Surface Soil Moisture Estimate by Integration of Remote Sensing and Low-Cost Field Sensor Network for Shallow Landslide Stability Assessment in Unsaturated Soils

Soil Moisture (SM) estimate and monitoring play a key role in rain-induced shallow landslides risk assessment. In fact, this kind of ground instability can be markedly influenced by the propagation of the saturation front inside the unsaturated zone. This research aims at defining a methodology for the large-scale, high spatial resolution estimate of SM through the integration of low-cost, ground-based soil moisture measurements and Sentinel-2 A/B (S-2 A/B) multispectral images. At the end of 2019, two soil moisture monitoring networks, based on capacitive sensors (WaterScout SM100), were installed in the sites of Mendatica and Ceriana-Mainardo (Liguria, Italy), subjected in shallow slope instabilities triggered by rainfall. These sensors, calibrated for the soil specific characteristics of the sites, provided a volumetric soil water content dataset (θ [m3/m3 %]), monitored in a few points of the study areas (four points in Mendatica and five points in Ceriana-Mainardo) and at four different depths (-10 cm, -30 cm, -50 cm, -80 cm). Extensive experimental work, aimed at calibrating the sensors, choosing the most suitable installation methods and verifying the effectiveness of the measurements in the field, was crucial to ensure the validity of the field-monitoring data. Consequently, the obtained soil water content data was taken as ground-truth for the analysis of the correlation between soil moisture and reflectance of multispectral images at the available bands. The results of such analyses show a marked influence of land cover on Reflectance - Soil Moisture correlations. The characteristics of the Ceriana-Mainardo site made it possible, in a first analysis, to assume that land cover (complex cultivations: arable crops, vineyards, fruit orchards, olives, grassland, kitchen gardens) was homogeneous over the entire study area, allowing all soil moisture measurements made at the five measurement nodes to be combined into four datasets, one for each of the four measurement depths. The linear correlation between these data and reflectance values from S-2 A/B images, extracted in the proximity of the measurement nodes, were investigated. The results show that the mean of Red Edge 2, Red Edge 3, Broad NIR and NIR bands provide higher R2 coefficient for soil moisture (>20% m3/m3) referred to 10 cm depth, compared to single band correlations. This kind of analysis allows to have correlation (R2=0.56, number of data (ordered pair Reflectance - Soil Moisture): 302, STD=3.3% θ (m3/m3), MAE=2.1% θ (m3/m3)) useful to extract spatially distributed information of volumetric water content from S-2 A/B images on vegetated areas. A slightly different approach was applied for the Mendatica site, which has much more heterogeneous land cover conditions. Each of the four measurement nodes (M1, M3, M4, M5) in fact falls into a different land cover, so it was necessary to apply four different Reflectance - Soil Moisture correlations over the study area according to a specific land cover map with as many classes. Furthermore, based on the conclusions reached by analyzing the Ceriana-Mainardo site, the Mendatica analyses focused strictly on the top 10 cm depth. The characteristics of the four Reflectance - Soil Moisture correlations applied for the four measurement nodes in Mendatica are summarized as follows: • M1 (Band: mean of SWIR 1 and SWIR 2, R2=0.77, number of data: 53, STD=2.8% θ (m3/m3), MAE=2.3% θ (m3/m3), Land cover: agriculture) • M3 (Band: Broad NIR, R2=0.85, number of data: 53, STD=2.3% θ (m3/m3), MAE=2.3% θ (m3/m3), Land cover: woods) • M4 (Band: mean of Red Edge 2, Red Edge 3, Broad NIR and NIR, R2=0.82, number of data: 54, STD=4.2% θ (m3/m3), MAE=3.2% θ (m3/m3), Land cover: complex agriculture) • M5 (Band: Red Edge 3, R2=0.66, number of data: 83, STD=4.7% θ (m3/m3), MAE=4.0% θ (m3/m3), Land cover: agriculture near to woods). Considering a time window from two to three years (depending on soil moisture monitoring network functionality), the procedure resulted in a number of soil moisture maps of about 360 for Mendatica (two years observation) and 240 for Ceriana-Mainardo (three years observation). It is worth noting that the study areas are all vegetated (non-bare soil conditions). Therefore, the multispectral images used for SM estimate provide data on the reflectance of vegetation on the analyzed surfaces. However, this information is correlated with soil moisture conditions, and therefore vegetation reflectance is an indirect indicator of SM, at least in the first 10 cm depth. The soil moisture maps thus obtained were the main input to implement a model for shallow landslide stability assessment under partially saturated soil conditions. The model provides an estimate of the relationship between soil moisture and suction (Water Retention Curve (WRC)), which allows the apparent cohesion contribution to be calculated in the shear strength estimate. WRC assessment requires the estimate of some physical characteristics of the soil present on the study area, including the percentage of sand, clay and carbon. The final results are maps of soil shear strength and factor of safety according to a LEM (Limit Equilibrium Method) approach. With reference to two sites in Liguria (Ceriana-Mainardo and Mendatica), the soil moisture data deduced from the methodology presented in this dissertation were used to map the susceptibility to shallow landslides under partially saturated conditions, where stability is precisely governed by soil moisture affecting both driving and resisting forces. It is worth underlining that the workflow here presented can be extended to other areas, depending on the availability of field soil moisture measurements for the initial calibration. The workflow allowing to obtain surface soil moisture and slope factor-of-safety maps from Sentinel-2 images is fully automatable and will be made available on GitHub and a dedicated web service.