Uncertainty and Overfitting in Fluvial Landform Classification Using Laser Scanned Data and Machine Learning: A Comparison of Pixel and Object-Based Approaches

Abstract

Floodplains are valuable scenes of water management and nature conservation. A better understanding of their geomorphological characteristic helps to understand the main processes involved. We performed a classification of floodplain forms in a naturally developed area in Hungary using a Digital Terrain Model (DTM) of aerial laser scanning. We derived 60 geomorphometric variables from the DTM and prepared a geomorphological map of 265 forms (crevasse channels, point bars, swales, levees). Random Forest classification was conducted with Recursive Feature Elimination (RFE) on the objects (mean pixel values by forms) and on the pixels of the variables. We also evaluated the classification probabilities (CP), the spatial uncertainties (SU), and the overfitting in the function of the number of the variables. We found that the object-based method had a better performance (95%) than the pixel-based method (78%). RFE helped to identify the most important 13–20 variables, maintaining the high model performance and reducing the overfitting. However, CP and SU were not efficient measures of classification accuracy as they were not in accordance with the class level accuracy metric. Our results help to understand classification results and the specific limits of laser scanned DTMs. This methodology can be useful in geomorphologic mapping.

Automated Earthwork Detection Using Topological Persistence

“…For levee detection, found that PC for image segmentation, DEM analysis, and integrated analysis methods was about 30%, 95%, and 80%, and DT was about 45%, 65%, and 50%. Object-oriented and pixel-based classification methods applied by Csatáriné Szabó et al (2020) provided overall accuracies of 75%-80%. The PC and DT found using persistence rank among the highest-performing of the semi-automated methods they tested for levees.…”

“…Object‐oriented and pixel‐based classification methods applied by Csatáriné Szabó et al. (2020) provided overall accuracies of 75%–80%. The PC and DT found using persistence rank among the highest‐performing of the semi‐automated methods they tested for levees.…”

Elaborating Hungarian Segment of the Global Map of Salt-Affected Soils (GSSmap): National Contribution to an International Initiative

“…We selected RF not only because it has become a frequently applied machine learning technique in various fields (e.g. [53][54][55][56]), but it commonly outperforms other techniques as well [23,57]. Before fitting RF between the indicators and the environmental covariates listed in Table 3, we fine-tuned the hyper-parameter mtry of RF that is the number of input covariates selected randomly at each split.…”