Accuracy, Uncertainty, and Limitations of Satellite-Based Remote Sensing of Evapotranspiration.

Satellite-derived data are ideally suited for producing spatially continuous evapotranspiration (ET) over extensive areas.  Levels of sophistication in producing ET from imagery vary widely, ranging from scaling of vegetation indices derived from reflected short-wave data to full forms of surface energy balance techniques requiring thermal imagery in complement of reflected short-wave data.  As expected, accuracy of ET estimates varies with the level of spectral information utilized, with the completeness of algorithms describing the ET process, and with complexity or type of vegetation and water availability. A tiered-system of categorizing remote sensing of ET approaches is introduced at the end of this abstract to provide structure in assessing, selecting or reviewing the various approaches.

As background, two primary principles govern the process of ET and influence the successful estimation using remote sensing:  1) transformation of liquid water to vapor via ET requires substantial quantities of energy extracted from the environment, either as radiation energy, thermal convection or thermal conduction;  therefore, energy availability and consumption rates govern the rate of ET;  and 2) vegetation that is short on water in its root zone will not transpire in proportion to the amount of vegetation.   The second principle creates a challenge in applying vegetation index based methods with accuracy over areas larger than local areas and without specific calibration using ground-based flux data.  The first principle provides the foundation for using surface energy balance-based methods where ET rates are derived by estimating the rates of energy consumption at the surface and the second principle suggests that energy balance based methods will provide the necessary flexibility to estimate ET under water short conditions in both space and time. 

Satellite imagery is typically applied in a surface energy balance at the surface by estimating energy consumed by the ET process by calculating that energy as a residual of the surface energy equation:

                                                           LE = R_n – G – H                                          (1)

where LE is the latent energy consumed by ET, Rn is net radiation (sum of all incoming and outgoing shortwave and longwave radiation at the surface), G is sensible heat flux conducted into the ground, and H is sensible heat flux convected to the air.  In most approaches, including the common models METRIC, SEBAL, and DISALEXI, Rn is computed using satellite-measured narrow-band reflectance and surface temperature; G is estimated from Rn, surface temperature, and vegetation indices, and in some cases, H (METRIC₂₀₁₂); and H is estimated from surface temperature ranges within the image, surface roughness, and wind speed using buoyancy correction. 

METRIC (Mapping EvapoTranspiration at high Resolution with Internalized Calibration) is one of several satellite-based image-processing tools for calculating ET as a residual of the surface energy balance. METRIC, as with the SEBAL model, uses an innovation originally developed in SEBAL where a near surface temperature gradient, dT, placed in a blended layer just above the surface, is used in the calculation of H.  The dT expression is indexed to radiometric surface temperature (T_s), where surface temperature is used primarily to spatially distribute dT (and H), rather than being used to estimate H directly.  The dT – T_s relationship has eliminated the need for absolute surface temperature calibration, a major stumbling block in operational satellite ET.  METRIC uses the SEBAL technique for estimating dT, thereby eliminating the need for retrieving absolutely accurate aerodynamic surface temperature and the need for retrieving air temperature measurements, both of which are required in the classical method for estimating H for establishing a surface to air temperature gradient, and both of which are prone to substantial biases which at times can be larger than the gradient itself.

Common Errors and biases:

Most polar orbiting satellites orbit about 700 km above the earth’s surface, yet the transport of vapor and sensible heat from land surfaces is strongly impacted by aerodynamic processes including wind speed, turbulence and buoyancy, all of which are essentially invisible to satellites.  In addition, precise quantification of albedo, net radiation and soil heat flux is uncertain and potentially biased.  Although the determination of ET  via surface energy balance is the best approach for mapping ET, especially from water stressed areas or from areas having significant evaporation, E, that is not captured by vegetation indices, the application of surface energy balance can be compromised by the systematic biases associated with data and parameter estimates and with computational procedures.  These biases or uncertainties include:

• atmospheric correction impacts on surface reflectance retrievals
• atmospheric correction impacts on surface temperature retrievals
• albedo calculated from bi-directional reflectance rather than as directional-hemispherical, and broadband surface albedo calculated by integrating narrowband at-surface reflectances
• Incoming longwave radiation (R_L↓) used in net radiation calculation
• calibration of the satellite thermal band and surface emissivity
• air temperature gradient function used in sensible heat flux calculation
• estimation of surface roughness and impacts of near surface characteristics and features on assumptions made for establishing near surface temperature gradients
• stability functions for convective transport
• soil heat flux
• impacts of wind speed fields on convection and near surface temperature gradients

