Downscaling soil moisture to improve watershed and resource management using a UAS and an L-band radiometer
California Water Science Center, USGS
Dr. Michelle Stern, Hydrologist, California Water Science Center, USGS.
Real-time, fine-scale soil moisture data that represents the entire soil profile can critically improve reservoir forecasts, and enhance our understanding of drought, fire, and flood risk in California. Characterization of the spatial, temporal, and vertical distribution of soil moisture can help resource managers recognize interactions of weather and climate with watershed hydrology, forest health, and reservoir management. In these times of uncertain climate patterns with increased occurrences of extreme events and drought, effective management of reservoirs and land in California should rely on timely, fine scale, and accurate estimates of soil moisture throughout the soil profile. Precisely identifying whether antecedent soil moisture conditions are in deficit or surplus and where they occur on the landscape enables the accurate representation of hydrologic status in a watershed that is needed for successful management.
Accurate and spatially resolved estimates of soil moisture are particularly difficult to model in California due to the complex terrain and variable climatic patterns. Soil moisture station data are sparse across the state, especially in mountainous regions where most of the water supply originates. Point measurements of soil moisture data are temporally resolved, yet spatially sparse and do not typically represent the surrounding area, especially at scales from 10 meters to 100’s of meters. In contrast, satellite remotely sensed global soil moisture products are too coarse a resolution for regional to local applications, are not temporally resolved, and do not provide data beyond the top few centimeters of the soil profile. Unoccupied Aircraft System (UAS) data is spatially fine scale over an intermediate sized footprint. This data set in combination with a dense network of soil sensors in varied soil types can be used to interpolate between station locations and importantly, to extrapolate through the soil profile in order to calculate the soil water deficit or surplus needed for reservoir forecasts and the available water for plants and ecosystems.
Land and resource managers across California can benefit from more accurate, higher spatial and temporal resolution soil moisture data to inform reservoir operations, agricultural management, and water resources management. In addition, the development of rapidly updateable fine scale soil moisture data can be input directly into fire, drought, landslide, and flooding risk applications. Current forecasting models for reservoirs and risk index models either ignore soil moisture conditions or rely on estimates that are not fine scale or not resolved vertically through the soil profile.
The disparity between spatial and temporal scales of soil moisture data provides a unique challenge for understanding the driving factors behind soil moisture patterns and temporal trends. Our goal is to improve the understanding of soil moisture processes at multiple spatial scales and depths using remotely sensed fine scale soil moisture data from a UAS. Until recently, radiometer sensors have been too large and heavy to be attached to a drone, and therefore have not been available for research. Recent cutting-edge innovations have reduced the size and weight of such instruments, allowing radiometry data to be collected simultaneously with multispectral and infrared sensors. We hypothesize that using this shallow soil moisture information in concert with a spatially dense soil moisture sensor network that explores the entire soil profile will help to develop the detailed understanding of soil moisture processes at an intermediate scale (< 10s of meters) that can provide key information needed to develop accurate downscaling from large scale satellite data to more relevant scales in California and globally.
The primary objective is to develop fine-scale (10 meter) surface and root zone soil moisture maps for Pepperwood’s reserve. This study has three testable hypotheses:
1) L-band radiometer data can accurately predict 10-meter resolution surface soil moisture using machine learning and deterministic variables such as geology, soil texture, topographic characteristics, land use, vegetation, and climate,
2) surface soil moisture can be used to predict root zone soil moisture at a fine spatiotemporal scale, and
3) the relationships of soil moisture to each independent variable are different at different spatiotemporal scales. We will also study the interactions between plant and soil water content to separate the backscatter signals from UAS L-band flight data.
Multiple data types will be collected for this project, including L-band remotely sensed data, field soil moisture measurements, existing in situ soil moisture and climate time data, as well as gridded physical maps of geology, soil texture and type, vegetation, and topography. The field collection of data will consist of repeat L-band radiometer drone flights over Pepperwood completed by a contractor and paid for by a USGS grant. In addition, 100-200 field measurements of surface soil moisture coincident with each of the flights will improve the spatial distribution of validation sites. Carefully selected collection protocols will ensure a representative set of drone measurements that capture the natural climate variability in the system. Repeat surveys of the same area under different moisture conditions are highly desirable to study the dynamic temporal nature of soil moisture processes in seasonal cycles. Three flight windows will consider specific and useful seasonal soil moisture states. Two-week flight windows will be selected based on historical time series at soil moisture stations to capture the wet season (first two weeks in January) when plants are dormant or leafless and the soils are at field capacity, the dry season (mid-October) when plants are dormant and soils are dry, and in between during the early growing season (first two weeks in May) when soils are drying out. This semiannual schedule will allow a range of soil textures and vegetation types to be represented with varied soil moisture levels and rates of soil dry down. During each flight window, multiple flights may be possible, which will help evaluate the sensitivity of radiometer measurements to daily variation of soil moisture.
In addition to the drone flights that provide optical, thermal, near infrared, and L-band remotely sensed imagery, in situ soil moisture and climate time series data will be collected for the same period and the data will be checked for errors before including it in the analyses. Soil moisture from 22 sensors with 3 to 4 depths each, as well as climate data will be assembled for the period most representative of the drone flights to the nearest hour. Gridded climate data will be downloaded from the PRISM dataset (Parameter elevation Regression on Independent Slopes Model, www.prism.oregonstate.edu) for the daily and monthly analysis. PRISM is available at 4-kilometers at a daily and monthly time step. I will spatially downscale the data from 4-kilometers to 1-kilometer, 100-meters, and 10-meters to test scale dependency on these analyses. Other static gridded data sets will be collected and processed to a consistent projection and spatial resolution for each proposed spatial scale (1-km, 100-m, 10-m). These data sets include soil properties, geology, vegetation, and a digital elevation model (DEM) (Table 1). Data management will include geodatabases in ArcPro or Google Earth Engine to store and analyze the processed time series and gridded data sets. Python and/or R scripts will be used to run quality checks and automate processing and machine learning analyses. Data from the drone surveys, other remote sensing data, time series data, and gridded physical properties will be used to evaluate each testable hypothesis.
We will collect field soil moisture data and existing in-situ soil moisture data, extract remotely sensed and UAS optical, near infrared, thermal, and L-band data for the study area, prepare ancillary data, develop the geostatistical and machine learning models, and calibrate/validate using in-situ data. The ancillary data will include elevation, soil properties, geology, percent canopy cover, and vegetation/land use type. The in-situ soil moisture data will be selected based on UAS data availability. Incorporating additional satellite optical and microwave data will be explored to increase the number of calibration and validation days. To reduce multicollinearity for the non-regression tree methods, a multiple linear analysis and Principal Component Analysis will identify important variables at various scales and reduce redundant and spatially correlated independent variables. The resulting key independent variables, remote sensing data, and in situ data will be combined using machine learning to predict surface soil moisture maps. The machine learning models will be calibrated to 70% of soil moisture station data available for a given day, tested using the remaining 30%, and uncertainty will be assessed for each prediction by calculating root mean squared error and coefficient of determination metrics for out-of-sample validation estimates only.
This project is ongoing until June 2023.