Remote Sensing of Crop Yield and Crop Yield Estimation at Regional Scale Lauriane Cayet‐Boisrobert MSc in Remote Sensing, department of Geography, University College of London
[email protected]
Course work paper Principles of Remote Sensing: Matt Disney
Received 17 November 2005
Abstract Regional to global scale crop yield estimation and prediction is important for food security warning, food trade policy, and Carbon cycle research. Remote sensing of crop yield offers great potential for regional crop production and yield estimation. Optical sensors (Landsat/MODIS/AVHRR) are mostly used because they measure radiance in the Red and Near Infra Red (NIR) and because of the high correlation between the NIR and crop production. The crop yield estimation methodology consists in retrieving biophysical parameters from the vegetation indices, and then in deriving crop yield from either a statistical method or an efficiency method. This method is reliable as far as medium resolution (30‐250m) images are used. Therefore, AVHRR‐based crop yield estimation is not reliable. Furthermore, reliable results can be obtained at national scale with meteorological‐based crop growth models and promising methods such as crop yield estimation via SAR technology are emerging.
1
Introduction Crop yield is a crucial parameter to international and bi‐lateral organizations, EU commissions, and Governments of countries. This parameter is desirable for adjusting food pricing and determining trading policies, but also for ensuring food security in the developing countries. Additionally, estimating accurately crop yield is a key element to understand the global carbon balance and the location of missing carbon sinks, and also to help predict global climate change (Running et al., 2000). Crop yield assessment at the field level or non‐very accurate regional estimation can be done from the ground by direct methods. The aged method consists in harvesting and weighting small samples and then in extrapolating to the surface area considered. The non‐destructive method consists in obtaining allometric relationships between biomass and more readily obtained measures such as Leaf Area Index (LAI) or height (Running, 2000). These two methods require field surveys and the result is not accurate because of crop yield variability. Patchy soil types, water gradient, variable slopes, and different agricultural management types and other parameters cause yield variability. Agricultural statistics such as those from the USDA National Agricultural Statistics Service (NASS) cited by Doraiswamy et al (2005) are reliable in estimating crop yields by sampling the field measurements of standing crops, but this method is time‐consuming and very costly (Liang, 2004). Hence, remote sensing is a promising alternative primarily in developing countries and remote areas. This essay discusses how remote sensing can support crop yield estimation at regional scale. The most common methodology is to derive biophysical parameters from the combination of particular wavebands radiance and thus reflecting the process of biomass production by a simple statistical method or trough a simulation model. In the first part, the essay explains how remote sensing supports retrieval of the biophysical parameters essential to estimate crop yield. The second part sums up all the different crop yield models based on Remote Sensing techniques, and to finish, it gives a detailed assessment of the different methods’ accuracy and limitations.
2
1. Remote sensing of crop yield 1.1. Vegetation’s absorption property Plants synthesize their biomass in a production process which takes place in the green parts of the plant (mostly in the leaves): photosynthesis (Jensen, 2000). Photosynthesis is an energy‐storing process which converts light energy into chemical energy which is stored in a simple carbohydrate molecule (Glucose). Therefore, plants are one of the key elements of the Carbon Cycle. The radiations from Sun to the Earth range from 0.2μm (UV) to 2.5 μm (Middle Infrared), but only a certain part of this energy is used in the photosynthesis process. It is called the Photosynthetically Active Radiation (PAR) ranging from 0.4 μm to 0.7μm, which corresponds to the visible part of the light spectrum (Jensen, 2000). However, only a portion of PAR is absorbed (APAR) by green plants (preferably Red and Blue radiation). Then, some of it is either re‐radiated, lost as latent heat, or stored by the activity of photosynthesis in organic substances.