These potential biases plague essentially all surface energy balance computations that utilize satellite imagery as the primary spatial information resource, and even though best efforts are made to estimate each of these parameters as accurately and as unbiased as possible, some biases do occur and can impact the final estimate of ET estimated as the residual of the energy balance.  The relative magnitude of biases and uncertainties increases as the relative ET signal decreases, since error variances tend to remain relatively constant.  To reduce the negative impacts of systematic bias, the METRIC approach, and to some extent, SEBAL, calibrates the surface energy balance at a known, fully vegetated condition to a reference ET,  ETr , to effectively remove systematic biases.  The calibration does this by introducing a bias correction into the calculation of H, that is then removed during calculation of LE.  The end result is that biases inherent to Rn, G, and subcomponents of H are essentially cancelled by the subtraction of a bias-canceling estimate for H.  The result is an ET map having values ranging between near zero and near ETr, for images having a range of bare or nearly bare soil and full vegetation cover.  The ET_r value is calculated using the ASCE standardized Penman-Monteith (PM) equation applied to the tall (alfalfa) reference.  The alfalfa reference tends to approximate near maximum ET from extensive areas of full vegetation cover and thereby serves as an effective ‘anchor’ for the process calibration.  The Penman-Monteith equation is applied using measurements or even estimates of weather data from a preferably agriculturally-located weather station.  In effect, METRIC replaces bias from less-known remote-sensing based estimates by hopefully better known and more limited biases in the PM equation.  Many applications of the SEBAL approach utilize R_n – G for the calibration of the function for H and the dT vs. T_s function, where R_n and G can both have unknown biases stemming from the remotely sensed retrievals.  The ‘dry’ point used as a second calibration condition is common to both METRIC and SEBAL.

24-Hour Evapotranspiration (ET₂₄).  Daily values of ET (ET₂₄) are generally more useful than the instantaneous ET that is derived from the satellite image.  Various strategies are employed by remote-sensing models, including using the evaporative fraction, EF, and fraction of reference ET, ET_rF to extrapolate or interpolate in time.  EF is defined as the ratio of LE to so-called available energy, R_n – G.  ET_rF is defined as the ratio of ET to  ET_r. In the METRIC process, for example, ET₂₄ is estimated by assuming that the instantaneous ET_rF computed at image time is the same as the average ETrF over the 24-hour average for land use types such as irrigated agriculture and wetlands that have external water supplies above ambient precipitation.  Because ET_rF is based on the PM equation and because the PM equation does a good job in capturing effects of advection, use of ET_rF tends to capture advective impacts on ET for those system types.  This is important because advection can double ET rates relative to R_n – G.  In the case of rainfed agriculture and natural vegetation systems, METRIC uses the EF, since EF tends to not increase the estimate for ET₂₄ due to advection.  By definition, advection of energy is zero when the location evaluated has the same water availability characteristics as those of the region.  Research at the University of Idaho and elsewhere has shown ETrF to be relatively constant over 24-hour periods for irrigated crops.

Seasonal Evapotranspiration (ET_seasonal). When ET maps are needed at field-scale resolution, requiring 30 m resolution of Landsat, for example, coupled with the 60 – 120 m thermal imagery of Landsat, the number of images available over the course of a growing season can become relatively few, especially in areas having common cloudiness.  Energy-balance based techniques produce ‘snapshots’ of ET for the available image dates, only.  In the case of Landsat, when one satellite is available, potential images appear only once each 16 days.  With two Landsat satellites, images are available each 8 days.  Under conditions of typical cloud cover, often 30 days or more can transpire between usable images for any given pixel location.  Because it is important to track the impacts of changing vegetation phenology and water availability on ET, the more frequent the imagery, the better the definition of the ET process and amounts of water consumed.  Integration of ET over time is typically done by interpolating ETrF or EF in between available processed image dates using spline or other interpolation technique to create a daily time series of ETrF or EF for each pixel.  The result is very similar to the construction of a continuous ‘crop coefficient’ curve.  The daily ETrF is multiplied by daily ETr computed from daily or hourly weather data to capture the day to day impacts of weather on ET, even under clouded conditions.  When time intervals between images are greater than about 20 days, it is preferable and likely more accurate to use a spline function for interpolation, rather than interpolation, to better approximate the general shape of vegetation development curves.  Generally one satellite image per month is sufficient to construct an accurate ETrF curve for purposes of estimating seasonal ET.  Allen et al., (2007) described the tendency of random error in ETrF on individual overpass dates to partially self-cancel when extended over multiple overpass dates, similar to the reduction in the standard deviation of a mean value with increasing degrees of freedom.  This is particularly true when ETrF is impacted by evaporation from random irrigation events or cutting of forage crops.

