Pixel Perfect: Using Vision Transformers to Improve Road Quality Predictions from Medium Resolution and Heterogeneous Satellite Imagery

Abstract

Critical infrastructure, such as roads and electricity, are core systems that enable economic development. However, these crucial systems are frequently under-monitored in developing regions, resulting in lost opportunities for growth. Recent advances in remote sensing and machine learning have enabled monitoring and measurement of infrastructure faster and more frequently than traditional methods. However, ground data are often unavailable, resulting in a disconnect between labels and remotely sensed data. Furthermore, data from industrialized regions can only sometimes be transferred to regions with sparse data due to differences in the concept of quality between regions. Additionally, inconsistency in data and the complexity of ML models can introduce bias due to learned characteristics across diverse regions, leading to inaccurate predictions and recommendations for action. In this study, we train and compare traditional neural networks and vision transformers to predict road quality from medium-resolution satellite imagery and apply them to realistic data conditions: heterogeneous temporal-spatial resolutions. The best models (vision transformers) achieve AUROC scores of 0.934 and 0.685 for binary and five-class classification tasks, respectively, exhibiting results appealing for inference in otherwise unmeasured areas. Furthermore, these experiments and results showed that proper training techniques could produce accurate models from limited, heterogeneous, and low-resolution data.