Spatially Adaptive Convolutional Networks with Coordinate-Conditioned Layers

In this study, we present a convolutional neural network (CNN) architecture, GeoConv, designed to improve the accuracy and adaptability of deep learning models using satellite imagery. Traditional CNNs, such as ResNet18, employ fixed-weight convolutional layers - i.e., layers that leverage the same set of weights for each input observation. However, these models can struggle to capture context-specific features inherent in satellite images, which may vary significantly across different geographic regions. To address this challenge, the GeoConv model utilizes dynamic weights that adapt based on the input image coordinates, allowing the model to tailor its feature extraction process to the unique characteristics of different geographic regions. Through experiments, we illustrate the utility of this approach in a case study which leverages satellite imagery to estimate household wealth across 11 countries, with GeoConv explaining an additional 10.12% of the variance in the data compared to a ResNet18 model. These results underscore the importance of incorporating spatially adaptive mechanisms in handling the variability present in satellite imagery. Code is available at: https://github.com/heatherbaier/geoconv.