Object Height Estimation on Monoscopic Satellite Images Using Deep Learning

Abstract

Conventional methodologies for deriving 3D information from satellite optical images often rely on intricate algorithms requiring substantial human supervision, numerous external parameters, and the acquisition of multiple images captured from disparate perspectives. This study aims to eliminate the need for manual parameter tuning and multiple image acquisition and completely automate the extraction of 3D object information from monoscopic satellite imagery by leveraging advanced deep learning algorithms. The Fused Height Estimation (Fused-HE) deep network model, which leverages the inherent characteristics of satellite imagery, is proposed within this scope. The model, featuring dual encoders, processes monoscopic satellite image inputs individually through the convolutional encoder for local feature extraction within neighboring pixels and the vision transformers encoder for global feature extraction using the relationships between objects in the image. The distinct feature outputs from the two encoders are concatenated in feature fusion blocks. Then, fused features are passed to the decoder blocks, and the height head produces the predicted heights. The proposed fused network is further improved by introducing an additional segmentation head to the model to assign height values to the correct pixel, and this model is named the Fused Segmentation Height Estimation (FusedSeg-HE). Comprehensive evaluations of individual models in the literature and the proposed fused networks demonstrate the proposed models provide both local and global feature extraction, reduce root mean squared error by approximately 5%-13%, and increase accuracy by 4%-9% in delta threshold according to the metric results