While any fractional cover estimate will be subject to error, we believe MESMA is an improvement over these other models for reasons that will be discussed in Section 4.2. Finally, we do not believe that the errors in temperature estimation that result from errors in fractional estimation will preferentially affect any specific crop species. Therefore, while there will be pixel-level error due to fractional cover estimations, these errors should balance out and allow for relative stress comparisons when analysis of crop temperatures is aggregated to the field and/or species levels. Fifth, differences in flight timing will contribute to noise in the interpretation of results. Crop stress fluctuates with time of day, and plant transpiration has been found to plateau midday and then decrease in the afternoon as soil water content and soil water potential decrease . Therefore, temperature residuals are not only a function of the overall health of the crop during the year and season the imagery was collected but also the water availability on the day and time it was captured. While flight timing should be considered when interpreting findings, we do not believe that the effect is significant enough to overwhelm the yearly signal of increasing stress that is likely due to drought. Temperature residuals were shown to increase from 2013 to 2015. If this trend were due to the timing that the data were collected, we would expect the timing to have a similar trend as the residuals. However, the flight in 2014 was flown latest in the day yet did not have the highest residuals. Additionally, 2013 and 2015, which were flown at similar times of day, had the greatest differences in average LST residual, which is the opposite result as would be expected if flight timing were the main driver of residuals. Therefore, flight timing is an important factor to consider when interpreting LST patterns, but its effect on this study is assumed to be small. If designing a study to compare water stress between months or years,blueberries in containers consistency in timing of acquisition would enhance interpretability of results. Sixth, this study assumed that environmental variables such as air temperature, radiation, and wind did not deviate significantly across the study scene.
In our study area, which has little variability in elevation and has similar vegetation types throughout, we assume the climate is relatively stable. However, some variability will exist across space that will cause error. If this model were applied to a study area that encompassed multiple climate zones, thermal endmembers would need to be calculated separately for each zone to account for known differences in environmental conditions. For this reason, we suggest that this method be implemented over relatively homogeneous areas, with respect to topography and climate. Finally, this model requires a priori information about the landscape and informed knowledge to select appropriate thermal groups. In this study, a crop map that was compiled on-the-ground at the county level was used to inform thermal classes. In lieu of a similar map, crop information could also be gathered from a classification of VSWIR imagery or from a national crop map such as the Cropland Data Layer . With that information, crops could be separated into groups that are similar physiologically and bio-physically such that they are expected to have similar thermal behavior. One of the limitations of the model is its sensitivity to choice of thermal groups. While we expect this method to be transferable to other regions, types of landscapes, and thermal groupings, further work would be necessary to test this hypothesis. This study found a significant inverse relationship between green vegetation and LST. Other studies have observed similar trends between temperature and vegetation . However, unlike other models that use vegetation indices to account for this relationship, such as the VHI or WDI, our model uses a mixing model. The mixing model is an improvement upon the vegetation indices as it leads to a more accurate estimate of fractional cover in a diverse landscape such as our study area. NDVI is sufficient for estimating fractional cover in simple landscapes with little spectral variability, but a mixture model is better suited for estimating fractions in spectrally and spatially diverse landscapes . NDVI is also sensitive to variability in soil background reflectance, which is accounted for with a mixing model. MESMA, in particular, has the added benefit of using multiple endmembers per surface cover component, which allows these components to be grouped by spectral similarity. We found that there is significant thermal variability within the broad surface covers of GV, soil, and NPV.
These findings imply that a detailed knowledge of the landscape, beyond basic surface cover fractions, should be considered for interpretation of LST in agricultural areas. We explored thermal patterns in agricultural orchards and found that core thermal differences exist between crop groups of citrus, perennial fruits, and nuts that may not be attributable to stress. These findings indicate that physiological differences between crops result in different thermal behaviors that will impact interpretability of stress. Results are also in agreement with the observation of Roberts et al. , who found that LST and GV cluster by dominant plant species in LST/GV space. Moreover, we found that there is thermal variability in soil that is correlated to soil albedo and in NPV that is hypothesized to vary by structure. This result is consistent with studies of soil moisture, which have observed that moisture will lower the albedo and temperature of a dry soil . The method detailed in this paper acknowledges the variability within soil, NPV, and GV and uses a crop map and MESMA endmembers to account for some of this thermal variability. Current remote sensing methods that estimate agricultural stress either require field specific inputs that limit the scale of applicability or are wide-reaching but too simplistic in their assumptions such that all GV, soil, and NPV are treated similarly regardless of structural complexity, albedo, or functional group. Field-level models such as the WDI and CWSI account for differences in species by requiring crop-specific data, but these intensive inputs limit broad spatial analysis. Other models that have been developed for analysis across large areas, such as satellite-based ESI and VHI, do not account for the degree of thermal variability that was found in this study to be present in an agricultural landscape. While ESI and VHI do use NDVI and, in the case of ESI, also LAI to account for thermal differences within green vegetation, these parameters will not account for thermal variability from soil type or moisture, NPV, or even species-level thermal variability. An advantage of this method is that it segments GV, soil and NPV into groups that should have similar thermal behaviors while requiring no site-specific inputs other than a crop map. Its internalized calibration makes this method scalable across time and space.
