View article

Domain shifts during seasonal variations are an important aspect affecting the robustness of aerial scene classification and so it is crucial that such variation is captured within aerial scene datasets. This is more evident in geographic locations in the global South, where aerial coverage is scarcer and the rural and semi-urban landscape varies dramatically between wet and dry seasons. As current datasets do not offer the ability to experiment with domain shifts due to seasonal variations, this work proposes a labelled dataset for classifying land use from aerial images, comprising both wet and dry season data from Ghaziabad in India. Moreover, we conduct a thorough investigation into how image features, namely colour, shape, and texture, influence the accuracy of scene classification. We demonstrate that a combination of an architecture that extracts salient features, with the implementation of a larger receptive field improves classification performance when applied to both shallow or deep architectures by extracting invariant feature representations across domains.