Object-based correction of LiDAR DEMs using RTK-GPS data and machine learning modeling in the coastal Everglades

• Published in: Environmental modelling & software : with environment data news 2019-02, Vol.112, p.179-191
• Main Authors: Cooper, Hannah M., Zhang, Caiyun, Davis, Stephen E., Troxler, Tiffany G.
• Format: Article
• Language: English
• Subjects: Artificial intelligence; Artificial neural networks; Bias; Coastal environments; Coastal wetlands; Computer simulation; DEMs; Digital Elevation Models; Environment models; Error correction; Error reduction; Global positioning systems; GPS; Imagery; Learning algorithms; LiDAR; Machine learning; Monte Carlo; Monte Carlo simulation; Neural networks; Object-based image analysis; Positioning devices (machinery); Root-mean-square errors; Satellite navigation systems; Support vector machines; Wetlands
• ISSN: 1364-8152; 1873-6726
• DOI: 10.1016/j.envsoft.2018.11.003

Light Detection and Ranging (LiDAR) Digital Elevation Models (DEMs) are frequently applied in modeling coastal environments. We present an object-based correction approach for accurate and precise DEMs by integrating LiDAR point data, aerial imagery, and Real Time Kinematic-Global Positioning Systems. Four machine learning techniques (Random Forest, Support Vector Machine, k-Nearest Neighbor, and Artificial Neural Network) were compared with the commonly used bias-correction method. The Random Forest object-based model produced best predictions for two study areas: Nine Mile (Mean Bias Error (MBE) reduced 0.18 to −0.02 m, Root Mean Square Error (RMSE) reduced 0.22 to 0.08 m) and Flamingo (MBE reduced 0.17 to 0.02 m, RMSE reduced 0.24 to 0.10 m). A Monte Carlo model was developed to combine errors into the object-based machine learning corrected DEMs, and uncertainty maps spatially revealed the likelihood of error. The object-based correction approach provides an attractive alternative to the bias-correction method. •Object-based correction reduces both bias and error when compared to bias-correction.•Object-based correction addresses accuracy and error difference within a plant community.•Object-based Random Forest model performed best in two separate study areas.•Monte Carlo simulation combines bias and error into corrected DEMs for more reliable products.