Eric Umuhoza
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Eric Umuhoza has a total of 45 co-authors. Here are the top co-authors by number of shared publications:
- Marco Brambilla
Politecnico di Milano, Milan, Italy
8 shared publications - Andrea Mauri
Université Claude Bernard Lyon 1, Lyon, France
3 shared publications Carlo Bernaschina
Politecnico di Milano
2 shared publications- Dalia Aime Jean Bwiza
Carnegie Mellon University, Kigali, Rwanda
2 shared publications Jordi Cabot
ICREA, Spain / Open University of Catalonia, Spain
2 shared publicationsMoise Busogi
2 shared publications
Latest Publications
- Open Access
- 9 October 2026
- Published by Elsevier in SSRN Electronic Journal
Field-scale maize yield estimation using a combination of satellite-based machine learning and crop growth model
Accurate field-scale crop yield estimation remains challenging in smallholder agroecosystems, where fields are highly heterogeneous and ground reference data are scarce. Although high-resolution satellite-derived vegetation indices (VIs) can characterize crop canopy dynamics, they may not adequately represent the weather and soil-water conditions driving yield variability. Whether crop growth model (CGM)-derived variables improve upon VIs alone and how they compare with direct integration of environmental information into machine learning (ML) remain poorly evaluated, mainly in sub-Saharan Africa. This study evaluated three Random Forest (RF) configurations for field-level maize yield estimation in Chivi District, Zimbabwe: (i) VIs-only, (ii) VIs + WOFOST-derived crop-growth and water-stress variables, and (iii) VIs + weather + SMAP root-zone soil moisture (RZSM). The VIs-only model explained 28% of the observed yield variation (RMSE = 861 kg/ha, NRMSE = 37.4%), while adding WOFOST-derived variables increased R2 to 0.54 and decreased RMSE to 685 kg/ha (NRMSE = 29.8%). This configuration also outperformed open-loop WOFOST (R2 = 0.43, RMSE = 765 kg/ha). The VIs + weather + SMAP RZSM achieved the highest performance (R2 = 0.57, RMSE = 663 kg/ha, NRMSE = 28.8%), but only modestly surpassed VIs + WOFOST. Permutation feature importance showed a shift from GCVI as the dominant predictor in the VIs-only model to WOFOST-derived transpiration reduction and water-stress variables in the hybrid configuration, and to temperature and RZSM in the direct environmental model. These findings demonstrate that augmenting high-resolution satellite data with process-based crop responses or direct weather and soil moisture information better capture field-level maize yield variability than VIs alone. Importantly, VIs + WOFOST model achieved predictive performance comparable to VIs + weather + SMAP RZSM model while improving process-level interpretability, which can help identify environmental features that affect crop productivity and inform agricultural interventions at the field level.
- 9 October 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessField-scale maize yield estimation using a combination of satellite-based machine learning and crop growth model
Accurate field-scale crop yield estimation remains challenging in smallholder agroecosystems, where fields are highly heterogeneous and ground reference data are scarce. Although high-resolution satellite-derived vegetation indices (VIs) can characterize crop canopy dynamics, they may not adequately represent the weather and soil-water conditions driving yield variability. Whether crop growth model (CGM)-derived variables improve upon VIs alone and how they compare with direct integration of environmental information into machine learning (ML) remain poorly evaluated, mainly in sub-Saharan Africa. This study evaluated three Random Forest (RF) configurations for field-level maize yield estimation in Chivi District, Zimbabwe: (i) VIs-only, (ii) VIs + WOFOST-derived crop-growth and water-stress variables, and (iii) VIs + weather + SMAP root-zone soil moisture (RZSM). The VIs-only model explained 28% of the observed yield variation (RMSE = 861 kg/ha, NRMSE = 37.4%), while adding WOFOST-derived variables increased R2 to 0.54 and decreased RMSE to 685 kg/ha (NRMSE = 29.8%). This configuration also outperformed open-loop WOFOST (R2 = 0.43, RMSE = 765 kg/ha). The VIs + weather + SMAP RZSM achieved the highest performance (R2 = 0.57, RMSE = 663 kg/ha, NRMSE = 28.8%), but only modestly surpassed VIs + WOFOST. Permutation feature importance showed a shift from GCVI as the dominant predictor in the VIs-only model to WOFOST-derived transpiration reduction and water-stress variables in the hybrid configuration, and to temperature and RZSM in the direct environmental model. These findings demonstrate that augmenting high-resolution satellite data with process-based crop responses or direct weather and soil moisture information better capture field-level maize yield variability than VIs alone. Importantly, VIs + WOFOST model achieved predictive performance comparable to VIs + weather + SMAP RZSM model while improving process-level interpretability, which can help identify environmental features that affect crop productivity and inform agricultural interventions at the field level.
