Predicting wheat grain yield and spatial variability at field scale using a simple regression or a crop model in conjunction with Landsat images

Highlights

• Model estimated yield was not more accurate than a simple regression estimation.

• Assimilating remote sensing data into a crop model correctly estimated yield spatial variability.

• Remote sensing-model data assimilation could be useful to recreate yield maps from archive data.

• Adjusted model parameters during assimilation could provide useful derived metrics of the site.

Abstract

Early prediction of crop yields has been a challenge frequently resolved through the combination of remote sensing data and crop models. The aim of this study was to evaluate two different methods based on remote sensing data for predicting winter wheat (Triticum aestivum L.) yield at field scale. We compared the accuracy of: (i) a simple regression method between different vegetation indices at anthesis and grain yield, and (ii) a crop model method based on optimization of two parameters (specific leaf nitrogen and initial aboveground-biomass) using time series of vegetation indices. Vegetation indices were derived from Landsat-7 Enhanced Thematic Mapper Plus (ETM+) and Landsat-8 Operational Land Imager (OLI) images acquired for two growing seasons (2013, 2014) across 22 fields in south western Uruguay with an average size of 128 ha. At all sites, leaf area index (LAI) was measured during a field campaign, and grain yield was measured with yield monitors on harvesters. The simple regression method (SRM) achieved higher accuracy than the model-based method (CMM) for the estimation of yield at field scale (RMSE = 966 kg ha⁻¹ and RMSE = 1532 kg ha⁻¹, respectively). When deviations between observed and estimated yields were evaluated at pixel (30 × 30 m) level, the model-based method was better at detecting existing spatial variability in grain yield and at identifying areas of different yield potential. Even though both methods have limited utility to estimate yield at field scale with very high accuracy due to large RMSE, the methodologies are suitable to predict harvest volumes at large agricultural areas or at country level, and to construct synthetic yield maps reflecting within field variability. Higher temporal resolution of images would improve accuracy in estimating yield and spatial variability at field scale.

Introduction

The concerns about national food security and sustainable agricultural development have increased in recent years and therefore, a key component is the precise estimation of supply and demand, in particular for major crops such as wheat. During the past four decades the prediction of crop yields has been undertaken by combining remote sensing data and crop models (Maas, 1988, Moulin et al., 1998, Fang et al., 2008, Casa et al., 2012, Jin et al., 2016, Jin et al., 2018). However, the application of crop models over large areas has been hampered by the lack of adequate and accurate information about parameters and input data, (e.g. variability in initial conditions and in soil properties) making it difficult to achieve high precision using a crop model at field scale (Moulin et al., 1998). For this reason, the development of crop growth models has been focused on the estimation of growth and yield in agricultural areas where the soil, weather and management practices were well known (Maas, 1988, Jin et al., 2018).

In Uruguay, like in many other parts of the world, soil properties of agricultural areas are very heterogeneous and, as a result, growing conditions for crops are highly variable (Berger et al., 2018). To enable the usage of crop simulation models over large areas, it is necessary to first recognize homogeneous areas with comparable initial conditions and parameter values. Consequently, the application of models to predict crop growth and yield at the field scale continues being a challenge which has been partially solved. Reflectance data and vegetation indexes are useful tools to quantify crop production and spatial variability in small areas (Baret and Guyot, 1991, Launay and Guerif, 2005, Moulin et al., 1998). While remote sensing provides a quantification of the state of the attributes of the canopy (e.g. leaf area index, underground biomass accumulated, nitrogen content) in discrete time, growth models provide continuous description of plant growth (Maas, 1988, Moulin et al., 1998, Delécolle et al., 1992, Hatfield et al., 2008).

The complementarity of crop growth models and remote sensing data to improve the accuracy of estimated yield is widely known and has been suggested and developed by several authors (Maas, 1988, Delécolle et al., 1992, Moulin et al., 1998, Padilla et al., 2012, Jin et al., 2016, Silvestro et al., 2017). Delécolle et al. (1992) described three schemes to assimilate remote sensing data into a crop model: (i) forcing, which consists in using a variable estimated by remote sensing directly; (ii) updating, where state variables of model (such as LAI) are renewed from remote sensing data as it becomes available; and (iii) re-calibration, where the relationship between remote sensing state variables and the simulated state variables is used to recalibrate parameters of the crop model. This last method is based on the hypothesis that the model equations precisely represent crop growth, but the parameters or input variables must be adjusted, and the new observational data that becomes gradually available allows for this adjustment. Even though such recalibration method requires high computation time, several published studies during the last years have been reporting the estimation of yield based on the combination of crop models and remote sensing data (Fang et al., 2008, Jin et al., 2016). Some studies have shown that the use of prediction models in conjunction with remote sensing data can give better results than sophisticated crop models alone (Kogan et al., 2013). In addition, in a data assimilation context, simple crop growth models (e.g. SAFY) are more suitable than complex biophysical models (Silvestro et al., 2017) due to a reduced number of parameters to estimate and a more parsimonius description of crop growth.

Among the different approaches used for forecasting yield during the last years, the ones based on the application of a regression between a vegetation index and yield (forcing) were more frequent and achieved variable success. Becker-Reshef et al. (2010) achieved accuracy between 7 and 10% when they estimated yield one month to a month and a half prior to harvest using a regression between the NDVI peak (maximum) and yield in Kansas and Ukraine. This study demonstrated that with simple methods and limited surface data, high precision in the estimation of regional wheat production is feasible.

Another important element is the scale of estimation, which has a direct effect on the aggregation of input data for modeling purposes (Hoffmann et al., 2016) or in the aggregation of remote sensing data (i.e. small pixel size vs large pixel size). Although the lower spatial resolution of imagery should have less variability than the higher due to aggregation, some studies have demonstrated that the use of low spatial resolution (more than 1.1 km⁻²) did not assure high precision in yield estimation across different crops (Doraiswamy et al., 2005). Bastiaanssen and Ali. (2003) obtained relative root mean square error (RRMSE) of 26% using Advanced Very High Resolution Radiometer (AVHRR) data for wheat yield estimation in Pakistan. While Doraiswamy et al. (2005) obtained an error around 10% at county level when they integrated crop growth models with MODIS imagery at 250 m resolution. In contraposition, Azzari et al. (2017) did not found advantage in terms of RRMSE for increased spatial resolution when they compared two methods (simple regression using a peak of vegetation index and APSIM model) to estimate wheat yield in northern India with two spatial and temporal resolutions (MODIS and Landsat).

Less effort has been dedicated in estimating yield at field scale (i.e. within the field) using methods based on remote sensing or crop modeling, and much less combining the two. An example is Silvestro et al. (2017) who applied an updating method to assimilate Landsat images into a simple crop model (SAFY). This author shows the feasibility of using medium resolution satellite data provided by HJ1A/B or Landsat 8 OLI, to estimate yield at field scale, with potential applications for precision agriculture. Time series of satellite images has a potential to enhance the estimation of crop yield through forcing an ecophysiological model with recalibration (Morel et al., 2014). However, achieving high accuracy in estimating the spatial variability at field scale using high spatial resolution continues to be a challenge.

The objective of this study was to compare high spatial resolution observed yield data (harvester yield monitor) with two methods for estimating wheat grain yield at field-scale using Landsat images: (i) an empirical method where the relationship between Landsat-derived vegetation index at anthesis and measured grain yield was used, and (ii) a semi-empirical method where time series of Landsat-derived vegetation indices were used to derive two parameters of a crop model.