Albedo, fAPAR, Land Surface Temperature, Light Use Efficiency (LUE), NDVI , Precipitation, Solar radiation, Statics and Weather data

Land Surface Temperature

The Land Surface Temperature (LST) is the radiative skin temperature of the land surface, as measured in the direction of the remote sensor. It is an indicator of root zone soil moisture content, as land surface temperature will be lower when sufficient water is available for evaporation and transpiration.

WaPOR application

Land surface temperature is used in WaPOR to calculate the relative root zone soil moisture content.

Methodology

Because most sensors are not recording LST available for the spatial resolutions in the WaPOR portal (100m and 30m), the LST needs to be resampled to those resolutions. For the previous versions the LST has been resampled using a bilinear resampling strategy. This works for homogeneous areas, but for heterogeneous areas it may have unwanted side effects. Especially for irrigated areas in arid conditions, the difference between LST of the irrigated areas (cool) and the surrounding land (warm) is large. Bilinear resampling may lead to an overestimation of the LST of the irrigated areas, and a subsequent underestimation of the root zone soil moisture content, because higher temperatures at similar vegetation cover indicates a lower root zone soil moisture content.

Data Mining Sharpener

Within the current literature a large body of work exists on the use of high-resolution optical data to sharpen thermal infrared imagery (Mao et al., 2021). The methodology chosen here has been introduced by Gao et al. (2012) and is described as a data mining sharpener (DMS) technique. The DMS approach builds regression trees between TIR band brightness temperatures and shortwave / optical spectral reflectances with additional information coming from elevation maps. The methodology comes with the caveat that sharpening cannot replace actual thermal band imagery at high resolutions. An actual measurement will always be better than a sharpened result. Therefore it is recommended to only use this methodology if there are no LST measurements available for the resolution needed.

An implementation of this algorithm in Python has been made available on GitHub and is described by Guzinski et al. (2019). First, the high resolution features are resampled to the same resolution as the original LST data. Any high resolution dataset can be taken as input for the high resolution features. For WaPOR high resolution feastures are derived from the raw reflectance’s in the optical and shortwave infrared, but also indices derived from these reflectance (vegetation indices, drought indices etcetera) and information derived from high resolution elevation maps. For every resampled pixel the coefficient of variance is calculated.

Where cv is the coefficient of variance and n is the number of features used for the regression and σi is the standard deviation of the feature and μi is the mean of the feature.

A threshold is used to determine the most homogeneous pixels that can be used for training the regression model. The paper of Gao et al (2012) proposes to use a threshold of 0.20 (20%). Regression trees are then modelled using the training data. This is done for the whole scene (global regression) as well as on a moving window within the scene (local regression).

The results of the global and local regression are combined based on residuals. The residuals are based on a comparison between the original coarse resolution LST and the modelled LST at the same resolution by using the regression. This analysis provides weighing coefficient which are used in the final regression result.

Where p is the result of the overall regression and pi is the result of the local/global regression and wi,c is the weights attributed to the local and global regression. The use of the global and local regression makes it suitable for processing by choosing custom global and local windows. This opens up the possibility of processing larger areas than just single images. That makes it suitable for use in the WaPOR context where the area under consideration is usually larger than a single scene.

The weights are calculated as:

The prediction (local or global) with the smaller residual r at a given coarse resolution pixel represents a better prediction and thus is weighted higher using the reversed squared residual.

Processing approach

For WaPOR version 3 the high resolution featues used to predict LST were created from (1) spectral bands, (2) indices based on the spectral bands, and (3) elevation features:

• NDVI (Normalized difference vegetation index)
• PSRI (Plant senescence reflectance index)
• LWCI (Leaf water content index)
• MSI (Moisture stress index)
• GVMI (Global Vegetation Moisture Index)
• NDWI (Normalized Difference Water Index)
• VSDI (Visible and Shortwave infrared Drought Index)
• SWCI (Surface Water Capacity Index)
• SIMI (Shortwave Infrared Moisture Index)
• NDMI (Normalized Difference Moisture Index)
• NMDI (Normalized Multi-band Drought Index)
• SAVI (Soil Adjusted Vegetation Index)
• NDBI (Normal Difference Built-up Index)
• BSI (Bare Soil Index)
• EVI (Enhanced Vegetation Index)
• FVC (Fraction Vegetation Cover)
• BI (Bare soil Index)
• MSAVI (Modified Soil Adjusted Vegetation Index)
• NDDI (Normalized Difference Drought Index)
• MNDWI (Modified Normalized Difference Water Index)
• OSAVI (Optimal Soil Adjusted Vegetation Index)
• IBI (Index-based built-up Index)
• IVI (Index-based vegetation Index)
• UI (Urban Index)
• NDSI (Normalised Difference Snow Index)
• VARI_RED_EDGE (Visible Atmospherically Resistant Index Red Edge)
• VARI_GREEN (Visible Atmospherically Resistant Index Green)
• MNDWI (Modified Normalized Difference Water Index)

After evaluating the features with some example datasets, a limited set of features was selected for the thermal sharpening methodology. Features which were correlated highly (correlation coefficient > 0.8) with other features were removed to avoid redundant information. This resulted in the selection of 2 out of the 11 evaluated spectral bands, the selection of all elevation features (elevation, aspect, slope and the cosine solar zenith angle), and the selection of 5 out of the 28 evaluated Sentinel-2 indices:

The Land Surface Temperature (LST) is obtained from VIIRS Brightness Temperature layer (band I5) that has a spatial resolution of 375m. For level 1 brightness temperature is resampled to 300m using bilinear resampling. For level 2 and level 3 it is resampled to respectively 100m and 20m using a thermal sharpening approach that uses Sentinel-2 features and digital elevation model (DEM) features to improve the spatial resolution. After resampling, the brightness temperature is converted into land surface temperature (see VIIRS for a description of the approach).

For some L3 areas that require a longer time series, e.g. the L3 areas in Libya and Yemen, LST data is obtained from Landsat which has a longer time series but fewer (less frequent) observations than VIIRS but a higher spatial resolution (100m). LST is either sourced from Landsat or VIIRS to avoid sensor related fluctuations in the soil moisture dataset.

Challenges

The assumed statistical relationship between LST and spectral indices is more complex in reality: for example water stress may be reflected in the thermal infrared weeks before it is visible in the spectral indices. Mitigating factor is that some of the spectral features also include shortwave infrared bands which are sensible to moisture content. Another issue lies in the topography of the terrain, which could lead to shadow effects, which are not represented well by optical data. For that purpose features based on elevation information are included.

Functions and flowcharts

The functions to derive the indices are provided in the ETLook code base in the indices module.

The flowchart of the whole process shows the inputs of the process (high resolution DEM, high resolution Sentinel-2 data, coarse resolution LST).
- Features are calculated based on the high resolution DEM and Sentinel-2 data. These features are being resampled to the same resolution of the LST input.
- Subsequently homogeneous pixels are selected using the coefficient of variation approach. The pixels selected will be coupled the corresponding coarse resolution LST pixels and the regression will be carried out. The regression is done on two levels (global and local).
- Based on the regression results for global and local a coarse resolution LST is calculated.
- Using the difference between the original LST and both modelled coarse resolution LST the regression coefficient for both the local and global approach are being weighted. These weights are then applied to the high resolution features derived in the first step and will lead to a high resolution LST.

Figure 1 Flowchart of the thermal sharpening procedure