- Open Access
- 2 September 2026
- Published by Association for Computing Machinery (ACM)
- p. 524-532
Towards fairer public transit: Real-time tensor fusion of video and audio for fare evasion and fraud detection
- 2 September 2026
- Published by Association for Computing Machinery (ACM)
- p. 524-532
Open AccessTowards fairer public transit: Real-time tensor fusion of video and audio for fare evasion and fraud detection
- Open Access
- 20 April 2026
- Published by Elsevier in SSRN Electronic Journal
GeoAI-Based Crop Yield Estimation in Africa: A Systematic and Bibliometric Literature Review With Comparisons to Major Agricultural Producers
Early and accurate crop monitoring and yield estimation are vital for improving food security and sustainable agriculture, yet conventional monitoring approaches based on censuses and crop cutting experiments remain costly and difficult to scale. GeoAI, which integrates EO data with ML algorithms, offers a cost-effective and scalable complement to conventional methods. However, the extent to which GeoAI approaches have advanced toward operational yield prediction in Africa remains unclear. This study presents the first comprehensive, systematic, and bibliometric review of GeoAI-based crop yield prediction in Africa, with a comparative perspective on major agricultural producers (MAP). Following PRISMA 2020 guidelines, 43 Africa- and 64 MAP-focused studies were analyzed to examine (i) the geospatial data sources used for yield estimation, (ii) the ML models and their reported performance, (iii) the geographic, institutional, authorship, and funding structures shaping the research landscape, and (iv) methodological trends, gaps, and future research directions. Results show that Sentinel-2 has become the dominant satellite data source for ML-based yield estimation in Africa, largely replacing coarse-resolution MODIS due to its improved spatial and spectral (red-edge bands) resolutions. Across studies, NDVI, precipitation, and temperature remain the most widely used predictors, while enhanced spectral indices (EVI, GCVI, MTCI, NDRE) and ancillary variables such as soil properties, LAI, evapotranspiration, and crop management information improve sensitivity and predictive performance. Random forest consistently performs well under data-scarce conditions, generally surpassing linear models. While deep learning remains underutilized and do not demonstrate systematic performance gains because of limited and noisy training datasets. Hybrid approaches integrating crop growth models with GeoAI show promise in mitigating sparse ground observations. Bibliometric analysis further reveals a highly donor-dependent and externally anchored research landscape, with geographically uneven study distribution and limited leadership by local institutions, in contrast to the diversified data infrastructures and nationally funded research ecosystems observed in MAP. We identify also critical gaps and actionable directions, including multi-year dataset harmonization, multi-sensor fusion, hybrid GeoAI-CGM modeling, domain adaptation strategies, and stable domestic research ecosystems to provide a sustainable roadmap for scalable and operational crop-yield prediction in data-scarce landscapes.
