellite imagery, historical agricultural statistics, and socioeconomic indicators. Methods: We present the Tri-Stream Latent-Dynamic Transformer (TLDT), unifying spatial-spectral Sentinel-1/2 features, temporal phenology, and exogenous economic signals via a purpose-designed cross-attention mechanism, validated on 24 Lake Chad Basin locations (2018–2023) using group-aware, leave-one-year-out cross-validation across 18 independent (country, year) units. Results: TLDT achieves RMSE = 0.360 ± 0.103 t/ha ( R 2 = −0.10 ± 0.56) for yield, comparable to — not clearly better than — a standard Transformer baseline (0.350 ± 0.057 t/ha). Price-stability classification (56.6% ± 19.2% accuracy) is statistically indistinguishable from a logistic-regression baseline on the raw exogenous features alone (58.3% ± 19.4%). Discussion: This indicates the hand-crafted price-stability label remains largely predictable from its own constituent variables — reported as an open limitation rather than a genuine forecasting capability. The Adaptive Positional Encoding and Tri-Stream Cross-Attention remain architecturally motivated but do not yet demonstrate a clear quantitative advantage on this dataset.
- 30 September 2026
- Published by Frontiers Media SA in Frontiers in Earth Science
- Vol. 14, 1895667
Open AccessScaling spatiotemporal transformers for regional food security: a tri-stream latent-dynamic approach to pre-harvest yield and price forecasting
Introduction: Pre-harvest yield forecasting and price stability prediction are critical for food security planning in climate-vulnerable regions, yet existing approaches struggle to integrate heterogeneous data streams spanning satellite imagery, historical agricultural statistics, and socioeconomic indicators. Methods: We present the Tri-Stream Latent-Dynamic Transformer (TLDT), unifying spatial-spectral Sentinel-1/2 features, temporal phenology, and exogenous economic signals via a purpose-designed cross-attention mechanism, validated on 24 Lake Chad Basin locations (2018–2023) using group-aware, leave-one-year-out cross-validation across 18 independent (country, year) units. Results: TLDT achieves RMSE = 0.360 ± 0.103 t/ha ( R 2 = −0.10 ± 0.56) for yield, comparable to — not clearly better than — a standard Transformer baseline (0.350 ± 0.057 t/ha). Price-stability classification (56.6% ± 19.2% accuracy) is statistically indistinguishable from a logistic-regression baseline on the raw exogenous features alone (58.3% ± 19.4%). Discussion: This indicates the hand-crafted price-stability label remains largely predictable from its own constituent variables — reported as an open limitation rather than a genuine forecasting capability. The Adaptive Positional Encoding and Tri-Stream Cross-Attention remain architecturally motivated but do not yet demonstrate a clear quantitative advantage on this dataset.
- Open Access
- 28 June 2026
- Published by Springer Nature in Research Square
FusionAttNet Framework for Hierarchical Attention Driven Sentinel 1 and Sentinel 2 Fusion for Semi Arid Land Cover Classification in Far North Cameroon
Accurate land cover mapping in semi-arid regions remains challenging due to spectral homogeneity of sparse vegetation, seasonal variability, and persistent cloud cover. In this study, we propose a FusionAttNet, a novel deep learning framework integrating Sentinel-1 SAR and Sentinel-2 optical data through modality-aware hierarchical attention for land cover classification case study of Far North Cameroon. Our approach employs parallel sensor-specific processing streams (generating cross-modal indices like NDVI/VV ratios) and a three-tiered attention mechanism (pixel-patch-landscape) to resolve spatial ambiguities in semi-arid landscapes. Enhanced by modality-aware augmentation and focal loss with label smoothing, FusionAttNet achieved 96.75% overall accuracy and 0.9680 F1-score on a multi-seasonal dataset over a 12-month temporal window spanning the year 2021, outperforming feature-stacking (85.1%) and early fusion (87.6%) baselines. Key innovations include: (1) landscape-level attention capturing phenological transitions in Sahelian ecotones, (2) cross-modal indices mitigating cloud-induced optical data gaps, and (3) SAR-optimized augmentation preserving backscatter textures. Results demonstrate a 14.7% reduction in misclassification of mixed bare soil/grassland interfaces compared to state-of-the-art methods, establishing a new paradigm for semi-arid land cover monitoring.
