Beyond exposure: Understanding household vulnerability configurations in semi-arid agroforestry systems through interpretable machine learning

Household vulnerability is commonly explained through exposure to environmental stress, yet households facing similar climatic conditions often experience markedly different outcomes. This discrepancy raises a persistent question in vulnerability research: to what extent do environmental pressures themselves shape vulnerability, and to what extent does it reflect differences in livelihood resources? Addressing this question, the present study examines household vulnerability in semi-arid agroforestry systems in Southwest Sumba, Indonesia, using data from 122 purposively selected households through an integrated analytical framework that combines multidimensional vulnerability measurement and K-means clustering with systematic cluster validation and external profile characterization . Alternative solutions from K = 2 to K = 6 were evaluated using complementary measures of compactness, separation, and stability, with a four-cluster solution providing the most defensible configuration. The resulting profiles reveal substantial heterogeneity among households operating within the same semi-arid agroecological context: one profile combined high exposure with physical constraints, another combined high exposure with limited economic resources, a third combined lower exposure with higher sensitivity and weaker ecological resources, while the fourth combined high exposure with relatively strong livelihood resources and adaptation. External characterization showed significant differentiation in adaptation, agroforestry model, and village, while agroforestry experience, social capital, productivity, and gender showed weaker or non-significant differentiation. The findings indicate that households operating within similar agroecological conditions may occupy substantially different vulnerability configurations because environmental pressures intersect with unequal livelihood resources . The study contributes a validated, theory-informed profiling approach that moves beyond additive and exposure-centred assessments by revealing multidimensional vulnerability configurations without using the same variables to construct and subsequently predict profile membership. These findings support context-sensitive vulnerability assessment and differentiated adaptation planning, while avoiding causal claims about the effects of particular livelihood interventions.