*Article* **Monitoring 10-m LST from the Combination MODIS**/**Sentinel-2, Validation in a High Contrast Semi-Arid Agroecosystem**

**Juan M. Sánchez 1,\* , Joan M. Galve <sup>1</sup> , José González-Piqueras <sup>1</sup> , Ramón López-Urrea <sup>2</sup> , Raquel Niclòs <sup>3</sup> and Alfonso Calera <sup>1</sup>**


Received: 25 March 2020; Accepted: 1 May 2020; Published: 4 May 2020

**Abstract:** Downscaling techniques offer a solution to the lack of high-resolution satellite Thermal InfraRed (TIR) data and can bridge the gap until operational TIR missions accomplishing spatio-temporal requirements are available. These techniques are generally based on the Visible Near InfraRed (VNIR)-TIR variable relations at a coarse spatial resolution, and the assumption that the relationship between spectral bands is independent of the spatial resolution. In this work, we adopted a previous downscaling method and introduced some adjustments to the original formulation to improve the model performance. Maps of Land Surface Temperature (LST) with 10-m spatial resolution were obtained as output from the combination of MODIS/Sentinel-2 images. An experiment was conducted in an agricultural area located in the Barrax test site, Spain (39◦03′35" N, 2 ◦06′ W), for the summer of 2018. Ground measurements of LST transects collocated with the MODIS overpasses were used for a robust local validation of the downscaling approach. Data from 6 different dates were available, covering a variety of croplands and surface conditions, with LST values ranging 300–325 K. Differences within ±4.0 K were observed between measured and modeled temperatures, with an average estimation error of ±2.2 K and a systematic deviation of 0.2 K for the full ground dataset. A further cross-validation of the disaggregated 10-m LST products was conducted using an additional set of Landsat-7/ETM+ images. A similar uncertainty of ±2.0 K was obtained as an average. These results are encouraging for the adaptation of this methodology to the tandem Sentinel-3/Sentinel-2, and are promising since the 10-m pixel size, together with the 3–5 days revisit frequency of Sentinel-2 satellites can fulfill the LST input requirements of the surface energy balance methods for a variety of hydrological, climatological or agricultural applications. However, certain limitations to capture the variability of extreme LST, or in recently sprinkler irrigated fields, claim the necessity to explore the implementation of soil moisture or vegetation indices sensitive to soil water content as inputs in the downscaling approach. The ground LST dataset introduced in this paper will be of great value for further refinements and assessments.

**Keywords:** Downscaling; thermal infrared; land surface temperature; disaggregation; Copernicus

#### **1. Introduction**

Time series of fine spatial and temporal resolution Thermal Infrared Images (TIR) are essential in a variety of agricultural applications, water resources management or irrigation scheduling, based on surface energy balance modeling [1–4]. However, spatio-temporal resolution of the operational TIR satellite sensors results are insufficient for some applications and services, including agriculture. The importance of high-resolution TIR images is being claimed [5–9]. The limitation in the TIR domain remains, since the revisit time for high spatial resolution TIR sensors is typically poor, while the spatial resolution for those with a high revisit frequency is too coarse. In practice, the spatial resolution requirements of satellite-derived surface temperature for agricultural applications are <50 m to face certain physical limitations related to the sensor's point spread function in TIR observations [2,10,11]. As for the temporal resolution, daily TIR observations are desired, although this requirement could be relaxed to 3 days as a minimum threshold [7,11].

The Copernicus conceptual mission LSTM [12] could complement other planned high-resolution TIR missions (e.g., the JPL-NASA Landsat 9-10 or the Indian-French TRISHNA mission [13]) and fulfill the spatio-temporal requirements stated above. In the meantime, downscaling methods are contributing to filling this gap by downscaling the TIR coarse resolution to finer resolutions [3,14–17]. Several techniques have been proposed in the literature to enhance the spatial resolution of the TIR domain over vegetated areas by linking TIR and reflectance information in the Visible Near Infrared (VNIR) [18–21]. These techniques are generally based on the assumption that there exists a relation between the vegetation cover and the LST. According to these approaches, a relation between the TIR and VNIR bands is first obtained at coarse spatial resolution, and then applied at the finer resolution of the VNIR bands, assuming that this relation is scale invariant.

