Ensemble Encoder–Decoder Models for Predicting Land Transformation

Land development is a dynamic and complex processinfluenced by a system of interconnected driving variables. Predicting such a process is important in mitigating severe climate situations and improving the resiliency of communities. Current predictive models in land transformation have not paid a serious attention to capturing and exploiting the interchannel relationships. Moreover, these models often have problems with generalization, which results in poor performance during testing. In this study, we use a novel multichannel data cube, constructed from socioeconomic attributes, terrain characteristics, and landscape traits, to predict land transformation in a watershed in the US. In particular, we introduce methods for projecting impervious land transformations using these data cubes, using 2-D and 3-D convolutional neural networks (CNNs) and their ensembles. We apply fusion at decision, score, and feature levels to improve the generalization ability and robustness of the proposed predictive models. Performance is assessed using the Dice coefficient, receiver operating characteristic curves, data visualization, and running time. Our study shows that the use of 2-D and 3-D CNN ensembles improved the performance of the models in terms of model stability, precision and recall, and Dice coefficient.