Stereollax Net: Stereo Parallax-Based Deep Learning Network for Building Height Estimation

Accurate estimation of building heights is crucial foreffective urban planning and resource management as it providesessential geometric information about the urban landscape. Manyend-to-end deep learning-based networks have been proposedfor image-to-height mapping using high-resolution nonopticaland optical remote sensing imagery. In this study, we developa novel deep-learning architecture that incorporates a stereoparallax-based mathematical formulation for building heightestimation. We estimate stereo formulation parameters includedifferential parallax (1P) image, average photo-base (b), andsatellite height (hs). The final height map is computed byutilizing these parameters in the stereo parallax equation, thuscombining closed-form solutions within the learning paradigm.Moreover, to improve the estimation of1P, we also introduce amultiscale differential shortcut connection (MSDSC) module. TheMSDSC module integrates high-frequency components into lowerresolution baseline decoder features while converting them intohigh-resolution decoder features. To establish the efficacy of ourproposed stereo parallax-based deep learning network (StereollaxNet), we train and evaluate our method on densely populatedcities of China (42-Cities dataset) and on the IEEE Data FusionContest 2018 (DFC2018) dataset. Our proposed Stereollax Netis trained only with RGB imagery and compared with thestate-of-the-art (SOTA) methods that utilize both panchromaticand multispectral (RGB and near-infrared) satellite imagery.The qualitative and quantitative results demonstrate that ourStereollax Net surpasses existing SOTA algorithms, achievingsuperior performance with fewer data and training parametersby a considerable margin. The code will be made publiclyavailable via the GitHub repository