These studies have shown a link between plant temperatures and canopy conductance, leaf water potential and irrigation regimes, and have shown that thermal imagery can detect even mild stress . However, while highly useful for refined understanding of stress within imaged trees, the methods are not scalable to larger areas due to their crop-specific ground inputs and need for imagery at a fine spatial resolution for tree crowns to be targeted and soil and shade effects minimized. Many of these studies use the Crop Water Stress Index , which relates the difference between air and canopy temperatures to a baseline temperature difference for a non-water stressed crop as a function of the atmospheric pressure deficit. However, Idso et al. found that due to differences in crop physiology, well-watered crops have unique thermal relationships and baseline non-water stressed temperatures. For this reason, when the CWSI is applied to agricultural areas with a diversity of crop species , the model parameters used to identify stress need to be optimized for each crop type by using varying non-water stressed thermal baselines. Further, the CWSI does not account for fractional cover and therefore has a major limitation that it can only be applied to fully vegetated fields or isolated tree crowns. One way to overcome this limitation is to use Moran et al.’s Water Deficit Index , which alters the CWSI so that it can be applied to partially vegetated fields.However, required inputs of net radiation, vapor pressure deficit, wind speed, air temperature, maximum possible plant height, minimum soil roughness, minimum and maximum Soil-Adjusted Vegetation Index, maximum possible leaf area index ,large pots plastic and minimum and maximum possible stomatal resistances make the method difficult to implement over large areas .

On the other hand, satellite-based stress metrics based on TIR data, such as the Vegetation Health Index and the Evaporative Stress Index , use coarse resolution imagery and a wide spatial scope to monitor vegetation health over large areas. The VHI uses a combination of LST and the Normalized Difference Vegetation Index to assess stress with the core assumption that NDVI and temperature are inversely correlated while the ESI calculates stress as an anomaly between actual evapotranspiration and potential ET, as calculated using the Atmosphere-Land Exchange Inverse surface energy balance model. While highly valuable for drought monitoring , the coverage and resolution of these indices inhibit consideration of physiologic differences between plant stomatal behaviors, field-scale variability in soil moisture, or ground covers in a pixel other than soil and vegetation, and are therefore unsuitable for crop stress interpretation at a finer scale. While recent work by Yang et al. has addressed some of these limitations by demonstrating the potential of using multi-sensor fusion to apply ESI at the field scale, this paper offers an alternate, potentially complementary approach for resolving thermal variability inherent in an agricultural landscape. Here, we use Airborne Visible/Infrared Imaging Spectrometer and MODIS/ASTER Airborne Simulator imagery from the Hyperspectral Infrared Imager Airborne Campaign and Geographic Information Science data of crop species to improve understanding of how medium-resolution remote sensing imagery can provide a compromise between complexity and scalability that allows for regional assessments of crop health. We propose an approach for evaluating crop stress that controls for thermal variability within crop species, soil and NPV without the need for site-specific environmental inputs. We analyze thermal patterns in orchards in the Central Valley of California during the 2013-2015 drought as they vary by surface cover component , fractional cover, crop species, and time. We then model the expected unstressed temperature of each pixel, given its fractional composition and species, and compare the expected temperature to the measured temperature as a way of assessing relative stress between orchard species.

The results will highlight the importance of accounting for thermal variability within surface covers when analyzing temperatures and are especially salient in light of the successfully deployed thermal ECOSTRESS instrument and the proposed Surface Biology & Geology mission .A spectral mixture analysis was run on each of the three AVIRIS images to obtain subpixel fractional cover of GV, NPV such as bark or dead leaves, and soil. While multisource energy balance models traditionally characterize soil and GV, this method includes NPV as an additional important cover class that will affect the temperature and turbulent coupling of the land with the atmosphere . The SMA method employed was Multiple Endmember Spectral Mixture Analysis . MESMA uses a linear mixture model to decompose pixels into their fractional components, and provides advantages over other SMA methodologies by allowing the number and types of endmembers to vary on a per pixel basis. In this way, MESMA is able to better account for variability within endmember classes. MESMA was chosen for this study for its ability to model the spectral properties of soils, NPV, and crops using multiple endmembers that can be used to group surface covers into thermal groups. A combined library of forty image endmembers was developed from the three images to capture the diversity of GV, NPV, and soil within the study site and across the dates. Image-selected endmembers were added manually to the spectral library until variation across time and space could be accounted for. This multiyear library was used to unmix each of the three images in order to increase consistency in fractions across the years. The forty endmember library was composed of 22 GV, 8 NPV, and 10 soil endmembers. MESMA unmixed each pixel as a mixture of GV, NPV, soil, and shade using between two and four endmembers to model each pixel. The shade fraction, used to account for differences between the brightness of the endmember and the spectrum being modeled, was restricted to between zero and 80 percent. Each endmember was restricted to between zero and 100 percent, and the residual mean squared error to a maximum of 0.025 for each pixel. In order to obtain physically reasonable fractions, the unmixed results were then shade normalized by dividing each non-shade component, GV, NPV, and soil,square planter pots by the sum total of all non-shade components in that pixel .