Fig1: Spectral Reflectance characteristics of green vegetation. From www.crisp.nus.edu.sg/~research/tutorial/optical.htm If we look closely at the reflectance spectrum of vegetation (Fig. 1), we can notice that there is a sharp edge in the reflectance over the wavelengths 0.68‐0.75μm. This has been called the ‘red edge’ (Barrett et al., 1999). Additionally, it has been proven that a strong relationship exists between chlorophyll and the red‐infrared portions of the spectrum (Barrett et al., 1999). Actually, this edge is the basement of vegetation remote detection and Vegetation Indices (VI) are derived from this edge property resulting of chlorophyll photosynthetic activity. 1.2. Space‐borne instrument suitability We will see later that the APAR can be estimated by VI’s reflectance in the Red and NIR parts of the spectrum. For this reason, sensors must be able to differentiate the amount of energy coming out from different wavebands. Hence space‐borne remotely sensed data have an advantage over standard aerial photography in that they can sample wavelength (Running et al., 2000). Additionally, yield monitoring is possible as Earth resource satellites such as Landsat systems are launched successively for data continuity (Jensen et al., 2000 and Running et al., 2000), thus yield monitoring can be done continuously. Additionally, space‐borne remote sensed imagery provides a standardized digital product, rapid‐repeat coverage, and regional to global dataset (Running et al., 2000). Landsat TM for example has a 185 km wide swath, and its repeating cycle is every 16 days. With larger coverage capacity (swath width = 2700 km), but lower resolution (1.1 × 1.1 km), NOAA‐AVHRR is also a relevant tool for assessing biomass production over regions. Since, NOAA has a high re‐visiting frequency (twice a day), unlike Landsat, it can measure daily NPP.
3
1.3. Retrieving biophysical parameters Biophysical parameters such as Leaf Area Index (LAI) and fractional Photosynthetically Active Radiation (fAPAR) are proxies that can support estimating the energy uptake by the photosynthesis. It seems that most of the recent crop yield estimation research focuses on fAPAR and most of the methods used to retrieve that parameter rely on statistical regression based on Vegetation Indexes. 1.3.1. Statistical method LAI can be retrieved because of the linear relationship between NDVI values (Eq. 1). SR has been also used for estimating crop LAI according to Liang (2004). LAIi = LAImax (NDVIi – NDVImin)/(NDVImax + NDVI min)
(1)
fAPAR can be calculated from SR as equation (Eq. 3). According to Lobell et al (2003), several studies suggest that replacing SR with NDVI (in Eq. 2) results in more accurate fAPAR estimates. Additionally, some authors (Lobell et al., 2003 and Tao et al., 2005) have calculated fAPAR as the average fAPAR from SR and NDVI. fAPAR = (SR – SRmin ) (fAPARmax – fAPARmin)/(SRmax + SR min) + fAPARmin
(2)
1.3.2. Satellite image reflectance adjustment with a look up table method (LUT) This method creates a LUT of modeled reflectance made from LAI measurements. Doraiswamy et al. (2005) used this method to invert LAI from MODIS 8‐day composite image via a radiative transfer model called SAIL. SAIL model served to create a LUT relating modeled canopy reflectance for MODIS band 1 and 2 to LAI. Sail input model requires all its input parameters from the field: Visible and NIR values of leaf reflectance and leaf transmittance, LAI, canopy leaf angle distribution (LAD), hot spot parameter, background reflectance, solar zenith and azimuth angles, sensor view angle, and proportions of direct and diffuse short wave lengths radiation. 2. Crop yield estimation 2.1. Crop production rate Primary productivity and yield quantify crop production rate. Primary production over a given period is termed primary productivity (g/m2/year) and primary productivity declines itself into Gross Primary Productivity (GPP) and Net primary productivity (NPP). GPP is all of the light energy that is converted to chemical energy by the primary producers considered. NPP is the gross primary productivity minus that respiration. In this essay, yield or NPP is employed indifferently because yield only differs from NPP by a factor of multiplication. Some example to calculate yield from productivity are shown below (Eq. 3 and 4). Yield can also be derived from biomass using a statistical model (Eq. 5). Tao et al, 2004 (maize across China)
