Rooftop PV modules mapped without per-module labels: MV-PVMapper recovers 7,710 of 7,711 modules across three plants
Synopsis
On five-direction UAV oblique surveys at roughly 1 cm nadir ground sampling distance, MV-PVMapper couples the frozen vision foundation model SAM 3 with photogrammetric geometry to output georeferenced polygons for individual modules without task-specific annotation or weight updates. Across three operational rooftop PV plants holding 7,711 reference modules, two feedback rounds match 7,710 with no false positives at an IoU threshold of 0.5; macro-averaged AJI reaches 97.55%, against 60.73% for an orthomosaic segmentation baseline and 41.24% for a multi-view pixel-voting baseline.
Interpretation
A module map can now give each module its own boundary and location, rather than only marking the area covered by PV arrays. Earlier routes either depended on rule assumptions about colour, line structure, shape or array layout, or required annotated UAV images to train detectors and segmentation networks. On five-direction surveys of three operational rooftop plants, two feedback rounds match 7,710 of 7,711 reference modules with no false positives at an IoU threshold of 0.5, and macro-averaged area precision and recall reach 99.00% and 98.52%.
The method first uses a text prompt to let SAM 3 propose module candidates on the original images, then transfers their contours into one common object-space frame for geometric verification. Per-module slope correction fits a local plane to visible surface points with random sample consensus and intersects contour rays with that plane, avoiding the edge bending that comes from intersecting directly with the digital surface model. Across the three plants, slope correction projection reaches an AJI of 97.55% and average precision of 95.67% at an IoU threshold of 0.95, against 97.24% and 93.46% for affine-based projection and 89.94% and 46.56% for collinearity-based projection.
Accepted module instances return to every image covering that location as box prompts, adding spatial cues that a text prompt alone cannot supply, and enter the next round of segmentation, verification and fusion. Returning the fused result only to its source image, or letting each image re-prompt from its own predictions, does not reach the boundary accuracy of covering-view routing. Covering-view routing reaches 95.67% average precision at an IoU threshold of 0.95, against 86.77% for source-view routing and 79.88% for image-space self re-prompting; at Site 3, where omissions concentrate, covering-view routing misses one module while the other two routes miss 117 and 258.
Perspective
The demonstrated setting is five-direction UAV oblique surveys of three operational rooftop PV plants at a nadir ground sampling distance of about 0.93 to 1.11 cm per pixel, with original images, reconstructed camera poses, a digital surface model and one representative module size as inputs. The method targets offline map generation, and each feedback round reprocesses every sliding window, so the present implementation is not suited to real-time inspection.
The evaluated sites do not cover rooftops with highly irregular installations and several substantially different module specifications, where a single representative size may reject valid modules or admit fragments of larger ones. Lower image resolution, severe occlusion and larger installation tilts still require further evaluation, and the segmentation backbone is held fixed at SAM 3 without comparing several models.