Estimation of corn yield using multi-temporal optical and radar satellite data and artificial neural networks

https://doi.org/10.1016/j.jag.2016.12.011Get rights and content

Highlights

  • A diagnostic and a real-time approaches are proposed to estimate the yield of corn.

  • Artificial neural networks are combined with multi-temporal optical and radar images.

  • Best yield estimation performance is obtained with red reflectance before harvest.

  • Combination of optical and radar signals provides accurate and early yield estimate.

Abstract

The yield forecasting of corn constitutes a key issue in agricultural management, particularly in the context of demographic pressure and climate change. This study presents two methods to estimate yields using artificial neural networks: a diagnostic approach based on all the satellite data acquired throughout the agricultural season, and a real-time approach, where estimates are updated after each image was acquired in the microwave and optical domains (Formosat-2, Spot-4/5, TerraSAR-X, and Radarsat-2) throughout the crop cycle. The results are based on the Multispectral Crop Monitoring experimental campaign conducted by the CESBIO (Centre d’Études de la BIOsphère) laboratory in 2010 over an agricultural region in southwestern France. Among the tested sensor configurations (multi-frequency, multi-polarization or multi-source data), the best yield estimation performance (using the diagnostic approach) is obtained with reflectance acquired in the red wavelength region, with a coefficient of determination of 0.77 and an RMSE of 6.6 q ha−1. In the real-time approach the combination of red reflectance and CHH backscattering coefficients provides the best compromise between the accuracy and earliness of the yield estimate (more than 3 months before the harvest), with an R2 of 0.69 and an RMSE of 7.0 q ha−1 during the development of the central stem. The two best yield estimates are similar in most cases (for more than 80% of the monitored fields), and the differences are related to discrepancies in the crop growth cycle and/or the consequences of pests.

Introduction

During the last fifty years, the world production of corn has increased at a rate of approximately 13 million tons per year, exceeding 1 billion tons in 2013 (FAO, http://faostat.fao.org/). To produce this amount, the area allocated to cultivation has regularly increased (by more than 1 million hectares per year), making corn the second most abundant crop in terms of area (after wheat). In France, corn is one of the main crops cultivated from spring to autumn, particularly in the Languedoc-Roussillon-Midi-Pyrénées region, where the yields observed during the 15 past years are comparable to those observed across the country (according to the statistics of the Direction Régionale de l'Alimentation, de l'Agriculture et de la Forêt, DRAAF). Many studies have demonstrated the usefulness of satellite images for agricultural purposes, taking advantage of both extensive coverage and regular revisits to map land use (McNairn et al., 2014), to detect irrigation (Fieuzal et al., 2011), to estimate biomass (Claverie et al., 2012, Battude et al., 2016) and to survey crop health (Yang et al., 2015). Nevertheless, the early monitoring of crops remains limited, and the few studies addressing this topic are focused on crop mapping (McNairn et al., 2014, Marais Sicre et al., 2016), while estimates of yield are usually obtained once the harvest has been done or by using in situ agronomic approaches, which are limited by their spatial sampling (Kalluri et al., 2003).

Remote sensing studies addressing the monitoring of crop phenology are mainly based on images acquired in the visible and near-infrared wavelengths. These data are often used to calculate vegetation indices, such as the normalized difference vegetation index (NDVI, Rouse et al., 1974) and other vegetation indices (Haboudane et al., 2004), which are mainly linked to the leaf area index (LAI) or to the fraction of absorbed photosynthetically active radiation (Asrar et al., 1984, Baret and Guyot, 1991). The main limitations of these approaches are related to the properties of optical images (almost unusable in conditions of heavy cloud cover). In the microwave domain, studies have demonstrated the application of radar data for crop monitoring, in particular the contribution of the multi-frequency, multi-polarization and/or multi-incidence aspects (Fieuzal et al., 2013, Fieuzal and Baup, 2016, Larranaga et al., 2013, Moran et al., 2012), such as the opportunities offered by the polarimetric and/or interferometric indices (Baup et al., 2016, Betbeder et al., 2016b, Lopez-Sanchez et al., 2012, Yang et al., 2014). The main limitations of using backscattering coefficients are their sensitivity to soil parameters (i.e., top soil moisture and roughness when vegetation is sparse) and the difficulty to constitute dense image series. Recently, multiple satellite missions (like Sentinel) have enabled the acquisition of regular Earth observations, providing quasi-synchronous optical and radar images that are useful to the understanding and comparison of the sensitivities of remote sensing signals.

