Advancing unpaved road assessment in Africa: Leveraging multimodal machine learning and large language-and-vision assistants across satellite imagery resolutions

Over 53 % of African road network is unpaved, yet systematic monitoring remains limited. This study introduces a cost-effective machine learning (ML) solution to help local authorities monitor and plan road maintenance. Building on earlier work using high-resolution satellite imagery in Tanzania, the analysis extends to Madagascar, incorporating medium- and low-resolution imagery to reduce costs. Two distinct methodologies were evaluated: traditional ML and multimodal ML. The multimodal ML model achieves 93.2 % accuracy with high-resolution imagery and maintains satisfactory performance with medium (84.0 %) and low-resolution (85.3 %) imagery, aided by transfer learning. The framework demonstrates robust cross-resolution performance across Tanzania and Madagascar contexts. Additionally, a pilot study explored a fine-tuned Large Language-and-Vision Assistant (LLaVA) model, which demonstrated potential for natural language-based condition reporting and maintenance recommendations, offering an interpretable alternative to quantitative classification outputs. Whilst LLaVA currently exhibits lower classification accuracy than the multimodal ML model, multi-turn conversational approaches show promise for enhancing performance whilst maintaining natural language interpretability. This study contributes to Sustainable Development Goal 9.1 by delivering a scalable, affordable strategy to support resilient infrastructure and economic development in low-income regions.