In lieu of a crop map, MESMA endmember groups, such as were used for NPV and soil, may be a suitable substitute for grouping GV into thermal classes. However, further study would be necessary to test this hypothesis. Monitoring crop stress is important in anticipating the future of the agricultural landscape and can provide insight into plant water status and plant stress that could help to identify unhealthy crops to mitigate impacts that could lead to decreased yields and economic losses. We found that thermal imagery collected at only one date per year over three years of drought was able to identify the species that were facing the highest degrees of stress, in agreement with county-level yield data. Thermal remote sensing has been observed to correlate with fruit quality in orchards with open canopies and has been used as an indicator of regional agricultural drought as measured by crop yields . Subsequent analysis of our model would be necessary to determine if the results are robust at the field level,planting blueberries in pots but even the regional correlation with measured crop yields has important implications for farmers, policymakers, and scientists analyzing food and water resources. Moreover, temperature patterns were correlated with expected ET rates for crops in California in a dry year, further bolstering the hypothesis that LST patterns can be used to infer information about crop water use and stress. We hypothesize that deviations in the expected linear relationship between ET and LST are the result of irrigation management decisions such as irrigation method, timing, or applied amount that will affect the health and productivity of the crops. ET rates were calculated under the assumption that crops were watered with surface irrigation systems, suggesting that expected ET rates would increase by 3-6% if drip or micro irrigation were applied . As the frequency of drip irrigation for orchards has increased dramatically in the past couple decades, this change in method may be a factor in deviations from the ET/LST relationship. We also hypothesize that deviations in the ET/LST relationship are a factor of drought management techniques such as reduced watering. In line with this assumption, walnuts consistently had the lowest temperatures and residuals of the three nut crops in the study. This finding is consistent with suggested drought irrigation management techniques that recommend against deficit irrigation for walnuts as they are highly susceptible to damage if faced with water stress and are not as tolerant to these practices as either almonds or pistachios . The lower walnut temperature may, therefore, be the result of continued, consistent irrigation in comparison to almonds and pistachios, which receive deficit irrigation.
The temperature residuals capture the difference between expected and measured temperature and therefore act to as an important source of information about ET rates, the irrigation management practice, and stress. These thermal analyses are important for prioritizing water resources, especially in times of drought when water is limited. Additionally, maps of thermal stress could be valuable to assess the representativeness of in situ measurements of carbon dioxide, water vapor, sensible heat, or other fluxes over a heterogeneous landscape . This method of quantifying stress could also be complementary to surface energy balance models such as Disaggregated ALEXI and Mapping Evapotranspiration with Internalized Calibration . As our approach provides a relative measure of stress and DisALEXI and METRIC estimate actual ET, the surface energy balance models could be used to test the sensitivity of our approach to changes in ET. Alternately, the segmentation of soil and NPV into thermal classes in our method may be of use for refining the evaporation component of DisALEXI or METRIC for increased accuracy in crop ET estimation. Our approach has the added benefit of requiring only VSWIR imagery, thermal imagery, and a crop map that provides a level of detail of, at minimum, plant functional groups, and these minimal inputs allow for ease of implementation. Given these inputs, the approach suggested in this paper is probably best suited for agricultural applications at a spatial scale where environmental variables do not vary highly, such as the study scene in this paper. Within a relatively homogenous area, this method could be applied routinely using consistent thermal groups in order to identify which fields are most stressed and/or to gain information about irrigation management practices, particularly during drought as studied in this paper. In order to improve the ability of this method to capture within crop change in drought stress through time, further analysis would need to be conducted with a dataset that collects at approximately the same time for every capture. With our dataset that contained large differences in time of capture, these timing differences made such analyses beyond the scope of this paper. This study showed that thermal signatures of agricultural crops are correlated with crop species and fractional cover. Therefore, LST data on its own without information about surface structure and composition is challenging to interpret in the context of crop stress. The SBG mission would provide spatially and temporally paired thermal and VSWIR imagery globally which would allow for detailed analysis of LST patterns that take into account fractional cover and surface type. The ability of SBG to monitor crop stress would be enhanced by a crop classification that could group crop fields into relevant thermal classes, removing the need for accurate GIS data layers.