- 20 April 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessGeoAI-Based Crop Yield Estimation in Africa: A Systematic and Bibliometric Literature Review With Comparisons to Major Agricultural Producers
Early and accurate crop monitoring and yield estimation are vital for improving food security and sustainable agriculture, yet conventional monitoring approaches based on censuses and crop cutting experiments remain costly and difficult to scale. GeoAI, which integrates EO data with ML algorithms, offers a cost-effective and scalable complement to conventional methods. However, the extent to which GeoAI approaches have advanced toward operational yield prediction in Africa remains unclear. This study presents the first comprehensive, systematic, and bibliometric review of GeoAI-based crop yield prediction in Africa, with a comparative perspective on major agricultural producers (MAP). Following PRISMA 2020 guidelines, 43 Africa- and 64 MAP-focused studies were analyzed to examine (i) the geospatial data sources used for yield estimation, (ii) the ML models and their reported performance, (iii) the geographic, institutional, authorship, and funding structures shaping the research landscape, and (iv) methodological trends, gaps, and future research directions. Results show that Sentinel-2 has become the dominant satellite data source for ML-based yield estimation in Africa, largely replacing coarse-resolution MODIS due to its improved spatial and spectral (red-edge bands) resolutions. Across studies, NDVI, precipitation, and temperature remain the most widely used predictors, while enhanced spectral indices (EVI, GCVI, MTCI, NDRE) and ancillary variables such as soil properties, LAI, evapotranspiration, and crop management information improve sensitivity and predictive performance. Random forest consistently performs well under data-scarce conditions, generally surpassing linear models. While deep learning remains underutilized and do not demonstrate systematic performance gains because of limited and noisy training datasets. Hybrid approaches integrating crop growth models with GeoAI show promise in mitigating sparse ground observations. Bibliometric analysis further reveals a highly donor-dependent and externally anchored research landscape, with geographically uneven study distribution and limited leadership by local institutions, in contrast to the diversified data infrastructures and nationally funded research ecosystems observed in MAP. We identify also critical gaps and actionable directions, including multi-year dataset harmonization, multi-sensor fusion, hybrid GeoAI-CGM modeling, domain adaptation strategies, and stable domestic research ecosystems to provide a sustainable roadmap for scalable and operational crop-yield prediction in data-scarce landscapes.
- Open Access
- 13 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-7
Co-Data: Cultivating Effective Human-LLM Collaboration for Collaborative Data Processing
- 13 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-7
Open AccessCo-Data: Cultivating Effective Human-LLM Collaboration for Collaborative Data Processing
- Open Access
- 12 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 55-58
Enhancing Road Safety in Resource-Constrained Cities: Leveraging Road User Participation in Road Traffic Crash Reporting
Road traffic crashes remain a major public health challenge in low- and middle-income countries, where underreporting and delayed documentation limit timely response and evidence-based policymaking. We present a citizen-powered crash reporting system for resource-constrained settings, enabling real-time reporting via both a smartphone application and a USSD interface for feature phones. The system captures structured incident data and optional evidence, timestamps and geotags submissions, and supports role-based access for stakeholders such as police, emergency responders, and insurers. We describe stakeholder-informed requirements from Rwanda, the system design and prototype implementation, and deployment-oriented recommendations for integrating institutional registries and improving verification and analytics while preserving accessibility and privacy.
- 12 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 55-58
Open AccessEnhancing Road Safety in Resource-Constrained Cities: Leveraging Road User Participation in Road Traffic Crash Reporting
Road traffic crashes remain a major public health challenge in low- and middle-income countries, where underreporting and delayed documentation limit timely response and evidence-based policymaking. We present a citizen-powered crash reporting system for resource-constrained settings, enabling real-time reporting via both a smartphone application and a USSD interface for feature phones. The system captures structured incident data and optional evidence, timestamps and geotags submissions, and supports role-based access for stakeholders such as police, emergency responders, and insurers. We describe stakeholder-informed requirements from Rwanda, the system design and prototype implementation, and deployment-oriented recommendations for integrating institutional registries and improving verification and analytics while preserving accessibility and privacy.