The Normalized Difference Vegetation Index (NDVI) or the Fractional Vegetation Cover (FVC) are the most commonly used inputs in sharpening techniques, although some studies have recently explored the possibility of implementing other combinations of reflectance values that can better characterize the surface response [17,22,23]. There are also some efforts attempting to integrate soil moisture delineated vegetation indices [24], and even radar-derived soil moisture [25] in the formulations of the LST downscaling.

For years, the Moderate Resolution Imaging Spectroradiometer (MODIS) or the Advanced Along-Track Scanning Radiometer (AATSR) were combined with Landsat or Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) imagery to downscale LST from 1000 m × 1000 m to ~1 ha (10,000 m<sup>2</sup> ) scale. Higher resolution VNIR sensors, such as Formosat or the Satellite pour l´Observation de la Terre (SPOT), have been also used to improve the disaggregated LST pixel size [10,26].

The synergistic use of Copernicus Sentinel-2 (S2) and Sentinel-3 (S3) imagery could offer the desired solution of high spatial and temporal resolution [8,26,27]. Although no TIR information is provided, the Sentinel-2A and -2B tandem offers a 3–5-day repeat cycle, and a 10–20 m spatial resolution in the VNIR bands. Revisit time for S3 reduces to 1–2 days, with a spatial resolution of 1000 m for their thermal channels. The relationship between TIR and VNIR bands could be extracted from S3 and then applied to S2 VNIR data. Sobrino et al. [27] explored the conceptual combination of the MultiSpectral Instrument (MSI), on board the Sentinel-2, and the Ocean and Land Color Instrument/Sea and Land Surface Temperature Radiometer (OLCI/SLSTR), on board the Sentinel-3, to show an improvement in LST products derived from AATSR at that time, before the Sentinel-3 data were available. High spatial resolution data from S2 was used to improve the characterization of the sub-pixel heterogeneity through a better parameterization of surface emissivity, although no downscaling was applied by these authors. A machine learning algorithm was proposed by Guzinski and Nieto [8] to sharpen low-resolution TIR observations from S3 using high-resolution VNIR S2 imagery. Huryna et al. [28] applied the methodology introduced by Agam et al. [19] to the combination of S3–S2 imagery. However, the methodology was tested using Terra/MODIS or Landsat observations in both works, due to the lack of high-resolution TIR data to use for cross-validation.

Despite the extraordinary growth of downscaling studies in the past decade, the assessment of the thermal sharpening techniques has been traditionally conducted by cross-validation with derived LST products at original Landsat or ASTER TIR spatial resolutions, 60–100 m and 90 m, respectively [1,8,17,21,28]. Comprehensive ground validations of disaggregated LST are quite scarce, due to the lack of robust datasets covering high contrast heterogeneous areas.

This paper continues the work initiated by Bisquert et al. [1]. These authors tested the application of different downscaling techniques in an experimental site in Barrax (Albacete, Spain) from the combination MODIS-Landsat to provide LST at fine spatial and temporal resolutions, to fulfill the requirements in the estimation of surface energy fluxes and evapotranspiration in the agricultural areas of semi-arid regions, where small land holdings dominate. Bisquert et al. [1] analyzed both classical methods based on the VNIR-LST regression, as well as more advanced approaches based on Neural Networks (NN) or Data Mining (DM). Linear, quadratic and exponential relationships, proposed in the literature, were tested and results were compared to those obtained by applying NN and regression trees in a DM approach using reflectance values from all the spectral bands available. These authors observed that NN and DM, as well as the nonlinear regression tested, have the risk of overfitting, being very sensitive to noise in the samples. They concluded that the simpler NDVI-LST linear regression led to the better results in this case. Bisquert et al. [1] explored the technique results for the different land covers in the Barrax area, and found the largest uncertainties for irrigated croplands, especially in summer when cover heterogeneity and irrigation effects are stressed. As a follow-up, Bisquert et al. [26] extracted disaggregated LST maps at a 10-m spatial resolution for the first time, using high-resolution SPOT-5 images in the framework of the Spot-5 Take 5 project. Results shown in [26] were encouraging for the further application of the model to operational S2 images.

