Potential of water balance and remote sensing-based evapotranspiration models to predict yields of spring barley and winter wheat in the Czech Republic

Highlights

• Results of the study demonstrated that ET-based indicators can be used for yield prediction prior to harvest.

• ET-based indicators can be employed as a useful tool for assessment of drought and its impact on agricultural crops.

• ET-based indicators can be used in real time during the season and thus have great potential for decision making at regional and district levels.

Abstract

Indicators based on evapotranspiration (ET) provide useful information about surface water status, response of vegetation to drought stress, and potential growth limitations. The capability of ET-based indicators, including actual ET and the evaporative stress index (ESI), to predict crop yields of spring barley and winter wheat was analyzed for 33 districts of the Czech Republic. In this study, the ET-based indicators were computed using two different approaches: (i) a prognostic model, SoilClim, which computes the water balance based on ground weather observations and information about soil and land cover; (ii) the diagnostic Atmosphere–Land Exchange Inverse (ALEXI) model based primarily on remotely sensed land surface temperature data. The capability of both sets of indicators to predict yields of spring barley and winter wheat was tested using artificial neural networks (ANNs) applied to the adjusting number and timeframe of inputs during the growing season. Yield predictions based on ANNs were computed for both crops for all districts together, as well as for individual districts. The root mean square error (RMSE) and coefficient of determination (R²) between observed and predicted yields varied with date within the growing season and with the number of ANN inputs used for yield prediction. The period with the highest predictive capability started from early-June to mid-June. This optimal period for yield prediction was identifiable already at the lower number of ANN inputs, nevertheless, the accuracy of the prediction improved as more inputs were included within ANNs.The RMSE values for individual districts varied between 0.4 and 0.7 t ha^–1 while R² reached values of 0.5–0.8 during the optimal period. Results of the study demonstrated that ET-based indicators can be used for yield prediction in real time during the growing season and therefore have great potential for decision making at regional and district levels.

Introduction

Agricultural drought can significantly impact crop yields, as demonstrated both recently (e.g. Blauhut et al., 2016; Jakubínský et al., 2019) and in the more distant past, e.g. in the pre-instrumental period (e.g. Brázdil et al., 2019), and can have global implications for food production and security (Trnka et al., 2019). The impacts of agricultural drought on crops can be most directly identified by the reduction of yield (Fiala et al., 2014); however, not only the quantity but also the quality of production can be affected by drought occurrence (Jensen et al., 1996). Understanding of the relationship between water status indicators and crop yields is crucial for assessing the drought impacts on agricultural crops (Anderson et al., 2016a, Yang et al., 2018). While the final impacts of drought on agricultural crop yields can be fully assessed only at harvest (yield observation), early estimates of yield (yield forecasting) are very useful for farmers, traders and policy makers (Zhang and Kovacs, 2012). Yield forecasting is typically realized using a set of real-time predictor time series, with forecast accuracy gradually increasing as the growing season progresses toward the peak performance in the interval prior to harvest (Bognár et al., 2017). In this way, the final crop yield can be to some degree predicted several weeks up to months before harvest, providing useful support for planning of agricultural operations (Hlavinka et al., 2019, Nagy et al., 2018).

One of the most widely used approaches for monitoring crop conditions is based on remote sensing (Tucker et al., 1980, Johnson, 2016; Wardlow et al., 2012), which utilizes the connection between the spectral reflectances of certain plant species and their yield (e.g. Tucker et al., 1980). This approach has been used since 1972 when Landsat 1 was launched (Rembold et al., 2013), representing the first civil satellite orbiting the Earth. In addition to satellites, data from airborne imaging and unmanned aerial vehicle surveys can be used for monitoring of vegetation (Zhang and Kovacs, 2012). Commonly used remote sensing indicators embody vegetation indices that have the capability to track crop progress and evolution in green biomass (Becker-Reshef et al., 2010), including the normalized difference vegetation index (NDVI), the enhanced vegetation index (EVI), or its two-band version (EVI2).

Another effective indicator of crop health is evapotranspiration (ET), which over agricultural lands represents water consumed by crops in the process of transpiration and water evaporated from the soil, leaves and other surfaces. During the growing cycle, anomalously low rates of actual ET relative to maximum expected ET may indicate depleted soil moisture and stress-related reductions in transpiration fluxes (Anderson et al., 2007a). Therefore, ET-based indicators have been used to monitor agricultural drought (Anderson et al., 2011, Anderson et al., 2016b, Potopová et al., 2015) and impacts on crop yields (Anderson et al., 2016a, Crocetti et al., 2020, Yang et al., 2018) including indices based on actual ET (ET_a) and ET_a normalized by reference ET (ET_o). The relationship between ET and crop health and productivity has been studied for a long time and provides basis for various approaches to monitor crop water stress. These included e.g. the crop coefficient model (Allen et al., 1998) combined with the approach based on radiation interception and light use efficiency (Monteith, 1977) or a crop simulation model based on the concept of water use efficiency (Raes et al., 2009, Steduto et al., 2012).

