Notes on explainable AI, computer vision, and what deep learning models actually learn — mostly written around the DeepPVMapper case study.
The Blind Men and the Solar Panel
Grid connection registries, remote sensing pipelines, and open mapping communities each describe "a PV installation" differently — and each is right, given what it's built to see. Reconciling them means agreeing on a shared ontology first, not just comparing numbers.
What Does a Solar Panel Detector Actually See?
While developing DeepPVMapper, the recurring question was never how accurate the model was, but whether it was actually detecting solar panels. Answering that turned out to require a new way of explaining models — decomposing predictions into wavelet scales — and revealed that the model was, in fact, an excellent grid detector.