- 28 June 2026
- Published by Springer Nature in Research Square
Open AccessFusionAttNet Framework for Hierarchical Attention Driven Sentinel 1 and Sentinel 2 Fusion for Semi Arid Land Cover Classification in Far North Cameroon
Accurate land cover mapping in semi-arid regions remains challenging due to spectral homogeneity of sparse vegetation, seasonal variability, and persistent cloud cover. In this study, we propose a FusionAttNet, a novel deep learning framework integrating Sentinel-1 SAR and Sentinel-2 optical data through modality-aware hierarchical attention for land cover classification case study of Far North Cameroon. Our approach employs parallel sensor-specific processing streams (generating cross-modal indices like NDVI/VV ratios) and a three-tiered attention mechanism (pixel-patch-landscape) to resolve spatial ambiguities in semi-arid landscapes. Enhanced by modality-aware augmentation and focal loss with label smoothing, FusionAttNet achieved 96.75% overall accuracy and 0.9680 F1-score on a multi-seasonal dataset over a 12-month temporal window spanning the year 2021, outperforming feature-stacking (85.1%) and early fusion (87.6%) baselines. Key innovations include: (1) landscape-level attention capturing phenological transitions in Sahelian ecotones, (2) cross-modal indices mitigating cloud-induced optical data gaps, and (3) SAR-optimized augmentation preserving backscatter textures. Results demonstrate a 14.7% reduction in misclassification of mixed bare soil/grassland interfaces compared to state-of-the-art methods, establishing a new paradigm for semi-arid land cover monitoring.
- Open Access
- 16 April 2026
- Published by California Digital Library (CDL) in EarthArXiv
Scaling Spatiotemporal Transformers for Regional Food Security: A Tri-Stream Latent-Dynamic Approach to Pre-Harvest Yield and Price Forecasting
Pre-harvest yield forecasting and price stability prediction are critical for food security planning in climate-vulnerable regions, yet existing approaches struggle to integrate heterogeneous data streams spanning satellite imagery, historical agricultural statistics, and socioeconomic indicators. We present the Tri-Stream Latent-Dynamic Transformer (TLDT), a novel deep learning architecture that unifies spatial-spectral features from Sentinel-1/2 satellite image time series, temporal phenological dynamics, and exogenous economic signals through a purpose-designed cross-attention mechanism. Unlike conventional fusion strategies that treat data streams independently, our Tri-Stream Cross-Attention module employs spatial features as queries, economic context as keys, and temporal phenology as values, enabling the model to learn how market conditions modulate the relationship between observed crop growth and final yields. We validate TLDT on a comprehensive dataset covering 24 locations across Cameroon, Chad, and Nigeria in the Lake Chad Basin from 2018–2023, incorporating Sentinel-1/2 imagery from Google Earth Engine, CHIRPS rainfall data, FAO yield statistics, and WFP market prices. On held-out test data, TLDT achieves an RMSE of 0.244 t/ha for yield prediction (representing a 2.9% improvement over LSTM baselines and 9.3% over standard Transformers) with an R2 of 0.197, while simultaneously attaining 93.2% accuracy and 0.933 weighted F1-score on three-class price stability classification (Stable, Vulnerable, Crisis). Our Adaptive Positional Encoding component effectively handles the irregular temporal sampling inherent in cloud-affected optical satellite data. The integrated Harvest-Expectation Index provides interpretable 90-day pre-harvest forecasts suitable for early warning systems. These results demonstrate that tri-stream fusion with economic context awareness substantially improves predictive performance in data-sparse, conflict-affected agricultural regions, offering a scalable framework for operational food security monitoring.