Regardless of the spectral domain (i.e., optical or microwave), previous studies have taken advantage of the temporal dynamics of signals to estimate yield by assimilating crop parameters (e.g., LAI or crop biomass) derived from satellite images into crop models (Betbeder et al., 2016a, Fieuzal et al., 2016, Dempewolf et al., 2014, Dente et al., 2008, Duchemin et al., 2015, Kouadio et al., 2014, Rinaldi et al., 2013, Xin et al., 2013, Battude et al., 2016, Claverie et al., 2012). In these approaches, the accuracy is closely linked to the errors associated with the inversion of the intermediate assimilated parameter, in order to limit the multiplicative bias on the final variable. Other approaches based only on ground measurements enable direct monitoring of specific crop parameters and yield prediction by training statistical algorithms on the field data (Martin et al., 2012, Sharma and Franzen, 2014, Yin et al., 2011, Yin et al., 2012). The representativeness of the collected dataset often limits the application of these empirical approaches at the regional or larger scale, the range of validity being specific to the observed agricultural practices and environmental conditions. Among the wide range of statistical algorithms, artificial neural networks (ANN) offer the major advantage of improved prediction capability (with significantly better performance than multiple linear regressions, Lek et al., 1996), especially when relations are complex (as in the case of yield, which is related to variety, cultural practices, and climatic and edaphologic conditions). The prediction capabilities of ANN have been used in many fields (Lek and Guegan, 1999, Svozil et al., 1997), and their combination with the capabilities of Earth observation satellites is promising for land mapping and surface parameter retrieval (Villmann et al., 2003, Rodriguez-Fernandez et al., 2015).

The objective of this study is thus to take advantage of dense satellites series acquired in the optical and microwave domains over corn to estimate and forecast yields using a statistical method (ANN).

Section snippets

Study area

The study area is located in southwestern France near Toulouse (Fig. 1), a 420 km2 area centered on the coordinates: 43°29′36”N, 01°14′14”E. The network of monitored fields is located in the alluvial plain near the meteorological station of Lamasquère (the distance between the fields and the meteorological station is less than 7 kilometers). The region is governed by a temperate climate, with an annual rainfall of approximately 600 mm and mean daily air temperature ranging from a few degrees in

Methodology

The proposed methodology is illustrated by the following flowchart (Fig. 4). The useful steps for processing the images delivered by the different sensors are described first, with a focus on the angular normalization of SAR signals (Section 3.1). Then, the two approaches implemented to estimate corn yields (diagnostic approach and real-time approach) are presented, together with the statistical algorithm (section 3.2).

Estimation of the yield using the diagnostic approach

Fig. 5 presents an overview of the performance obtained with the multi-sensor satellite data (RMSE and R2) for the estimate of the yield of corn using all the images acquired during the crop cycle. The tested sensor configurations exhibit a wide range of performance, as illustrated by the R2 values (between 0.05 and 0.77) and RMSE (between 6.6 and 15.5 q ha−1). Among the tested optical wavelengths, the best results are obtained using optical signals acquired in the red wavelength region. These

Conclusion

The aim of this study was to estimate the yield of corn using dynamic satellite signals acquired by optical (Formosat-2 and Spot-4/5) and/or SAR (TerraSAR-X and Radarsat-2) satellite sensors between the sowing and harvesting of corn.

