Abstract. Building height and footprint are two fundamental urban morphological features required by urban climate modelling. Although some statistical methods have been proposed to estimate average building height and footprint from publicly available satellite imagery, they often involve tedious feature engineering which makes it hard to achieve efficient knowledge discovery in a changing urban environment with ever-increasing earth observations. In this work, we develop a deep-learning-based (DL) Python package – SHAFTS (Simultaneous building Height And FootprinT extraction from Sentinel imagery) to extract such information. Multi-task deep-learning (MTDL) models are proposed to automatically learn feature representation shared by building height and footprint prediction. Besides, we integrate digital elevation model (DEM) information into developed models to inform models of terrain-induced effects on the backscattering displayed by Sentinel-1 imagery. We set conventional machine-learning-based (ML) models and single-task deep-learning (STDL) models as benchmarks and select 46 cities worldwide to evaluate developed models’ patch-level prediction skills and city-level spatial transferability at four resolutions (100, 250, 500 and 1000 m). Patch-level results of 43 cities show that DL models successfully produce discriminative feature representation and improve the coefficient of determination (R2) of building height and footprint prediction more than ML models by 0.27–0.63 and 0.11–0.49, respectively. Moreover, stratified error assessment reveals that DL models effectively mitigate the severe systematic underestimation of ML models in the high-value domain: for the 100 m case, DL models reduce the root mean square error (RMSE) of building height higher than 40 m and building footprint larger than 0.25 by 31 m and 0.1, respectively, which demonstrates the superiority of DL models on refined 3D building information extraction in highly urbanized areas. For the evaluation of spatial transferability, when compared with an existing state-of-the-art product, DL models can achieve similar improvement on the overall performance and high-value prediction. Furthermore, within the DL family, comparison in building height prediction between STDL and MTDL models reveals that MTDL models achieve higher accuracy in all cases and smaller bias uncertainty for the prediction in the high-value domain at the refined scale, which proves the effectiveness of multi-task learning (MTL) on building height estimation.
Weather radar plays an important role in accurate weather monitoring and modern weather forecasting, as it can provide timely and refined weather forecasts for the public and for decision makers. Deep learning has been applied in radar nowcasting tasks and has exhibited a better performance than traditional radar echo extrapolation methods. However, current deep learning-based radar nowcasting models are found to suffer from a spatial “blurry” effect that can be attributed to a deficiency in spatial variability representation. This study proposes a Spatial Variability Representation Enhancement (SVRE) loss function and an effective nowcasting model, named the Attentional Generative Adversarial Network (AGAN), to alleviate this blurry effect by enhancing the spatial variability representation of radar nowcasting. An ablation experiment and a comparison experiment were implemented to assess the effect of the generative adversarial (GA) training strategy and the SVRE loss, as well as to compare the performance of the AGAN and SVRE loss function with the current advanced radar nowcasting models. The performances of the models were validated on the whole test set and inspected in two storm cases. The results showed that both the GA strategy and SVRE loss function could alleviate the blurry effect by enhancing the spatial variability representation, which helps the AGAN to achieve better nowcasting performance than the other competitor models. Our study provides a feasible solution for high-precision radar nowcasting applications.
<p>Buildings are common components in the urban environment whose 3D information is fundamental for urban hydrometeorological modeling and planning applications. In order to monitor building footprint and height across large areas on a regular basis, recent earth observation research has witnessed promising progress in mapping such information from publicly available satellite imagery by statistical methods using regression between multi-source remotely sensed data and target variables. However, most of them often involve tedious feature preprocessing, which constrains their capability to establish a comprehensive representation of an ever-changing and multi-scale urban system efficiently.</p> <p>Considering this bottleneck, this work develops a deep-learning-based (DL) Python package-SHAFTS (Simultaneous building Height And FootprinT extraction from Sentinel Imagery) to estimate 3D building information at various scales. SHAFTS provides Convolutional Neural Networks (CNN) with the Multi-Branch Multi-Head (MBMH) structure to automatically learn representative features shared by building height and footprint mapping tasks from multi-modal Sentinel imagery and additional background DEM information. Besides, to leverage the power of big data infrastructures, SHAFTS offers essential functionality including automatically collecting potential reference datasets by web scraping and filtering appropriate input imagery from Google Earth Engine, which can effectively ease model upgrading and deployment for large-scale mapping.</p> <p>To evaluate the patch-level prediction skills and city-level spatial transferability of developed models, this work performs diagnostic performance comparisons in 46 cities worldwide by using conventional machine-learning-based (ML) models and CNN with the Multi-Branch Single-Head (MBSH) structure as benchmarks. Patch-level results show that DL models successfully produce more discriminative feature representation and improve the coefficient of determination of building height and footprint prediction over ML models by 0.27-0.63, 0.11-0.49, respectively. Moreover, stratified error assessment reveals that DL models effectively mitigate severe systematic underestimation of ML models in the high-value domain. Additionally, within the DL family, comparison in spatial transferability demonstrates that the MBMH structure improves the accuracy of CNN and reduces the uncertainty of building height predictions in the high-value domain at the refined scale. Therefore, multi-task learning can be considered as a possible solution for improving the generalization ability of models for 3D building information mapping.</p>
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.