- 16 April 2026
- Published by California Digital Library (CDL) in EarthArXiv
Open AccessScaling Spatiotemporal Transformers for Regional Food Security: A Tri-Stream Latent-Dynamic Approach to Pre-Harvest Yield and Price Forecasting
Pre-harvest yield forecasting and price stability prediction are critical for food security planning in climate-vulnerable regions, yet existing approaches struggle to integrate heterogeneous data streams spanning satellite imagery, historical agricultural statistics, and socioeconomic indicators. We present the Tri-Stream Latent-Dynamic Transformer (TLDT), a novel deep learning architecture that unifies spatial-spectral features from Sentinel-1/2 satellite image time series, temporal phenological dynamics, and exogenous economic signals through a purpose-designed cross-attention mechanism. Unlike conventional fusion strategies that treat data streams independently, our Tri-Stream Cross-Attention module employs spatial features as queries, economic context as keys, and temporal phenology as values, enabling the model to learn how market conditions modulate the relationship between observed crop growth and final yields. We validate TLDT on a comprehensive dataset covering 24 locations across Cameroon, Chad, and Nigeria in the Lake Chad Basin from 2018–2023, incorporating Sentinel-1/2 imagery from Google Earth Engine, CHIRPS rainfall data, FAO yield statistics, and WFP market prices. On held-out test data, TLDT achieves an RMSE of 0.244 t/ha for yield prediction (representing a 2.9% improvement over LSTM baselines and 9.3% over standard Transformers) with an R2 of 0.197, while simultaneously attaining 93.2% accuracy and 0.933 weighted F1-score on three-class price stability classification (Stable, Vulnerable, Crisis). Our Adaptive Positional Encoding component effectively handles the irregular temporal sampling inherent in cloud-affected optical satellite data. The integrated Harvest-Expectation Index provides interpretable 90-day pre-harvest forecasts suitable for early warning systems. These results demonstrate that tri-stream fusion with economic context awareness substantially improves predictive performance in data-sparse, conflict-affected agricultural regions, offering a scalable framework for operational food security monitoring.
- Open Access
- 10 March 2026
- Published by Springer Nature in Research Square
FusionAttNet: Hierarchical Attention-Driven Sentinel- 1/Sentinel-2 Fusion for Semi-Arid Land Cover Classification in Far North Cameroon
Accurate land cover mapping in semi-arid regions remains challenging due to spectral homogeneity of sparse vegetation, seasonal variability, and persistent cloud cover. In this study, we propose a FusionAttNet, a novel deep learning framework integrating Sentinel-1 SAR and Sentinel-2 optical data through modality-aware hierarchical attention for land cover classification case study of Far North Cameroon. Our approach employs parallel sensor-specific processing streams (generating cross-modal indices like NDVI/VV ratios) and a three-tiered attention mechanism (pixel-patch-landscape) to resolve spatial ambiguities in semi-arid landscapes. Enhanced by modality-aware augmentation and focal loss with label smoothing, FusionAttNet achieved 96.75% overall accuracy and 0.89 F1-score on a multi-seasonal dataset (2020–2023), outperforming feature-stacking (85.1%) and early fusion (87.6%) baselines. Key innovations include: (1) landscape-level attention capturing phenological transitions in Sahelian ecotones, (2) cross-modal indices mitigating cloud-induced optical data gaps, and (3) SAR-optimized augmentation preserving backscatter textures. Results demonstrate a 14.7% reduction in misclassification of mixed bare soil/grassland interfaces compared to state-of-the-art methods, establishing a new paradigm for semi-arid land cover monitoring.
- 10 March 2026
- Published by Springer Nature in Research Square
Open AccessFusionAttNet: Hierarchical Attention-Driven Sentinel- 1/Sentinel-2 Fusion for Semi-Arid Land Cover Classification in Far North Cameroon
Accurate land cover mapping in semi-arid regions remains challenging due to spectral homogeneity of sparse vegetation, seasonal variability, and persistent cloud cover. In this study, we propose a FusionAttNet, a novel deep learning framework integrating Sentinel-1 SAR and Sentinel-2 optical data through modality-aware hierarchical attention for land cover classification case study of Far North Cameroon. Our approach employs parallel sensor-specific processing streams (generating cross-modal indices like NDVI/VV ratios) and a three-tiered attention mechanism (pixel-patch-landscape) to resolve spatial ambiguities in semi-arid landscapes. Enhanced by modality-aware augmentation and focal loss with label smoothing, FusionAttNet achieved 96.75% overall accuracy and 0.89 F1-score on a multi-seasonal dataset (2020–2023), outperforming feature-stacking (85.1%) and early fusion (87.6%) baselines. Key innovations include: (1) landscape-level attention capturing phenological transitions in Sahelian ecotones, (2) cross-modal indices mitigating cloud-induced optical data gaps, and (3) SAR-optimized augmentation preserving backscatter textures. Results demonstrate a 14.7% reduction in misclassification of mixed bare soil/grassland interfaces compared to state-of-the-art methods, establishing a new paradigm for semi-arid land cover monitoring.
