It is well recognised that the severity of pipeline external corrosion is highly related to the corrosivity of the surrounding soil environment. However, in practice, the explicit effects of the soil corrosivity variables on the spatial distribution of external corrosion defects are largely unknown. This paper presents a novel modelling and predicting approach for pipeline external corrosion defect count data in terms of spatial patterns using on a multivariate Poisson-lognormal (MVPLN) model. The MVPLN model can account for the over-dispersion and unobserved heterogeneity of the defect count data, as well as consider the stochastic correlation between corrosion defects with different spatial patterns. The developed model is applied to a pipeline inspection dataset consisting in-line inspection (ILI) data and corresponding soil corrosivity measurements. Its performance is validated by using cross-validation. A comparison study shows that the MVPLN model provides superior modelling results for the spatial distribution of external corrosion defects over the commonly used univariate count data models. In addition, the obtained model coefficients of the soil corrosivity measurements are discussed, and their estimated impacts on the spatial patterns of corrosion defects are verified qualitatively. The potential application to assess the corrosion severity of non-piggable pipeline segments is further demonstrated.