The accuracy assessment of LULC classification is presented in Table 2 by using the confusion matrix to compare the ground truth and DNN prediction. The overall prediction accuracy is 0.9073, where the impervious features, pervious land and trees achieved the prediction accuracy of 0.9634, 0.7422 and 0.8611, respectively. A robust LULC prediction model developed by Rwanga & Ndambuki (2017) was also assessed by using the confusion matrix and got a Kappa coefficient of 0.722. In this study, the Kappa coefficient (0.8217) qualifies as considerable, indicating that the proposed DL approach has consistent prediction performance among the defined LULC classes of the selected study catchments. The acceptable level of prediction accuracy varies depending on the available conditions, and most DL models consider 75% accuracy as the lowest boundary of excellent prediction (Li et al. 2019; He et al. 2020; Kiran 2020). In this study, the 0.9634 prediction accuracy of impervious features is sufficient to estimate parameters for simulating the rainfall–runoff process in urban catchments. Therefore, the outputs of impervious feature extraction are acceptably conveyed to the sampling process as the inputs of the probability-fitting study.