Results of the summarizing process for the arXiv paper: 2401.09481v1

, , , , In the field of earth observation, three-dimensional (3D) data has become increasingly prevalent in recent years. Despite this growth, many 3D surveys remain underutilized due to the lack of accessible and explainable automatic classification methods, particularly for emerging technologies like topo-bathymetric lidar data. To address this gap, a new workflow called 3DMASC (Multiple Attributes, Scales, and Clouds) has been introduced for explainable machine learning in 3D data classification. The 3DMASC workflow incorporates multi-cloud classification by utilizing dual-cloud features that encrypt local spectral and geometrical ratios and differences. It leverages classical multi-scale descriptors that are adaptable to various types of 3D point clouds, along with new descriptors based on spatial variations. This approach has been applied to the multi-class classification of topo-bathymetric lidar data in coastal and fluvial environments. Through the use of multivariate and embedded feature selection techniques, optimized predictor sets with reduced complexity have been developed using 3DMASC. The study identifies key features that are particularly relevant for describing coastal and riverine scenes, highlighting the significance of dual-cloud features, lidar return-based attributes averaged over specific scales, as well as statistics derived from dimensionality-based and spectral features. The results demonstrate that small to medium spherical neighborhood diameters (<7 m) can effectively build classifiers when combined with distance-to-ground or distance-to-water-surface features. Notably, without relying on optional RGB information and utilizing a maximum of 37 descriptors, classification accuracies ranging from 91% for complex multi-class tasks to 98% for lower-level processing have been achieved using models trained on fewer than 2000 samples per class. Comparisons with traditional point cloud classification methods reveal that the descriptive power of 3DMASC features significantly outperforms existing approaches. Furthermore, the contributions of this work are made accessible through a plugin in the CloudCompare software, enabling non-specialist users to create classifiers for various types of 3D data characterized by one or two point clouds (such as airborne or terrestrial lidar). Additionally, two labeled topo-bathymetric lidar datasets are available for further exploration on https://opentopography.org/.

• - Three-dimensional (3D) data is increasingly prevalent in earth observation
• - Lack of accessible and explainable automatic classification methods for 3D surveys, especially for emerging technologies like topo-bathymetric lidar data
• - Introduction of the 3DMASC workflow for explainable machine learning in 3D data classification
• - Utilization of dual-cloud features, multi-scale descriptors, and new descriptors based on spatial variations in the 3DMASC workflow
• - Application of 3DMASC to multi-class classification of topo-bathymetric lidar data in coastal and fluvial environments
• - Development of optimized predictor sets with reduced complexity using multivariate and embedded feature selection techniques
• - Identification of key features relevant for describing coastal and riverine scenes, including dual-cloud features, lidar return-based attributes averaged over specific scales, and statistics derived from dimensionality-based and spectral features
• - Effective building of classifiers with small to medium spherical neighborhood diameters combined with distance-to-ground or distance-to-water-surface features
• - Achievement of high classification accuracies ranging from 91% to 98% using models trained on fewer than 2000 samples per class without relying on optional RGB information
• - Outperformance of existing approaches by the descriptive power of 3DMASC features in comparison to traditional point cloud classification methods
• - Accessibility of contributions through a plugin in CloudCompare software for non-specialist users to create classifiers for various types of 3D data characterized by one or two point clouds such as airborne or terrestrial lidar

Summary- 3D data, which shows objects in three dimensions, is becoming more common in studying the Earth. - There aren't many easy ways to automatically classify this type of data, especially for new technologies like lidar. - A new method called 3DMASC helps explain how machines learn to classify 3D data. - This method uses different features and descriptors to understand and classify the data better. - By using 3DMASC, scientists can accurately classify different types of lidar data found near coasts and rivers. Definitions- Three-dimensional (3D): Showing objects with height, width, and depth. - Data: Information or facts collected for analysis or reference. - Lidar: A technology that uses lasers to measure distances on the Earth's surface.

Introduction

In recent years, the use of three-dimensional (3D) data in earth observation has grown significantly. However, many 3D surveys are not being fully utilized due to the lack of accessible and explainable automatic classification methods. This is especially true for emerging technologies like topo-bathymetric lidar data. To address this gap, a new workflow called 3DMASC (Multiple Attributes, Scales, and Clouds) has been introduced for explainable machine learning in 3D data classification.

The 3DMASC Workflow

The 3DMASC workflow incorporates multi-cloud classification by utilizing dual-cloud features that encrypt local spectral and geometrical ratios and differences. It also leverages classical multi-scale descriptors that are adaptable to various types of 3D point clouds, along with new descriptors based on spatial variations. This approach has been applied to the multi-class classification of topo-bathymetric lidar data in coastal and fluvial environments. Through the use of multivariate and embedded feature selection techniques, optimized predictor sets with reduced complexity have been developed using 3DMASC.

Key Features Identified

The study identifies key features that are particularly relevant for describing coastal and riverine scenes. These include: - Dual-cloud features - Lidar return-based attributes averaged over specific scales - Statistics derived from dimensionality-based and spectral features The results demonstrate that small to medium spherical neighborhood diameters (<7 m) can effectively build classifiers when combined with distance-to-ground or distance-to-water-surface features.

Performance Results

Using models trained on fewer than 2000 samples per class, the researchers were able to achieve high classification accuracies ranging from 91% for complex multi-class tasks to 98% for lower-level processing. This was achieved without relying on optional RGB information and utilizing a maximum of 37 descriptors. Comparisons with traditional point cloud classification methods reveal that the descriptive power of 3DMASC features significantly outperforms existing approaches.

Accessibility and Availability

The contributions of this work are made accessible through a plugin in the CloudCompare software, enabling non-specialist users to create classifiers for various types of 3D data characterized by one or two point clouds (such as airborne or terrestrial lidar). Additionally, two labeled topo-bathymetric lidar datasets are available for further exploration on https://opentopography.org/.

Conclusion

In conclusion, the 3DMASC workflow offers a new approach to explainable machine learning in 3D data classification. By incorporating multi-cloud classification and leveraging classical and new descriptors, it has shown promising results in classifying topo-bathymetric lidar data in coastal and fluvial environments. The accessibility and availability of this work make it a valuable tool for non-specialists looking to utilize 3D data in their research.