The first analysis, based on the diagnostic approach, shows that the yield estimates based on red reflectance have the best performance (with a coefficient of determination of 0.77 and a relative error of 6%). This approach offers the advantage of being solely based

Acknowledgments

The authors wish to thank the ESA (European Space Agency), DLR (German Space Agency), CSA (Canadian Space Agency), SOAR Project and CNES (Centre National des Etudes Spatiales) for their support, funding and satellite images (proposal HYD0611 and SOAR-EU and Categorie-1 ESA project no. 6843). In addition, the authors wish to thank the farmers (Mr. Blanquet, Mr. Bollati, Mr. Brardo, Mr. Pavan and Mr. Peres) for their time and precious discussion and the people who helped to collect the ground

References (60)

  • S. Lek et al.

    Artificial neural networks as a tool in ecological modelling. an introduction

    Ecol. Modell.

    (1999)
  • S. Lek et al.

    Application of neural network for nonlinear modeling in ecology

    Ecol. Modell.

    (1996)
  • C. Marais Sicre et al.

    Determination of the crop row orientations from Formosat-2 multi-temporal and panchromatic images

    ISPRS J. Photogramm. Remote Sens.

    (2014)
  • H. McNairn et al.

    Early season monitoring of corn and soybeans with TerraSAR-X and RADARSAT-2

    Int. J. Appl. Earth Obs. Geoinf.

    (2014)
  • A.K. Prasad et al.

    Crop yield estimation model for Iowa using remote sensing and surface parameters

    Int. J. Appl. Earth Obs. Geoinf.

    (2006)
  • D. Svozil et al.

    Introduction to multi-layer feed-forward neural networks

    Chemom. Intell. Lab. Syst.

    (1997)
  • T. Villmann et al.

    Neural maps in remote sensing image analysis

    Neural Netw.

    (2003)
  • H. Yang et al.

    Wheat lodging monitoring using polarimetric index from RADARSAT-2 data

    Int. J. Appl. Earth Obs. Geoinf.

    (2015)
  • L.J. Abendroth et al.

    Corn Growth and Development

    (2011)
  • G. Asrar et al.

    Estimating absorbed photosynthetic radiation and leaf-area index from spectral reflectance in wheat

    Agron. J.

    (1984)
  • A. Balenzano et al.

    Dense temporal series of C- and L-band SAR data for soil moisture retrieval over agricultural crops

    IEEE J. Select. Topics Appl. Earth Observ. Remote Sens.

    (2011)
  • F. Baup et al.

    Radar signatures of sahelian surfaces in Mali using ENVISAT-ASAR data

    IEEE Trans. Geosci. Remote Sens.

    (2007)
  • F. Baup et al.

    MCM'10: An experiment for satellite multi-sensors crop monitoring from high to low resolution observations

  • F. Baup et al.

    Sensitivity of X-band (σ°, γ) and optical (NDVI) satellite data to corn biophysical parameters

    Adv. Remote Sens.

    (2016)
  • J. Betbeder et al.

    Assimilation of LAI and Dry Biomass data from optical and SAR images into an agro-meteorological model to estimate soybean yield

    IEEE J. Select. Topics Appl. Earth Observ. Remote Sens.

    (2016)
  • J. Betbeder et al.

    Contribution of multi-temporal polarimetric SAR data for monitoring winter wheat and rapeseed crops

    J. Appl. Remote Sens.

    (2016)
  • H. Breit et al.

    TerraSAR-X SAR processing and products

    IEEE Trans. Geosci. Remote Sens.

    (2010)
  • DRAAF

    Direction Régionale De l'Alimentation

    (2016)
  • J. Dempewolf et al.

    Wheat yield forecasting for punjab province from vegetation index time series and historic crop statistics

    Remote Sens.

    (2014)
  • B. Duchemin et al.

    Impact of sowing date on yield and water-use-efficiency of wheat analyzed through spatial modeling and FORMOSAT-2 images

    Remote Sens.

    (2015)
  • Cited by (82)

    View all citing articles on Scopus
    View full text