In this context, the objective of this paper is to revise and adapt the downscaling technique to the combination MODIS-S2 to derive operational LST maps with a spatial resolution of 10 m. Some adjustments to the original formulation of the approach were introduced to reduce the model uncertainty by adding an additional image-based parametrization of the residual as a function of the VNIR response. Ground LST data from an experimental campaign carried out in the summer of 2018 were used for the model evaluation. The variety of croplands and the contrast in the surface conditions during the experiment in the selected area allowed a comprehensive analysis of the performance of the downscaling technique, not achieved before. Strengths and limitations of the models were discussed, and also some guidelines for the optimal use of this technique with Sentinel-3 and Sentinel-2 imagery are given.

This paper is structured as follows. Section 2 describes the study site, the field measurements and the satellite imagery used, as well as the downscaling methodology. Results of the ground validation and distributed assessment are shown in Section 3. Interpretation of the results and comparison with previous studies comprise Section 4. Finally, Section 5 summarizes the main conclusions of this work.

#### **2. Materials and Methods**

### *2.1. Study Site and Measurements*

This work was conducted in the semi-arid area of Barrax, southeast Spain (39◦03′ N, 2◦06′ W). This is a very flat area with an average altitude of 700 m a.s.l, close to Albacete (Figure 1), traditionally used by ESA (European Space Agency) as a test site in different international campaigns [29–31]. Irrigated and rainfed crops combine in this agroecosystem, with field size ranging from small terrains below 1 ha to large pivots over 50 ha (Figure 1). This large variety makes Barrax a perfect target to assess the performance of a downscaling technique and explore its strengths and weaknesses.

Ground measurements of LST (LSTg) were registered in "Las Tiesas" experimental farm during the summer of 2018, concurrent with EOS-Terra/MODIS overpasses, and covering a total of 10 different crops in 13 independent fields (Figure 1). The temperatures were measured using four hand-held infrared radiometers (IRTs) Apogee MI-210. These radiometers have a broad thermal band (8–14 µm), with a 22◦ field of view and an accuracy of ±0.2 K, according to the manufacturer (Apogee Instruments, Inc.). In fact, the similar Apogee SI-121 radiometers (same radiometer, but with a field of view of 18◦ and without datalogger) were calibrated against a National Institute of Standards and Technology (NIST) blackbody, during a comparison of TIR radiometers carried out in Miami by the Committee on Earth Observation Satellites (CEOS), and the accuracy was established at 0.2 K [32]. Special care was taken with the ground measurements in the sparse crops (vineyard and almond orchards), by averaging soil and canopy component temperatures to obtain representative values of the target LST. The radiometers were manually carried back and forth along transects on the fields pointing at nadir view, at a height of 1.5–2 m above the ground surface. Temperatures were registered at a rate of 5–10 measurements/min, in transect distances of 30–50 m/min, and then covering several hectares with each IRT. The 10-min averages centered at the satellite overpass time were considered. Radiometric temperatures were corrected from atmospheric and emissivity effects [33]. Downwelling sky radiance was measured with each radiometer and emissivity data were obtained through the Temperature-Emissivity Separation (TES) procedure [34,35] from in situ thermal radiance measurements using a multispectral radiometer CIMEL Electronique CE 312-2 [36].

**Figure 1.** Overview of "Las Tiesas" experimental farm. Measurement sites are located over a S2 false color composition corresponding to date 25 July 2018. Labels for the different study fields are explained in the adjacent Table, together with indication of crop type and field size.
