Fully convolutional networks for land cover classification from historical panchromatic aerial photographs

Abstract

Historical aerial photographs provide salient information on the historical state of the landscape. The exploitation of these archives is often limited by accessibility and the timeconsuming process of digitizing the analogue copies at a high resolution and processing them with a proper photogrammetric workflow. Furthermore, these data are characterised by limited spectral information since it occurs very often in a single band. Our work presents a first application of deep learning for the extraction of land cover from historical aerial panchromatic photographs of the African cities of Goma, Bukavu and Bujumbura. We evaluate the suitability of deep learning for land cover generation from a challenging dataset of photographs from the 1940s and 1950s that covers large geographical extents and is characterised by radiometric variations between dates and locations. A fully convolutional approach is investigated by considering two network architectures with different strategies of exploiting contextual information: one used atrous convolutional layers without downsampling, whereas the second network has both downsampling and learned upsampling convolutional layers (U-NET). The networks are trained to detect three main classes namely, buildings, high vegetation and a mixed class of bare land and low vegetation class. High overall accuracies of > 90% in Goma-Gisenyi and Bukavu, and >85 % in Bujumbura are obtained. This work provides a novel methodology that outperforms a baseline standard machine learning classifier for the exploitation of the vast archives of historical aerial photographs that can aid long-term environmental baseline studies. Future work will entail developing domain adaptation strategies in order to make the trained network robust for different image mosaics.

Characteristics and Distribution of Landslides in the Populated Hillslopes of Bujumbura, Burundi

“…The historical aerial photographs available at the Royal Museum for Central Africa (Belgium) for the periods 1957-1959 and 1974 were also used to obtain a stereoscopic vision of the landscape morphology. Note that the quality of the photographs is not always optimal as they have been preserved in paper format and, in some cases, suffer from low-quality imaging (e.g., blurring, under-and overexposure) and ageing effects [48]. A 1-m spatial resolution orthomosaic was built from the 1957-1959 photographs by applying recent Multiview Stereo Photogrammetry (MVS) approaches [48][49][50] in Metashape Pro [51].…”

“…Note that the quality of the photographs is not always optimal as they have been preserved in paper format and, in some cases, suffer from low-quality imaging (e.g., blurring, under-and overexposure) and ageing effects [48]. A 1-m spatial resolution orthomosaic was built from the 1957-1959 photographs by applying recent Multiview Stereo Photogrammetry (MVS) approaches [48][49][50] in Metashape Pro [51]. We also visually interpreted a set of products (slope angle, contour lines and hillshade) that were derived from a 1-m very high spatial resolution Digital Elevation Model (DEM).…”

“…The signatures of shallow landslide scars tend to disappear quickly from the landscape, sometimes already after a few months. Hence, all mapped shallow landslides have been detected in recent high-resolution satellite images and/or in the field; their identification on the historical aerial photographs being difficult because of their lower quality [48]. The smallest landslides observed in the field were not visible on the satellite images either, especially when they occurred alongside road cuts or river banks.…”

MAGI: Multistream Aerial Segmentation of Ground Images with Small-Scale Drones

“…The AMSM is composed of 3 branches, each performing convolutions with different dilation rates, while the AFM fuses the information of shallow and deep feature maps generated by the several branches. Lastly, in [41] an FCN with skip connection was designed for segmentation in panchromatic aerial images, obtaining an overall accuracy of >90% in the best case.…”

GIS-Based Weights-of-Evidence Model for Landslide Susceptibility Mapping and Assessment in the Hillslopes of Bujumbura, Burundi

“…2). Land use is classi ed in three units (1) the built-up area with a high extension due to the high demographic pressure over the last few decades, following the rural exodus of the massive population to settle in the urban and peri-urban area of Bujumbura, (2) the more abundant forest in the East which is gradually disappearing towards the west to cope with human activities and (3) agricultural land with low to moderate vegetation and scattered trees (Mboga et al 2020).…”