Deep Learning–Based Approach for Short- and Long-Term IRI Forecasting Using Long-Term Pavement Performance Data

<jats:title>Abstract</jats:title> <jats:p> This study presents a deep learning–based approach for forecasting the international roughness index (IRI) using a long short-term memory (LSTM) network enhanced with an attention mechanism. Leveraging data from the long-term pavement performance program, the proposed LSTM–attention model is applied to both short-term (three-step) and long-term (six-step) prediction horizons and is systematically evaluated under different historical input lengths (lags). The results show that the model can generate accurate multistep IRI forecasts, with a 7-year lag providing the best trade-off between predictive accuracy and data availability for long-term predictions. Stratified analyses further reveal that predictive performance is higher for pavements in good-to-fair ( <jats:inline-formula> <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" alttext="0 en dash 1.5 normal m divided by k m" display="inline" overflow="scroll"> <mml:mn>0</mml:mn> <mml:mi>–</mml:mi> <mml:mn>1.5</mml:mn> <mml:mtext>  </mml:mtext> <mml:mi mathvariant="normal">m</mml:mi> <mml:mo stretchy="false">/</mml:mo> <mml:mi>km</mml:mi> </mml:math> </jats:inline-formula> ) condition than for those in more advanced stages of deterioration, indicating increasing variability in IRI progression as pavements degrade. The study also demonstrates that traditional evaluation metrics such as <jats:inline-formula> <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" alttext="upper R squared" display="inline" overflow="scroll"> <mml:msup> <mml:mi>R</mml:mi> <mml:mn>2</mml:mn> </mml:msup> </mml:math> </jats:inline-formula> can be misleading for deep nonlinear models and highlights Chatterjee’s correlation coefficient ( <jats:inline-formula> <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" alttext="xi Subscript n" display="inline" overflow="scroll"> <mml:msub> <mml:mrow> <mml:mi>ξ</mml:mi> </mml:mrow> <mml:mrow> <mml:mi>n</mml:mi> </mml:mrow> </mml:msub> </mml:math> </jats:inline-formula> ) as a more reliable measure of predictive association in this context. </jats:p>