Quantifying Nature-Based Flood Risk Reduction Through LSTM Modeling: Evidence from Pakistan's Swat River Basin

Escalating flood frequency and intensity, driven by anthropogenic land use modifications and climate variability, pose critical challenges to watershed management worldwide. This study examines the hydrological impacts of the Billion Tree Tsunami (BTT) reforestation initiative, implemented in 2014 across Pakistan's Khyber Pakhtunkhwa province, with specific focus on flood attenuation dynamics in the Swat River catchment. We integrate land use and land cover (LULC) change analysis spanning three decades (1990–2024) with Long Short-Term Memory (LSTM) neural networks to assess historical discharge patterns and project future hydrological conditions through 2050. LULC analysis reveals substantial landscape transformation, including significant forest expansion, marked reduction in barren land, and agricultural land modifications. Statistical evaluation demonstrates notable flood mitigation effects post-intervention, with 15% reduction in peak flows and decreased discharge intensification rates. The LSTM models exhibit strong predictive performance (R² = 0.87), forecasting a 20–25% reduction in peak discharge events by 2050 under continued reforestation scenarios. These findings underscore the efficacy of large-scale reforestation as a nature-based solution for flood risk reduction and demonstrate the value of integrating machine learning approaches with conventional hydrological modeling for enhanced watershed management strategies.