For regional monitoring of ET, two modeling approaches are generally used (e.g. Yilmaz et al., 2014; Anderson et al., 2007a): prognostic and diagnostic models, each with its advantages and disadvantages. ET as a water balance component can be estimated by the use of prognostic land surface models (Yilmaz et al., 2014, Mendiguren et al., 2017, Arnold et al., 1998, Abbott et al., 1986). In this study, we considered the case of the SoilClim water balance model widely used in the Czech Republic (Hlavinka et al., 2011, Trnka et al., 2020). Other water balance models include the Variable Infiltration Capacity (VIC) model (Wood et al., 1992), the Soil and Water Assessment Tool (SWAT) (Arnold et al., 1998) and the FAO model AquaCrop (Raes et al., 2009, Steduto et al., 2012). Prognostic assessment of ET can be also performed using land surface models (LSMs) such as Noah LSMs (e.g. Ek et al., 2003). Prognostic models use basic weather variables (e.g. precipitation, wind, radiation, air temperature) together with information about soil (e.g. field capacity and wilting point) and vegetation (e.g. leaf area index – LAI, rooting depth, crop coefficient) to model the complete water and energy balance in the soil, plant and atmosphere interface. Predicted ET rates depend strongly on the accuracy and availability of input weather variables and parameters related to soil and vegetation, as well as on the accuracy of the model itself (Beljaars et al., 1996). Many of these inputs can be difficult to specify at regional or continental level, and biases may result in cumulative errors in soil water content and associated fluxes. These biases can even become larger in the long-term perspective if there are no mechanisms for periodic correction (Schaake et al., 2004).

In contrast, diagnostic models of ET require less information about soil and vegetation properties and generally do not require precipitation as input. To estimate ET, these models rely on remote sensing retrievals of land surface temperature (LST), albedo, vegetation cover fraction or LAI. Therefore, diagnostic models can provide the information about the plant water status whenever remote sensing data are available independently on the previous imaging. As estimation of ET is based on remote sensing data, these models are largely independent from ground measurements and ET-based indicators can be calculated over the large areas, from country to continent (e.g. Europe) to global scales. On the other hand, diagnostic models based on thermal and visible/near-infrared remote sensing data inherently have one significant limitation compared to prognostic ground-based models: input data can be obtained only during clear-sky conditions. Examples of diagnostic ET models include the Surface Energy Balance System (SEBS) (Su, 2002), the Surface Energy Balance Algorithm for Land (SEBAL) (Bastiaanssen et al., 1998) and the Two-Source Energy Balance model (TSEB) (Norman et al., 1995). The diagnostic model considered here is the Atmosphere–Land Exchange Inverse (ALEXI) model (Anderson et al., 1997, Anderson et al., 2007a), which is based on TSEB and uses time changes in LST derived from remote sensing to partition the surface energy budget. The evaporative stress index (ESI) is an ALEXI based product representing standardized anomalies in the ratio of actual to reference ET (ET_a/ET_o) (Anderson et al., 2007b). ESI has been demonstrated to be consistent with standard precipitation-based drought indices while additionaly capable to provide quick response to changes in vegetation health during rapid drought onset events (Otkin et al., 2016).

Recent efforts toward developing operational yield forecasting systems based on different predictors have been conducted within several regions in Central Europe. In Hungary, for example, several studies have investigated the relationship between remote sensing data and crop yields (Bognár et al., 2017, Kern et al., 2018, Nagy et al., 2018). In the Czech Republic, ESI was tested for spring barley and winter wheat yields (Anderson et al., 2016a), the two most widely cultivated cereal crops in the region. Based on the study of Anderson et al. (2016a) and several analyses focusing on vegetation indices (e.g. Johnson, 2016), a prototype remotely sensed forecasting method for key crops for the whole country was put into use in 2018. This system was initially based on vegetation indices (NDVI, EVI, EVI2) from the Moderate Resolution Imaging Spectroradiometer (MODIS) with 250 m spatial resolution. Calibration of yield forecasts for 2018 were based on yield data and vegetation indices from 2000 to 2017. This operational forecasting method has capability to predict the final yields more than two months prior to harvest.

In this study, we tested the predictive capability of ET-based indicators (ET_a and ESI) toward incorporating them into the operational yield forecasting system. The ET-based indicators were tested for the period 2001–2018 over 33 districts in the Czech Republic and were generated with models representing two different approaches: prognostic ground-based (SoilClim) versus diagnostic remote sensing-based model (ALEXI). In the previous study of Anderson et al. (2016a), linear regressions between yields and ESI from ALEXI were tested at regional and district levels. In the current study, the method of the artificial neural networks (ANNs) at the district level was investigated. The main objective was to (i) to determine whether an indicator from a prognostic or diagnostic model is preferable to be included into the operational yield forecasting system. The other objective was (ii) to identify the optimal period during the growing season when the capability to predict yield is the highest. This was tested by the adjusting number and timeframe of ANN inputs during the growing season.