Please use this identifier to cite or link to this item:
Title: A generalized approach to soil strength prediction with machine learning methods
Authors: Seman, Peter M.
Keywords: Airplanes
Neural networks (Computer science)
Soils--Remote sensing
Publisher: Cold Regions Research and Engineering Laboratory (U.S.)
Engineer Research and Development Center (U.S.)
Series/Report no.: ERDC/CRREL ; TR-06-15.
Abstract: Current methods for evaluating the suitability of potential landing sites for fixed-wing aircraft require a direct measurement of soil bearing capacity. In contingency military operations, the commitment of ground troops to carry out this mission prior to landing poses problems in hostile territory, including logistics, safety, and operational security. Developments in remote sensing technology provide an opportunity to make indirect measurements that may prove useful for inferring basic soil properties. However, methods to accurately predict strength from other fundamental geotechnical parameters are lacking, especially for a broad range of soil types under widely varying environmental conditions. To support the development of new procedures, a dataset of in situ soil pit test results was gathered from airfield pavement evaluations at 46 locations worldwide that encompass a broad variety of soil types. Many features associated with soil strength—including gradation, moisture content, density, specific gravity and plasticity—were collected along with California bearing ratio (CBR), a critical strength index used to determine the traffic loading that the ground can support. Machine learning methods—with advantages in nonlinear relationship mapping, nonparametric distribution treatment, superior generalization, and implicit modeling—were applied, hypothesizing these characteristics might make them better suited to geotechnical problems. Artificial neural network and k-nearest neighbor techniques were tested on plastic and non-plastic subsets of data and compared to conventional regression and existing CBR prediction methods. The machine learning models were able to halve the baseline error rate for plastic soils, but non-plastic soils showed no significant improvement. For both groups, normalized root mean square error (NRMSE) for generalization to new cases was approximately fifty percent for the best models. The high degree of variability for direct soil strength measurement methods limits the lowest possible NRMSE to approximately 25%, even before introducing any additional errors expected with remote sensing.
Description: Technical Report
Appears in Collections:Technical Report

Files in This Item:
File Description SizeFormat 
CRREL-TR-06-15.pdf1.21 MBAdobe PDFThumbnail