Assessing Potential at Scale

Assessing whether a property can host solar panels requires evaluating roof geometry, tree shade, existing obstructions (such as HVAC, plumbing etc.), utility connections, and setback constraints. For solar installers, gathering this data manually across hundreds of sites per campaign is costly and error-prone. At StationA, we faced the challenge of automating this assessment at scale. Can we use remotely sensed imagery and computer vision to automatically detect and model solar and energy storage potential on hundreds of thousands of properties across the US on demand and en masse.

GeoAI for Rooftop and Carport Assessment

We built a multi-task ML pipeline using open-source pre-trained models fine-tuned on annotated satellite and aerial imagery. The objective is to detect what's on a roof and what's not accessible. The entities to be extracted include:
- Active solar panels (existing installations to avoid overlap)
- Keepouts (HVAC units, vents, skylights, antennas or any areas unsuitable for mounting)
- Tree shade (obstructions that reduce production)
- Roof and carport segmentation (constructing spatially contiguous rooftops/carports from surrounding clutter)

We trained models using Roboflow, an annotation-first platform that let us efficiently label tens of thousands of satellite chips, organize datasets, and manage multiple model versions. Roboflow's data augmentation capabilities also helped us address geographic diversity. A rooftop in Seattle looks different from one in Arizona, and augmentation reduced the labeling burden.

Both the rooftop and carport detection pipelines follow the same workflow: pre-trained instance segmentation models fine-tuned on labeled imagery and validated against ground-truth surveys from completed installations.

Training, Annotation, and Data Sourcing:

A non-trivial challenge in GeoAI projects is the choice of training and inference data. A rule of thumb is: if a feature is not visible to the naked eye, then a computer vision model will also not be able to extract it. This leads us to an optimal spatial resolution requirement window of centimeter-level imagery sources such as NAIP. Thus the annotation workflow consists of the following steps:
- Collect raw satellite/aerial imagery (USGS orthophoto, Sentinel-2, licensed providers)
- Use Roboflow to annotate bounding boxes and masks for each detection class
- Tile large images into manageable chunks.
- Version datasets and track model performance across training runs
- Continuous retraining: as installations complete, we validate model predictions against actual deployed systems and add high-value outliers back to the training data

Model Deployment and Inference:

Once the model is trained and meets expected performance metrics, it is exported, tagged and containerized. A FastAPI-based application server provides an interface to execute the model, convert inference results to vectors and serve them back to the clients. The model and the service run on managed Kubernetes services alongside other micro services.

Lessons Learned and Where We are Headed:

Recent developments in transfer learning and segmentation models such as SAM (and SAM Geo) greatly benefit computer vision applications for renewable energy. The choice of training and inference imagery data plays a significant part in the detection outcomes as much as the model architecture and flavor do. Because StationA is a fully integrated origination platform, the availability of actual solar panel and battery placements allows us to continuously refine and improve our computer vision models.

Space Optimization: Panel and BESS Packing

An adjacent problem in renewable energy is the determination of maximum energy potential for a given site once the available deployable area is determined using computer vision models. This problem involves packing solar panels and battery packs in the optimal configuration to maximize available space.

Our current algorithms utilize a grid-search approach to determine the best possible placements. We are poised to explore the application of deep learning models for this 2D space optimization problem. The resulting schematic should follow strict electrical and building code requirements while maximizing the usage of available space and avoiding the keep-out areas determined by the segmentation models.