CLIMATE CHANGE VULNERABILITY AND ADAPTIVE CAPABILITY OF SMALL-SCALE FARMERS IN PUNJAB, PAKISTAN: A MACHINE LEARNING APPROACH

Abstract

Smallholder farmers in Pakistan’s rain-fed north-central Punjab face intensifying climate stress, yet the institutional conditions that shape their adaptive capacity remain poorly understood. This study assesses household-level climate vulnerability and projects its trajectory to 2035 under three IPCC AR6-aligned scenarios. A cross-sectional survey of 200 smallholder farmers across Attack, Chakwal, Islamabad, and Jhelum districts was conducted. Three composite indices were constructed: a climate vulnerability index (CVI) grounded in the IPCC exposure–sensitivity framework, an adaptive strategy index (ASI), and an institutional and policy support index (IPS). A tuned random forest model (RFM) (1,000 trees; 5-fold cross-validation) was used to identify drivers of vulnerability, and scenario-based projections were generated to 2035. Results show that mean CVI is 0.645 on a image scale, placing the average household in the moderate-to-high vulnerability range. Random Forest analysis identified institutional support (IPS) as the dominant predictor of vulnerability (importance = 95), with education and district effects secondary. Critically, the data revealed an institutional vacuum: 96% of farmers reported no extension service access, no climate-resilience training, and no cooperative membership, despite 100% reporting observed weather changes and willingness to adopt new practices. Scenario projections indicate that by 2035, even under SSP2-4.5 with a 30% increase in adaptation and 40% increase in institutional support, 96% of farmers would fall into the high-vulnerability category statistically indistinguishable from the business-as-usual outcome (93.5%). Under SSP5-8.5, 99% of farmers become highly vulnerable. These findings suggest that incremental adaptation within the current institutional system cannot offset projected climate stress. Transformative policy response centered on building extension capacity, cooperative infrastructure, and mobile-based advisory at scale is required. The study contributes a replicable framework combining composite indexing, machine learning, and scenario projection for forward-looking vulnerability assessment.