Categories of Remote Sensing Approaches for Evapotranspiration

The following four tiers are being considered by the American Society of Civil Engineers (ASCE) Task Committee on Remote Sensing of Evapotranspiration to help users and describers of remote sensing systems and approaches for ET better describe the content of approaches and expected accuracies and to determine the appropriate types of applications for a particular method or method type.  These tiers are described here for purposes of summarizing the general families of approaches, expected accuracies and recommended types of applications that are congruent with the expected accuracy.

 

Tier 1. (lowest tier) - Cursory exploration of spatial distribution of water consumption by vegetation

-- using general ET vs. vegetation indices or general crop coefficients (K_c) vs. vegetation indices

-- having accuracy of about +/- 30% uncertainty (1 S.D.).       

-- Useful for:             

Identifying irrigated vs. nonirrigated areas,

showing greeness and wetness of riparian systems, etc.

for atlas-level type of work.

Tier 2. - Image-based products where some human overview is exercised, and where short-wave or thermally-based algorithms are utilized that have a limited physical basis.  

Methods may include

scaling from vegetation indices and

scaling of reference ET by surface temperature. 

Applications may include:

a. annual reporting having moderate accuracy where rapid computation is required and where approximate accuracy or relative use is sufficient

b. national or global surveys on water consumption or production of atlases

Tier 2 applications are encouraged to validate against tier 3 and 4 models and ground-based measurements.

Concerns include approaches and accuracy of time integration

Maximum time spans between interpolation points

Accuracy over diverse vegetation types

Tier 3. – Creating spatial maps of ET showing variation over time and space with sufficient accuracy at the monthly scale for use in parameterizing or driving:

a. hydrologic models including ground-water recharge and depletion estimation

b. surface water accounting on streams and streamflow depletion

c. general basin-wide water balances, and

d. developing crop coefficients. 

Tier 3 requires well-developed procedures for relatively accurate time integration between sparse images or successful and tested fusion of multiple spatial and temporal resolutions of imagery. 

Tier 3 also requires algorithms that apply well (accurately) to broad ranges of vegetation and land uses.

Tier 4. (top tier) – for supporting:

a. management of water rights,

b. water transfers,

c. litigation,

d. streamflow depletions for mitigation and multi-state agreements. 

Tier 4 applications employ

a large amount of human oversight and professional, expert review

substantial (or sufficient) physics in algorithms to quantify:

important impacts of vegetation or surface characteristics on the surface energy balance

transformation of energy to latent heat as impacted by:

surface albedo                                         

ground heat flux                                     

boundary layer bouyancy  (instability)

surface roughness

sensible heat flux

radiative emissivity

Methods should employ robust means for time-integration between image 'snapshots' that include either

adjustment for background evaporation from exposed soil due to precipitation wetting for the integrated period as opposed to that occurring at the time of the snapshot or

fusion with more frequent nonthermal or larger resolution imagery

Tier 4 applications must be able to pass muster in courts of law and among the common applications communities. 

A targeted accuracy for Tier 4 might be +/- 10% uncertainty (1 S.D.).

As with most ‘models’ and data handling, operation by trained experts having good physics background is a distinct advantage in achieving high accuracy from remote sensing of ET.  Having human review and oversight of the processing at the individual image level is often required to produce ET estimates having sufficient accuracy and credibility for the intended use.  Most remote sensing techniques are highly reliant on the operator’s ability to calibrate or tune the applications and to recognize situations when the process is not working well. Current work among various research groups includes the use of statistical and other relationships to automate the calibration process.  The application of energy balance to a wide mixture of agricultural crops and other vegetation is complex enough that there are still some areas of considerable empiricism and therefore potential for local refinement.