Radiometric Correction of Historical Aerial Imagery using Convolutional Neural Networks : A Metadata-Conditioned Deep Learning Approach

This thesis investigates the potential of convolutional neural networks (CNNs) to address radiometric distortions in historical aerial imagery. The imagery maintained by Lantmäteriet, while central to this study, is just one of many valuable resources for geospatial applications, documenting environmental and urban changes over dec-ades. However, distortions caused by factors such as vignetting and hot spots limit its utility. Current manual correction processes are labour-intensive, error-prone, and unsuitable for managing large archives, emphasising the importance of a scalable and automated solution.The proposed approach leverages image metadata to enhance the correction of these distortions. Specifically, a multi-scale CNN architecture (MIRNet-v2) is employed and extended into a metadata-conditioned variant (MIRNet-MC) by incorporating solar elevation and azimuth information via Feature-wise Linear Modulation (FiLM). The FiLM mechanism introduces metadata-dependent scaling and shifting at multi-ple layers of the network, effectively conditioning the model’s feature representa-tions on illumination context.A comprehensive evaluation compares the baseline MIRNet-v2 model to the pro-posed MIRNet-MC. Quantitative results demonstrate that MIRNet-MC outper-forms the metadata-agnostic baseline in peak signal-to-noise radio (PSNR) while de-livering nearly equivalent structural similarity index measure (SSIM) performance. Specifically, MIRNet-MC achieves a mean PSNR about 0.5 dB higher than MIRNet-v2 indicating a notable reduction in radiometric error, whereas SSIM values remain virtually identical. This suggests that incorporating solar metadata yields a modest but consistent improvement in fidelity, particularly in aiding the correction of ex-treme illumination artefacts such as hot spot maxima.Qualitative assessment further emphasises the benefits of metadata conditioning. A small-scale expert survey comparing CNN-corrected images to manually corrected ground truth images indicates that the outputs for MIRNet-MC are on par with manual corrections in overall visual quality. Participants observed that MIRNet-MC effectively mitigates vignetting and overexposure artefacts with its results often matching or even slightly exceeding the perceptual quality of hand-corrected im-ages. These findings indicate that the CNN-based approach can bridge the gap be-tween raw scans and expert-adjusted photographs offering a viable automated solu-tion for large historical image archives.Overall, this metadata-conditioned CNN approach marks a significant step toward scalable radiometric correction, reinforcing the value of historical aerial imagery as a sustainable geospatial resource.