- Open Access
- 19 January 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (2) , 986
Eye Gaze Detection Using a Hybrid Multimodal Deep Learning Model for Assistive Technology
This paper presents a novel hybrid multimodal deep learning model for robust and real-time eye gaze estimation. Accurate gaze tracking is essential for advancing human–computer interaction (HCI) and assistive technologies, but existing methods often struggle with environmental variations, require extensive calibration, and are computationally intensive. Our proposed model, GazeNet-HM, addresses these limitations by synergistically fusing features from RGB, depth, and infrared (IR) imaging modalities. This multimodal approach allows the model to leverage complementary information: RGB provides rich texture, depth offers invariance to lighting and aids pose estimation, and IR ensures robust pupil detection. Furthermore, we introduce a personalized adaptation module that dynamically fine-tunes the model to individual users with minimal calibration data. To ensure practical deployment, we employ advanced model compression techniques, enabling real-time inference on resource-constrained embedded systems. Extensive evaluations on public datasets (MPIIGaze, EYEDIAP, Gaze360) and our collected M-Gaze dataset demonstrate that GazeNet-HM achieves state-of-the-art performance, reducing the mean angular error by up to 27.1% compared to leading unimodal methods. After model compression, the system achieves a real-time inference speed of 32 FPS on an embedded Jetson Xavier NX platform. Ablation studies confirm the contribution of each modality and component, highlighting the effectiveness of our holistic design.
- 19 January 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (2) , 986
Open AccessEye Gaze Detection Using a Hybrid Multimodal Deep Learning Model for Assistive Technology
This paper presents a novel hybrid multimodal deep learning model for robust and real-time eye gaze estimation. Accurate gaze tracking is essential for advancing human–computer interaction (HCI) and assistive technologies, but existing methods often struggle with environmental variations, require extensive calibration, and are computationally intensive. Our proposed model, GazeNet-HM, addresses these limitations by synergistically fusing features from RGB, depth, and infrared (IR) imaging modalities. This multimodal approach allows the model to leverage complementary information: RGB provides rich texture, depth offers invariance to lighting and aids pose estimation, and IR ensures robust pupil detection. Furthermore, we introduce a personalized adaptation module that dynamically fine-tunes the model to individual users with minimal calibration data. To ensure practical deployment, we employ advanced model compression techniques, enabling real-time inference on resource-constrained embedded systems. Extensive evaluations on public datasets (MPIIGaze, EYEDIAP, Gaze360) and our collected M-Gaze dataset demonstrate that GazeNet-HM achieves state-of-the-art performance, reducing the mean angular error by up to 27.1% compared to leading unimodal methods. After model compression, the system achieves a real-time inference speed of 32 FPS on an embedded Jetson Xavier NX platform. Ablation studies confirm the contribution of each modality and component, highlighting the effectiveness of our holistic design.
- Open Access
- 1 January 2026
- Published by Wiley in Advances in Human-Computer Interaction
- Vol. 2026 (1) , 3300666
TinyGaze v2: A Systems Benchmark for Practical Gaze Estimation on Edge Hardware
The deployment of robust, real‐time gaze estimation on commodity hardware is critical for pervasive gaze‐aware augmented and virtual reality (AR/VR). While the computer vision community has made significant progress in accuracy, a pervasive focus on leaderboard performance has created a gap between research and the system‐level constraints of real‐world edge devices. High‐end smartphones and AR/VR headsets increasingly incorporate dedicated ASIC‐based eye tracking hardware, yet such solutions are not universally accessible, as they require custom sensor arrays, fixed camera placements and proprietary SDKs that preclude cross‐platform deployment. This paper presents TinyGaze v2, an open‐source benchmarking suite for edge‐based gaze estimation. We introduce a novel set of hardware‐centric metrics (FPS‐at‐5W, Sustained FPS) and a standardised evaluation protocol. As a concrete instantiation of this benchmark, we provide an optimised gaze estimation model that achieves 58 FPS at 4.8° error on a Jetson Nano while consuming under 4 W. Through rigorous evaluation on a diverse hardware testbed (Jetson Nano, Raspberry Pi + TPU, Google Pixel 7), our multiobjective bootstrap analysis shows that TinyGaze v2 and FullFace are the only statistically nondominated models, occupying complementary Pareto regimes: FullFace for highest accuracy (4.10°, 8 FPS) and TinyGaze v2 for highest efficiency (4.81°, 58 FPS). Our work provides the community with the necessary tools and perspective—not a single model, but a reproducible benchmarking framework—to build feasible, real‐world gaze‐enabled systems.
