Super-resolving and composing building dataset using a momentum spatial-channel attention residual feature aggregation network

Abstract

Model generalizability is crucial in the deployment of deep learning (DL) techniques. When trained on specific datasets, generalizability problems arise across many applications of DL including building extractions. Apart from regularizing the training process, collecting data with distinctive characteristics or distributions can be a promising solution. Over the past decade, several open building datasets have been released. However, in practice, a single dataset cannot overcome the generalization error. By unifying the spatial resolution and spectral bands of different datasets, those datasets could be integrated to relieve the generalization error in building footprint extraction. In this work, we focused on the difference in the spatial resolution between different building datasets. We first examined state-of-the-art super-resolution methods and proposed our own method based on Residual Feature Aggregation Network (RFANet), which we named Momentum and Spatial-Channel Attention RFANet (MSCA-RFANet). We then benchmarked our MSCA-RFANet in a comparative study; our new method achieved higher performance on spatial resolution enhancement. Specifically, in the four times spatial resolution enhancement on the SWOOP 2010 Dataset, our MSCA-RFANet result’s peak signal-to-noise ratio (PSNR) of 30.72 dB exceeded that of RFANet (30.66 dB). Likewise, we achieved a lower mean squared error (MSE) of 36.64 compared to RFANet’s 36.94. With detailed benchmarks against Second-order Attention Network (SAN) and Residual Channel Attention Network (RCAN), we confirmed the superior performance of our method in enhancing the spatial resolution of high-spatial-resolution images. Then, we explored the impact of super-resolution resolution and data composition on building footprint extraction. Our building footprint extraction experiments demonstrated the positive impact of super-resolution and data composition. These promising results showed that our method is suitable to integrating existing public building dataset to overcome generalization error in DL-based building footprint extraction.