Terrain Characterization via Machine vs. Deep Learning Using Remote Sensing.

Area of Science:

• Geosciences
• Remote Sensing
• Artificial Intelligence

Background:

• Terrain traversability is crucial for ground vehicle mobility and mission success.
• Current in-situ soil measurements are time-consuming, costly, and potentially dangerous.
• Remote sensing offers a potential alternative for terrain property assessment.

Purpose of the Study:

• To investigate the use of UAV-based remote sensing (thermal, multispectral, hyperspectral) for terrain property prediction.
• To compare the efficacy of machine learning and deep learning algorithms in estimating soil moisture and strength.
• To develop rapid, cost-efficient, and safer methods for mobility mapping.

Main Methods:

• Utilized thermal, multispectral, and hyperspectral remote sensing data from a UAV platform.
• Applied various machine learning (Linear, Ridge, Lasso, PLS, SVM, KNN) and deep learning (MLP, CNN) algorithms.
• Correlated predicted soil properties (moisture, cone penetrometer strength) with vehicle performance metrics (wheel slip, speed).

Main Results:

• Deep learning models significantly outperformed traditional machine learning models.
• A Multi-Layer Perceptron (MLP) achieved the highest accuracy in predicting soil moisture (R²=0.97) and soil strength (CP06: R²=0.95, CP12: R²=0.92).
• Observed correlations between predicted soil strength and vehicle performance (wheel slip, speed) using a Polaris MRZR.

Conclusions:

• UAV-based remote sensing coupled with deep learning provides a viable, efficient, and safe alternative to in-situ measurements for terrain analysis.
• The developed prediction maps can be effectively applied to enhance ground vehicle mobility assessments.
• This methodology holds significant potential for military and civilian applications requiring rapid terrain evaluation.