Extended Data Fig. 3: Evaluation of rooftop area quantification using deep learning.
From: Worldwide rooftop photovoltaic electricity generation may mitigate global warming

This figure evaluates the prediction accuracy of our deep learning model, which quantified rooftop area compared with manually labelled data in an independent test dataset. This comparison was conducted using randomly selected image patches, each measuring 512 × 512 pixels. a, Overview of all test image patches (total number = 2,951). b-u, Breakdown of test image patches by geographic region, showing the specific number of patches evaluated in each area. b, Australia and New Zealand (36 patches). c, Pacific Islands (136 patches). d, Caribbean (173 patches). e, Northern America (36 patches). f, Central America (99 patches). g, South America (164 patches). h, Eastern Asia (87 patches). i, South-Eastern Asia (129 patches). j, Southern Asia (127 patches). k, Western Asia (266 patches). l, Central Asia (90 patches). m, Northern Africa (101 patches). n, Eastern Africa (206 patches). o, Southern Africa (74 patches). p, Western Africa (182 patches). q, Middle Africa (96 patches). r, Northern Europe (306 patches). s, Eastern Europe (220 patches). t, Southern Europe (252 patches). u, Western Europe (171 patches).