Abstract: This article provides a formal correction to the research paper "LightGBM hybrid model based DEM correction for forested areas," originally published in PLOS ONE. The correction rectifies layout errors regarding figure placement while highlighting the study's core contribution: a sophisticated machine learning pipeline designed to improve Digital Elevation Models (DEMs) by accurately filtering out vegetation interference in forested regions.
The Challenge: Seeing Through the Canopy
Digital Elevation Models (DEMs), such as the COP30 dataset, are fundamental tools for environmental modeling and infrastructure planning. However, satellite sensors often struggle to distinguish between the top of a tree canopy and the actual ground surface. In densely forested areas, this leads to significant elevation errors, where the recorded data reflects a "surface model" rather than the true terrain. Existing correction methods frequently fail to account for the complex, non-linear relationships between canopy density, terrain slope, and sensor bias.
Methodology: A Multi-Stage Hybrid Pipeline
To address these inaccuracies, the researchers developed a high-precision hybrid machine learning architecture. The model, termed DMFDEM, utilizes a multi-stage approach to "strip away" vegetation and reveal the true topography beneath.
As detailed in the model flowchart, the pipeline consists of four primary components:
- Feature Extraction: A Convolutional Neural Network (CNN) processes spatial features from raw DEM data and GIS factors (such as slope and canopy height) using a "9-box grid" to capture local terrain texture.
- Regression Engine: LightGBM (Light Gradient Boosting Machine) serves as the core predictor, chosen for its efficiency with large-scale datasets.
- Bio-inspired Optimization: The Firefly Algorithm (FA) and Sparrow Search Algorithm (SSA) are used in tandem for hyperparameter optimization, preventing the model from becoming stuck in local minima during training.
- Validation: The model was trained and tested against high-precision ICESat-2 LiDAR data, which serves as the "ground truth" reference.
Global Validation and Datasets
To ensure the model's robustness across diverse ecosystems, the researchers tested it in four distinct global biomes, ranging from tropical rainforests to boreal environments.
The study sites included:
- Kalimantan, Indonesia: Tropical rainforest.
- Florida, USA: Subtropical wetlands.
- California, USA: Temperate forests with high topographic relief.
- Pärnumaa, Estonia: Boreal forest environments.
Key Results and Insights
The hybrid FA-SSA-CNN-LightGBM model demonstrated a significant performance leap over traditional methods. While standard Support Vector Regression (SVR) models achieved an $R^2$ of 0.63, the proposed hybrid pipeline reached an $R^2$ of 0.91.
Key findings include:
- Spatial Context: The use of a CNN front-end allows the model to understand the surrounding terrain context rather than treating pixels in isolation.
- Generalizability: The model maintained high accuracy across vastly different biomes, suggesting it captures fundamental physical principles of remote sensing errors in vegetation.
- Visual Fidelity: Transect plots confirmed that the DMFDEM closely tracks ICESat-2 ground hits, effectively ignoring canopy returns that typically pull uncorrected models upward.
Conclusion
By combining deep learning with gradient boosting and bio-inspired optimization, this research provides a powerful new tool for generating accurate terrain models in forested regions. While computationally intensive, the model's ability to generalize across global biomes marks a significant step forward in the field of remote sensing and topographic mapping.
Summary updated