Near-Surface Soil Moisture Characterization in Mississippi’s Highway Slopes Using Machine Learning Methods and UAV-Captured Infrared and Optical Images

Abstract

Near-surface soil moisture content variation is a major factor in the frequent shallow slope failures observed on Mississippi’s highway slopes built on expansive clay. Soil moisture content variation is monitored generally through borehole sensors in highway embankments and slopes. This point monitoring method lacks spatial resolution, and the sensors are susceptible to premature failure due to wear and tear. In contrast, Unmanned/Uncrewed Aerial Vehicles (UAVs) have higher spatial and temporal resolutions that enable more efficient monitoring of site conditions, including soil moisture variation. The current study focused on developing two methods to predict soil moisture content (θ) using UAV-captured optical and thermal combined with machine learning and statistical modeling. The first method used Red, Green, and Blue (RGB) color values from UAV-captured optical images to predict θ. Support Vector Machine for Regression (SVR), Extreme Gradient Boosting (XGB), and Multiple Linear Regression (MLR) models were trained and evaluated for predicting θ from RGB values. The XGB model and MLR model outperformed the SVR model in predicting soil moisture content from RGB values. The R2 values for the XGB and MLR models were >0.9 for predicting soil moisture when compared to SVR (R2 = 0.25). The Root Mean Square Error (RMSE) for XGB, SVR, and MLR were 0.009, 0.025, and 0.01, respectively, for the test dataset, affirming that XGB was the best-performing model among the three models evaluated, followed by MLR and SVR. The better-performing XGB and MLR models were further validated by predicting soil moisture using unseen input data, and they provided good prediction results. The second method used Diurnal Land Surface Temperature variation (ΔLST) from UAV-captured Thermal Infrared (TIR) images to predict θ. TIR images of vegetation-covered areas and bare ground areas of the highway embankment side slopes were processed to extract ΔLST amplitudes. The underlying relationship between soil surface thermal inertia and moisture content variation was utilized to develop a predictive model. The resulting single-parameter power curve fit model accurately predicted soil moisture from ΔLST, especially in vegetation-covered areas. The power curve fit model was further validated on previously unseen TIR, and it predicted θ with an accuracy of RMSE = 0.0273, indicating good prediction performance. The study was conducted on a field scale and not in a controlled environment, which aids in the generalizability of the developed predictive models.

Failure Mechanisms and Protection Measures for Expansive Soil Slopes: A Review

“…With the development of science and technology, intelligence gradually penetrates all fields, geotechnical engineering being no exception. For expansive soil slope disasters, it is possible to transition from traditional monitoring methods to novel intelligent monitoring methods [202][203][204]. Puppala et al [198] successfully incorporate visualization tools and unmanned aerial vehicle platforms for studying and monitoring the health of civil infrastructure built on expansive soils.…”

Integrated Edge Deployable Fault Diagnostic Algorithm for the Internet of Things (IoT): A Methane Sensing Application

“…Four different measures were used to assess the performance of the proposed algorithm. Accuracy [35] was the first metric used, as defined in Equation (12). The comparative outcomes between the classifiers are shown in Figure 8.…”

“…From the four-layer architecture it is understandable that the data collected from the environment has a greater impact over the entire IoT architecture, as any corruption in the SED-data can result in overall system failure [ 11 ]. SED faults occur for a variety of reasons, such as wear and tear [ 12 ], calibration error, physical damage, hostile environment due to heat [ 13 ], vibration [ 14 ], network failure or intentional tampering [ 15 ]. A SED may sustain the damage but then the fault must be identified as soon as it originates to save the data-driven system.…”

Comparison of Machine Learning Methods for Predicting Soil Total Nitrogen Content Using Landsat-8, Sentinel-1, and Sentinel-2 Images

“…GBM and XGBoost are both iterative models, with each model's prediction based on the residuals of the previous model. The models are sensitive to outliers, as a large outlier may affect the residuals of each model and result in a wider predicted range of STN content [51]. Based on measured soil data, the STN content ranged from 0.052 to 2.396 g•kg −1 .…”