- Open Access
- 28 March 2026
- Published by Elsevier in SSRN Electronic Journal
Towards fairer public transit: Real-time tensor fusionof video and audio for fare evasion and fraud detection
This study developed a real-time multimodal system to detect fare evasionand fraud in public transportation by integrating video and audio streamsfrom CCTV. The system addressed the limitations of manual surveillancecurrently used in Rwanda, where fare enforcement is highly dependent onoperators monitoring multiple camera feeds simultaneously. To capture com-plex behavioral and acoustic indicators of fraud, we employed the VisionTransformer for Video (ViViT) for spatio-temporal video feature extractionand the Audio Spectrogram Transformer (AST) for transaction-related au-dio analysis. A Tensor Fusion Network (TFN) was adapted and extendedto explicitly model unimodal and bimodal interactions between visual andaudio modalities through a 2-fold Cartesian product. The model was trainedand evaluated on a custom dataset of 820 synchronized video–audio sam-ples, representing real-world boarding scenarios labeled as either legitimateor fraudulent. Using a stratified 5-fold cross-validation procedure, the sys-tem achieved an accuracy of 89.5%, a precision of 87.2%, a recall of 84.0%,and an F1-score of 85.6%, outperforming both unimodal baselines and earlyand latefusion methods. Ablation studies further demonstrated that explicittensor fusion improved the F1-score by 7.0% over traditional concatenationapproaches. These results highlight the potential of multimodal tensor-basedlearning to significantly improve automation, reduce revenue leakage, and im-prove operational compliance in public transportation systems. The findingsprovide a foundation for scalable deployment and motivate future work onintegrating additional data sources, increasing the size of the dataset, andimproving the robustness under real-world conditions.
- 28 March 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessTowards fairer public transit: Real-time tensor fusionof video and audio for fare evasion and fraud detection
This study developed a real-time multimodal system to detect fare evasionand fraud in public transportation by integrating video and audio streamsfrom CCTV. The system addressed the limitations of manual surveillancecurrently used in Rwanda, where fare enforcement is highly dependent onoperators monitoring multiple camera feeds simultaneously. To capture com-plex behavioral and acoustic indicators of fraud, we employed the VisionTransformer for Video (ViViT) for spatio-temporal video feature extractionand the Audio Spectrogram Transformer (AST) for transaction-related au-dio analysis. A Tensor Fusion Network (TFN) was adapted and extendedto explicitly model unimodal and bimodal interactions between visual andaudio modalities through a 2-fold Cartesian product. The model was trainedand evaluated on a custom dataset of 820 synchronized video–audio sam-ples, representing real-world boarding scenarios labeled as either legitimateor fraudulent. Using a stratified 5-fold cross-validation procedure, the sys-tem achieved an accuracy of 89.5%, a precision of 87.2%, a recall of 84.0%,and an F1-score of 85.6%, outperforming both unimodal baselines and earlyand latefusion methods. Ablation studies further demonstrated that explicittensor fusion improved the F1-score by 7.0% over traditional concatenationapproaches. These results highlight the potential of multimodal tensor-basedlearning to significantly improve automation, reduce revenue leakage, and im-prove operational compliance in public transportation systems. The findingsprovide a foundation for scalable deployment and motivate future work onintegrating additional data sources, increasing the size of the dataset, andimproving the robustness under real-world conditions.
- 2 January 2026
- Published by Springer Nature
- p. 439-448
Application of Machine Learning on Satellite Imagery for Crop-Type Classification in Sub-Saharan Africa
- 2 January 2026
- Published by Springer Nature
- p. 439-448
Application of Machine Learning on Satellite Imagery for Crop-Type Classification in Sub-Saharan Africa
- Open Access
- 2 September 2021
- Published by Wiley in Journal of Software: Evolution and Process
- Vol. 35 (5) , e2377
Design thinking and creativity of colocated versus globally distributed software developers
Designing software is an activity in which software developers think and make design decisions that shape the structure and behavior of software products. Designing software is one of the least understood software engineering activities. In a collaborative design setting, various types of distances can lead to challenges and effects that potentially affect how software is designed. To contribute to a better understanding of collaborative software design, we investigate how communication gaps caused by social and geographic distances affect its design thinking and the creativity of its discussions. To this end, we conducted a multiple‐case study exploring the design thinking and creativity of colocated and distributed software developers in a collaborative design setting. Compared with colocated developers, distributed developers spend less time on exploring the problem space, which could be related to different sociotechnical challenges, such as lack of awareness and common understanding. Distributed development does not seem to affect the creativity of their activities. Developers engaging in collaborative design need to be aware that problem space exploration is reduced in a distributed setting. Unless distributed teams take compensatory measures, this could adversely affect the development. Regarding the effect distance has on creativity, our results are inconclusive and further studies are needed.