- 1 January 2026
- Published by Wiley in Advances in Human-Computer Interaction
- Vol. 2026 (1) , 3300666
Open AccessTinyGaze v2: A Systems Benchmark for Practical Gaze Estimation on Edge Hardware
The deployment of robust, real‐time gaze estimation on commodity hardware is critical for pervasive gaze‐aware augmented and virtual reality (AR/VR). While the computer vision community has made significant progress in accuracy, a pervasive focus on leaderboard performance has created a gap between research and the system‐level constraints of real‐world edge devices. High‐end smartphones and AR/VR headsets increasingly incorporate dedicated ASIC‐based eye tracking hardware, yet such solutions are not universally accessible, as they require custom sensor arrays, fixed camera placements and proprietary SDKs that preclude cross‐platform deployment. This paper presents TinyGaze v2, an open‐source benchmarking suite for edge‐based gaze estimation. We introduce a novel set of hardware‐centric metrics (FPS‐at‐5W, Sustained FPS) and a standardised evaluation protocol. As a concrete instantiation of this benchmark, we provide an optimised gaze estimation model that achieves 58 FPS at 4.8° error on a Jetson Nano while consuming under 4 W. Through rigorous evaluation on a diverse hardware testbed (Jetson Nano, Raspberry Pi + TPU, Google Pixel 7), our multiobjective bootstrap analysis shows that TinyGaze v2 and FullFace are the only statistically nondominated models, occupying complementary Pareto regimes: FullFace for highest accuracy (4.10°, 8 FPS) and TinyGaze v2 for highest efficiency (4.81°, 58 FPS). Our work provides the community with the necessary tools and perspective—not a single model, but a reproducible benchmarking framework—to build feasible, real‐world gaze‐enabled systems.
- Open Access
- 28 July 2025
- Published by Springer Nature in Discover Applied Sciences
- Vol. 7 (8) , 1-28
Exploration of machine learning techniques for cloud removal and gap filling on sentinel-2 time series images for better exploitation in far North Cameroon
- 28 July 2025
- Published by Springer Nature in Discover Applied Sciences
- Vol. 7 (8) , 1-28
Open AccessExploration of machine learning techniques for cloud removal and gap filling on sentinel-2 time series images for better exploitation in far North Cameroon
- Open Access
- 21 April 2025
- Published by Springer Nature in Research Square
Exploration of Machine Learning techniques for Cloud Removal and Gap Filling on Sentinel-2 time series images for better Exploitation in Far North Cameroon
The direct exploitation and interpretation of optical satellite images, such as Sentinel-2 data, is significantly hampered by cloud cover. In this paper, we explore several machine learning algorithms in order to suggest a machine learning-based method for cloud removal and gap filling in Sentinel-2 satellite pictures for better utilization in the far north of Cameroon, concentrating on the city of Maroua. Our goal is to successfully fill these gaps produced by these cloud masks in the photos by using data from several photographs taken on various dates, assuming that cloud occurrence changes both geographically and temporally. The cloud-covered and cloud-free regions are analyzed using a variety of machine learning methods, such as Random Forest, VGG16 with Random Forest, VGG with dense layers, VGG16 with image data augmentation, SVM, and deep learning CNN models. We assess the correctness of each model and compare their performance through rigorous experimentation. Our findings show that the VGG16 model with the addition of picture data had the best accuracy, at about 95%.
- 21 April 2025
- Published by Springer Nature in Research Square
Open AccessExploration of Machine Learning techniques for Cloud Removal and Gap Filling on Sentinel-2 time series images for better Exploitation in Far North Cameroon
The direct exploitation and interpretation of optical satellite images, such as Sentinel-2 data, is significantly hampered by cloud cover. In this paper, we explore several machine learning algorithms in order to suggest a machine learning-based method for cloud removal and gap filling in Sentinel-2 satellite pictures for better utilization in the far north of Cameroon, concentrating on the city of Maroua. Our goal is to successfully fill these gaps produced by these cloud masks in the photos by using data from several photographs taken on various dates, assuming that cloud occurrence changes both geographically and temporally. The cloud-covered and cloud-free regions are analyzed using a variety of machine learning methods, such as Random Forest, VGG16 with Random Forest, VGG with dense layers, VGG16 with image data augmentation, SVM, and deep learning CNN models. We assess the correctness of each model and compare their performance through rigorous experimentation. Our findings show that the VGG16 model with the addition of picture data had the best accuracy, at about 95%.