This study aimed to meet the need for high-resolution (10-meter) building height estimation models by establishing a comprehensive spatial-spectral-temporal feature database. The database combined data from Sentinel-1 SAR, Sentinel-2 optical, and building footprint sources, extracting 160 statistical features.
To select the most relevant features, the study utilized a combination of feature importance methods, including Random Forest, Shapley Additive Explanations, and Permutation Feature Importance. The final 13 stable features were used to train a Random Forest regression model.
To address challenges such as the double-bounce effect and building shadows, the study implemented object-based modeling with a 50-meter buffer around buildings and a 50x50-meter moving window for the final pixel-based predictions. This approach helped mitigate the impact of these issues on the high-resolution data.
The model was trained on building height data from 12 cities across the United States and achieved an R-squared of 0.78 on the test set. The study then applied the model to estimate building heights across the state of Iowa, providing insights into the distribution of building heights at the county level.
The authors note that while the model performed well overall, there were some challenges in accurately estimating the heights of very low and very tall buildings, which they attribute to factors such as inaccuracies in the reference data, the double-bounce effect, and the spatial resolution limitations. Future work will explore the use of deep learning algorithms to further improve the model's performance.
To Another Language
from source content
arxiv.org
Deeper Inquiries