- 2 September 2021
- Published by Wiley in Journal of Software: Evolution and Process
- Vol. 35 (5) , e2377
Open AccessDesign thinking and creativity of colocated versus globally distributed software developers
Designing software is an activity in which software developers think and make design decisions that shape the structure and behavior of software products. Designing software is one of the least understood software engineering activities. In a collaborative design setting, various types of distances can lead to challenges and effects that potentially affect how software is designed. To contribute to a better understanding of collaborative software design, we investigate how communication gaps caused by social and geographic distances affect its design thinking and the creativity of its discussions. To this end, we conducted a multiple‐case study exploring the design thinking and creativity of colocated and distributed software developers in a collaborative design setting. Compared with colocated developers, distributed developers spend less time on exploring the problem space, which could be related to different sociotechnical challenges, such as lack of awareness and common understanding. Distributed development does not seem to affect the creativity of their activities. Developers engaging in collaborative design need to be aware that problem space exploration is reduced in a distributed setting. Unless distributed teams take compensatory measures, this could adversely affect the development. Regarding the effect distance has on creativity, our results are inconclusive and further studies are needed.
- 1 September 2020
- Published by Institute of Electrical and Electronics Engineers (IEEE) in SAIEE Africa Research Journal
- Vol. 111 (3) , 95-101
Using Unsupervised Machine Learning Techniques for Behavioral-based Credit Card Users Segmentation in Africa
- 1 September 2020
- Published by Institute of Electrical and Electronics Engineers (IEEE) in SAIEE Africa Research Journal
- Vol. 111 (3) , 95-101
Using Unsupervised Machine Learning Techniques for Behavioral-based Credit Card Users Segmentation in Africa
- Open Access
- 26 September 2017
- Published by Sociedade Brasileira de Computação in Journal of Internet Services and Applications
- Vol. 8 (1) , 246-253
Model-driven development of user interfaces for IoT systems via domain-specific components and patterns
Internet of Things technologies and applications are evolving and continuously gaining traction in all fields and environments, including homes, cities, services, industry and commercial enterprises. However, still many problems need to be addressed. For instance, the IoT vision is mainly focused on the technological and infrastructure aspect, and on the management and analysis of the huge amount of generated data, while so far the development of front-end and user interfaces for IoT has not played a relevant role in research. On the contrary, user interfaces can play a key role in the acceptance of IoT solutions by final adopters. In this paper we discuss the requirements and usage scenarios covering the front end aspects of IoT systems and we present a model-driven approach to the design of such interfaces by: defining specific components and design patterns using a visual modeling language for IoT applications; describing an implementation of the solution that comprises also automatic code generation from models; and by showing the solution at work.
- 26 September 2017
- Published by Sociedade Brasileira de Computação in Journal of Internet Services and Applications
- Vol. 8 (1) , 246-253
Open AccessModel-driven development of user interfaces for IoT systems via domain-specific components and patterns
Internet of Things technologies and applications are evolving and continuously gaining traction in all fields and environments, including homes, cities, services, industry and commercial enterprises. However, still many problems need to be addressed. For instance, the IoT vision is mainly focused on the technological and infrastructure aspect, and on the management and analysis of the huge amount of generated data, while so far the development of front-end and user interfaces for IoT has not played a relevant role in research. On the contrary, user interfaces can play a key role in the acceptance of IoT solutions by final adopters. In this paper we discuss the requirements and usage scenarios covering the front end aspects of IoT systems and we present a model-driven approach to the design of such interfaces by: defining specific components and design patterns using a visual modeling language for IoT applications; describing an implementation of the solution that comprises also automatic code generation from models; and by showing the solution at work.