Better Measures of Land Surface Temperatures Improve Monitoring for Drought, Stressed Crops, and Crop Productivity
This post written by Leiqiu Hu, University of Alabama in Huntsville, Department of Atmospheric and Earth Science
Weather and climate shocks like droughts, floods, and heatwaves can imperil climate-sensitive agricultural systems in food-insecure regions, threatening the livelihoods and nutritional status of vulnerable populations in these areas. Satellite remote sensing offers an inexpensive, timely solution to monitor conditions on the Earth’s surface and, therefore, has been increasingly used to generate data in decision making by private and public actors. One major satellite product widely used for drought and vegetation stress monitoring in agricultural systems is land surface temperature (LST). Spatial and temporal variation in LST are critical for identifying the governing physical process of land-atmosphere interaction. In particular, LST variability has been demonstrated to accurately predict plants’ evapotranspiration and surface moisture that are key determinants of vegetative stress associated with low crop productivity.Figure 1 A: Examples of Diurnal Cycle Model fitting for two cases: one in a clear-sky day and another with a brief cloud presence.
Even though the availability of satellite data has soared in recent decades, some major gaps remain for accurately representing vegetation stress and drought severity spatially and temporally. For example, a trade-off exists between spatial and temporal resolution of LST products, which is mainly determined by satellite mission platforms. Currently, available LST products have either low spatial resolution (products derived from geostationary satellites, with hourly observations, but at 3 kilometer or coarser spatial resolution) or low revisit frequency (products generated by sensors on near-polar orbiting satellites that provide observations at most once or twice a day but at 1 km spatial resolution). Thermal remote sensing for estimating LST is also challenged by missing data due to cloud cover obstruction. To help fill this gap and reduce the stark space-time tradeoff in existing LST series, we have developed a novel cloud cover-corrected and diurnal-based LST product that overcomes these limits, improving the accuracy of vegetative stress detection.
Specifically, we use a physics-based diurnal temperature cycle (DTC) model to synthesize multi-source information with a realistic and timely representation over Feed the Future countries. The DTC model allows us to numerically represent a complete diurnal cycle of LST for clear or pseudo-clear (i.e., cloudy) sky conditions based on a few observed values at any time of a day. Such information allows us to generate a true time-consistent daily mean or extrema (i.e., minimum or maximum) LST for seasonal and interannual trend analysis.Figure 1 B: The comparison of DTC modeled monthly mean LST and the simple composite without filling the clouds from geostationary satellites over Ghana and Guatemala shows the DTC can correct the bias due to uneven samples (due to clouds). The dark red color suggests all observations (24) in a diurnal cycle are available, and the blue color indicates fewer observations in a diurnal cycle.
Two freely available LST products are used: hourly-5km LST from geostationary satellites (Copernicus LST products, e.g., observations from NOAA’s GOES Imager and European MSG SEVERI) and day/night-1km NASA MODIS LST products. The geostationary hourly product offers rich temporal information to build and train the DTC models. Furthermore, we have tested the capability of DTC for filling temporal missing data due to cloud cover (or other reasons) and found that DTC-fitted results are almost the same as clear-sky observations. Meanwhile, the DTC model corrects the high or low bias of daily mean LST estimated using the traditional approach in partial cloudy days due to non-random sample selection (e.g., a brief presence of clouds in a diurnal cycle in Figure 1A). Next, we adapt the geostationary-observations-trained DTC model to fit MODIS observations that have higher spatial resolution but lower temporal frequency (at most four observations per day), resulting in a finer spatial resolution (1 km) product for LST daily mean and extrema from 2003-2018 (See Figure 1B). Such refined estimates are true 24-hr average rather than a simple average of uneven sampled day and night observations, which can more accurately represent surface climatic conditions than do existing instantaneous LST estimates directly observed during satellite overpasses.
We validated this approach using ground observations of more than 100 global flux sites from the FLUXNET. We found the correlation coefficients between ground observations and DTC-based LST (Figure 2) all exceed 0.95 over croplands. The initial assessment on spatial representation by comparing the new LST products with the simple composite maps have suggested a much-improved spatial continuity of our new products during the cloudy seasons and over the cloudy regions. We are further assessing its spatial accuracy with independent areal model-based products which contain the daily mean radiative temperature, e.g., ERA-Interim.Figure 2. Examples of the DTC-fitted LST (°C) for the monthly mean, maximum, and minimum over Ethiopia in January of 2011 from geostationary observations. Note that LST is different from air temperature, and its diurnal range can be much larger than the air temperature.
This new LST product captures diurnal features of LST at higher spatial and temporal frequency––and with less bias than existing products––thereby enhancing analysts’ ability to detect vegetation anomalies induced by droughts and heatwaves regionally and globally, especially as growing conditions are particularly sensitive to daily maximum and minimum temperatures more accurately reflected by this improved, physics-based method. Our more reliable, more representative product can also be used by the climate modeling communities for temperature trend comparisons. We anticipate using this new LST product, as well as the newly developed SIF product from our project, to better inform decisionmakers' understanding of crop productivity, as well as economic and nutritional outcomes in Feed the Future countries.
This work was produced as part of a USAID-funded project: Harnessing Big Data and Machine Learning to Feed the Future (PI: Chris Barrett, Cornell University)