Landslide risk is traditionally predicted by process-based models with detailed assessments or point-scale, attribute-based machine learning (ML) models with first- or second-order features, e.g., slope, as inputs. One could hypothesize that terrain patterns might contain useful information that could be extracted, via computer vision ML models, to elevate prediction performance beyond that achievable with low-order features. We put this hypothesis to the test in the state of Oregon, where a large landslide dataset is available. The image-processing convolutional neural networks (CNN2D) using 2D terrain data obtained either higher Precision or higher Recall than attribute-based random forest (RF1D) models, but could not improve both simultaneously. While CNN2D can be set up to identify more real events, it would then introduce more false positives, highlighting the challenge of generalizing landslide-prone terrain patterns and the potential omission of critical factors. However, ensembling CNN2D and RF1D produced overall better Precision and Recall, and this cross-model-type ensemble was better than other ways to ensemble, leveraging information content of fine-scale topography while suppressing its noise. These models further showed robust results in cross-regional validation. Our perturbation tests showed that 10m resolution (the smallest possible) produced the best model in a range of resolutions. Rainfall, land cover, soil moisture, and elevation were the most important predictors. Based on the results of the analysis, we generated landslide susceptibility maps, providing insights into spatial patterns of landslide risk.