GeoAI Tutorial Extracts Building Footprints From Aerial Imagery

A comprehensive tutorial has been released detailing a GeoAI workflow designed to extract building footprints from high-resolution National Agriculture Imagery Program (NAIP) aerial imagery. This workflow integrates several advanced deep learning models, including U-Net, Grounding DINO, Segment Anything Model (SAM), and Mask R-CNN, to achieve accurate and robust results. The tutorial begins by guiding users through the configuration of a geospatial deep learning environment, a crucial first step for any advanced geospatial analysis. This involves downloading the necessary raster imagery and corresponding vector labels, followed by an inspection of their spatial properties to ensure data integrity and compatibility. Subsequently, the process moves to generating georeferenced image chips and segmentation masks, which are essential for training and evaluating the deep learning models.

The core of the tutorial focuses on training a U-Net model, a popular convolutional neural network architecture for image segmentation tasks. Specifically, it details the use of a ResNet-34 encoder within the U-Net framework. Users are instructed on how to evaluate the model's learning behavior, a critical step in understanding its performance and identifying potential issues. Following training, the tutorial demonstrates the application of a sliding-window inference technique on an unseen scene, allowing the model to process larger areas of imagery than it was directly trained on. This method is vital for practical applications where imagery datasets are extensive.

Beyond basic semantic segmentation, the tutorial extends to more sophisticated post-processing and alternative segmentation approaches. It explains how to convert the predicted segmentation masks into cleaned and regularized building polygons, a necessary step for creating usable vector data. Performance evaluation is addressed through the calculation of Intersection over Union (IoU) and F1 metrics, standard measures for assessing the accuracy of segmentation models. The workflow also explores zero-shot segmentation capabilities using Grounding DINO and SAM, models that can identify objects based on textual descriptions without prior specific training on those object classes. The results from these zero-shot methods are then compared with those obtained from a pretrained Mask R-CNN instance segmentation model, providing a comparative analysis of different approaches.

Furthermore, the tutorial emphasizes the practical applicability of the developed pipeline by demonstrating its extension to real-world areas. This is achieved by utilizing NAIP imagery sourced from Microsoft Planetary Computer, a platform offering vast amounts of geospatial data, and building labels obtained from Overture Maps, an open data initiative. This integration highlights the workflow's adaptability to publicly available datasets, making it a valuable resource for researchers, developers, and practitioners in the field of geospatial artificial intelligence and remote sensing. The entire process is presented in a manner conducive to replication and further development.