Integrating community perspectives and a drone-based machine learning framework for risk profiling of potential aedes breeding sites

Dengue remains a major environmental health challenge in rapidly urbanising environments. This study evaluated the potential of drone technology and Machine Learning (ML) to support community-level dengue prevention in Malaysia. A cross-sectional mixed-methods design was employed, integrating three distinct analytical phases. First, a socio-behavioural assessment combined household surveys (n = 866) in urban and rural communities with expert interviews to evaluate public perception and readiness. Second, high-resolution red, green, and blue (RGB) drone imagery was acquired to conduct geospatial and microclimate modelling of the built environment. Finally, these environmental features were integrated into a Machine Learning pipeline to classify potential Aedes breeding habitats. Together, these methods form an integrated framework linking community readiness with data-driven spatial risk profiling. Under the first objective (socio-behavioural), urban residents presented higher acceptance of drone-based surveillance (65.4%) than rural residents (38.2%). Privacy concerns were significant (42.7%), particularly in rural areas (p < 0.001). Younger adults (18-30 years) were most willing to engage in drone-assisted control and download a surveillance application, whereas older adults favoured traditional training. Across states, a persistent knowledge-practice gap was revealed, with notable urban-rural heterogeneity. Expert interviews highlighted regulatory constraints, privacy, and community engagement as key barriers, with sociopolitical trust acting as a critical mediator. Under the second objective (geospatial), drone-based microclimate modelling revealed that vegetation-shade interactions and housing typology strongly modulate Aedes habitat risk. Low-density terrace housing exhibited substantially higher ecological risk scores than high-rise settings, with high-risk zones covering up to 46.5% of terrace areas. Model validation demonstrated high predictive performance (R² = 0.91), with the top 20% of predicted high-risk pixels successfully accounting for 65% of observed breeding-prone areas. Under the third objective (ML classification), single-site classification of RGB imagery achieved overall accuracies of 70-90% with moderate to strong Kappa coefficients (~0.5-0.9). However, cross-site transferability declined to 40-55% accuracy due to spectral and structural heterogeneity between housing estates. This study proposes an integrated framework that links community readiness, environmental risk profiling, and ML-based habitat classification to support targeted dengue prevention. Ultimately, the findings suggest that when embedded in meaningful community engagement and appropriate governance, drone and ML technologies can significantly enhance the precision and scalability of environmentally oriented dengue surveillance.