The resultant products mapped soil, NPV, and GV at a subpixel level throughout the study scene with an accompanying map showing which endmembers were used to unmix which pixels. Some spatial knowledge of what types of crops are being grown is necessary for accurate thermal estimation of each field. While this information can come from a variety of sources either derived from remote sensing data or as a separate ground data source, knowledge of the spatial distribution of crop species within this study area was obtained from county-level polygons provided by each of the four counties that intersect it: Fresno, Kern, Kings, and Tulare. These crop polygon maps were created by each county in accordance with a statewide pesticide permitting and use reporting program, which requires growers to register their field with information of crop type and the proposed pesticide application. These maps were used to match thermal data from MASTER with the corresponding crop on the ground. For the purposes of this study, we evaluated the thermal signatures of eleven fruit and nut crops: almonds , cherries , grapes , lemons , nectarines , oranges , peaches , pistachios , plums , tangerines , and walnuts . These crops were chosen due to their prevalence in the study area, high economic value, and long lifespan. Given these characteristics, the health of these crops is particularly important to monitor during drought for economic and cultural reasons. The number of fields of each crop species contained in the GIS data layers for each year are listed in Table 3.3. The data layers are not comprehensive of all crop fields in the study area. Therefore, although the 2015 data layer contains roughly 500 more fields than 2013 or 2014, this increase is not necessarily representative of a larger number of perennial fields planted in the study area but is rather more indicative of the completeness of the 2015 layer. We hypothesized that soil and NPV properties that lead to different VSWIR spectra, such as albedo, moisture, or structure, will create differences in temperature that will create distinct thermal classes that group by VSWIR endmember. Temperature distributions were evaluated by VSWIR endmember using a one-way analysis of variance . All pure NPV and soil pixels that had been classified using the same NPV or soil endmember, and therefore showed spectral similarity, were grouped together as a class for analysis. Post-hoc Tukey’s tests were run for both NPV and soil for each date to test for significant thermal differences between classes. To further test the hypothesis that MESMA endmembers model soil pixels into groups of similar soil moisture, we tested two hypotheses: 1) Pixels modeled with MESMA soil endmembers that have lower albedos will have lower temperatures and; 2) Pixels modeled with MESMA soil endmembers that have lower SWIR reflectance will have lower temperatures. First, we plotted the albedo of each of the ten endmembers against the average temperature of the pure soil pixels unmixed with that endmember to assess correlation.

Albedo is a measure of brightness defined as the ratio of reflected radiance from the surface to the incoming solar radiation to the surface . As moisture lowers the albedo and cools the temperature of soils, we expected to find a positive correlation between albedo and temperature. Albedo was calculated using the same approach as in Roberts et al., . Using this method, down welling irradiance is modeled for the specific location, date and time of a dataset using MODTRAN , then multiplied by surface reflectance from AVIRIS and summed across all wavelengths to calculate total reflected irradiance between 350 and 2500 nm . This sum is divided by down welling irradiance integrated across the same wavelength region. Albedo was calculated using the reflectance and modeled irradiance as described in Eq 1, resulting in a unitless measure of brightness for each endmember. Second, we assessed correlation between SWIR reflectance of each of the ten endmembers and the average temperature of the pure soil pixels unmixed with that endmember. Musick and Pelletier found that moist soils have lower than expected reflectance at the longer wavelengths of the SWIR, as examined through the ratio of the Landsat TM 5 and 7 bands. We used a comparable AVIRIS ratio of the 1662 nm band to the 2028 nm band and evaluated whether the pixels modeled by endmembers with smaller ratios had lower average LST values. For GV, we hypothesized that crop species would show multiple distinct thermal signatures that would persist through time. To test this, a one-way ANOVA and post-hoc Tukey’s tests were run to determine if significant differences exist between pure pixels of crop species. Further, to evaluate whether crops cluster by species within GV-LST space, we used the same method as Roberts et al. . A Multivariate Analysis of Variance was run as an initial test for clustering. After finding that significant differences existed by species, to further study one-on-one differences between pairs of crops, Bhattacharyya Distances were computed for each crop pair in GV-LST space. DB is a measure of the ability to separate two classes, calculated with the mean and covariance matrix of each class . Smaller distances indicate greater similarity and larger numbers greater dissimilarity. Distances were computed for all species distribution pairs in each of the three images to better understand which species show the greatest dissimilarity in GV-LST space and to analyze whether these patterns hold over time. This paper proposes an approach for estimating expected, non-stressed LST at the pixel level using internalized calibration of temperatures with species and scene-specific thermal endmembers. By calculating an expected temperature for each pixel, we can assess whether the measured temperature from MASTER is higher or lower than the modeled expected temperature for an unstressed crop, given its species and fractional cover. Differences may indicate stress, in the case of a higher than expected measured LST, and shade effects or high crop ET in the case of a cooler than expected LST.