Pragmatic detection and early warning model for reduction of landslide effects in communities at risk : a case of selected sub-counties in Bududa district, Uganda

Abstract

This study addressed the devastating and persistent threat of landslides in Bududa District, Uganda, by designing and evaluating a pragmatic Landslide Early Warning Model (LEWaM), recognizing that existing systems are widely perceived as ineffective. The research adopted design science as a paradigm and was a case study design where qualitative data informed the quantitative data with a sample of 199 participants. Data collection methods involved survey methed, focus group discussion, and interviews, documentation review, and observation. The research first diagnosed critical system shortcomings and local knowledge gaps. Key findings confirmed that over half of the participants (51.3%) rated the current EWS at only 25% effectiveness, citing technical failures such as data inaccuracy, power outages, and equipment vandalism, alongside reliance on slow manual observation. In contrast, the community relies heavily on indigenous knowledge, recognizing heavy rainfall (96.2%) and precursory signs like soil cracks (89.7%), and uses word-of-mouth (71.2%) for warnings. This confirmed a critical disconnect between formal detection tools and effective, context-specific hazard communication, underscoring the necessity for a customized solution. The resulting Landslide Early Warning Model (LEWaM) was designed and implemented using Unified Modeling Language (UML) and was subsequently evaluated using a Random Forest machine learning algorithm on a tailored, dynamic dataset. The model evaluation demonstrated robust performance, achieving an overall accuracy of 89% with a macro-averaged F1-score of 0.86. Critically, LEWaM achieved near-perfect precision and recall in detecting "Danger" conditions, proving its reliability for identifying imminent threats while maintaining a low false alarm rate (3%) for "Normal" conditions. The primary area for refinement was identified in the "Warning" class, where a recall of 54% suggested a need to boost early warning sensitivity. Based on these results, recommendations include institutionalizing LEWaM within Bududa’s disaster plans, initiating a phased rollout strategy in pilot sub-counties prone to landslides, and empowering community ambassadors for warning relay. Future work should prioritize training the ML component on localized, real-world Bududa data to refine the sensitivity and precision of the crucial early "Warning" class.