Yd = NPP × AR × HI / (1 ‐ m0 )
(3)
Lobell et al, 2002 (maize, wheat, soybean in the Yaqui Valley, Mexico)
Yd = APAR × ε × HI
(4)
Yd = a GUYRI + b
(5)
Roebeling et al, 2004
4
Where, Yd is yield in g. m‐2; AR is the ratio of above‐ground Biomass to total biomass; HI is the harvest indices, defined as the ratio of seed yield to above‐ground biological yield, m0 the moisture content of grains, ε is the light use efficiency in units of g biomass MJ‐1; APAR in MJ. 2.2. Crop yield estimation methods based on the biophysical parameters 2.2.1 AVHRR indices Many studies have been conducted to assess the correlation between AVHRR derived NDVI and crop yield. Mkabella and Mkabella (2005) found a positive linear correlation between cotton yield and cumulative maximum NDVI in the Lowveld of Swaziland in 2000. Hereafter (Eq. 6,7,8) are some examples of prediction formula used by Mkabella and Mkabella (2005). They differ from one region to another. Middleveld
Yd = 5.7045 (ΣAveNDVI) – 3.8171
(6)
Lowveld
Yd = 4.0224 (ΣAveNDVI) – 4.4442
(7)
Lubombo Plateau
Yd = 1.9695 (ΣAveNDVI) – 4.45591
(8)
Where here Yd is maize yield (t.ha‐1) and ΣAveNDVI is cumulative average NDVI. Ferencz et al (2004) derived a new vegetation index (General Yield Unified Reference Index (GYURI) based on AVHRR images and found very a good correlation (R2 = 0.84.6 – 87.2) between AVHRR‐ based GYURI and corn yield estimation at county level in Hungary (Eq. 9). Yd = ar GUYRI + br
(9)
Where ar and br are constants determined during the Calibration and development phase 2.2.2. Efficiency models Physiological models aim to integrate daily APAR over the growing season and incorporate it into a model based on crop light use efficiency (ε) to predict crop production. CASA (Eq.10) and GLO‐PME2 (Eq. 11) are two efficiency models. They have been used extensively by Lobell et al (2003) and Tao et al. (2005). CASA:
NPP = ε × APAR where APAR = Σ (PAR × fAPAR) Δt
(10)
Where ε is in units of g biomass MJ‐1. GLO‐PME2:
NPP = Σ 〔(σ ε*g )〕(PAR ×fAPAR) Pg Pm Δt
(11)
Where ε*g is the unstressed photosynthetic potential in terms of gross production (g CMJ‐1), the reduction of ε*g caused by stressors at time (t); Pg and Pm and are the growth and maintenance respiration terms representing proportion of photosynthetate remaining after carbon losses to these processes. 2.3. Climate‐based physiological models
Doraiswamy et al. (2005) assessed crop yield with a complex model. MODIS‐derived LAI model was used to adjust a climate‐based crop yield model.
5
3. Comparisons, limitations and alternatives (see annex 1) 3.1. Accuracy comparisons 3.1.1. To retrieve the biophysical parameters Doraiswamy et al. (2005) retrieved the biophysical parameter by simulating a MODIS‐based LAI model adjusted with SAIL. To prove the method’s reliability, they plot a reflectance scattergram of the MODIS image and SAIL (Fig. 3). They observed a good correspondence, except a few MODIS‐based points at the bottom were outside or the LUT envelop. They correspond actually to mixels (reflectance of soil and canopy). They also checked MODIS reflectance obtained with field measurements (8‐band cropscan radiometer with same bands as MODIS band 1 and 2). They obtained a coefficient determinant (R2) equal to 0.81 and RMSE to 0.81 which make the method reliable. Lobell et al (2003) used a statistical model but they did not assess its accuracy. However, they took the precaution to use the average fAPAR derived from two different VI obtained from AVHRR. Since this statistical method has been used for a long time (Asrar et al. (1984) used NDVI and Steinmetz et al. (1999) used SR) and recent studies made by Los et al. (2000) indicate that an average of SR and NDVI perform best, we can assume that the accuracy of AVHRR indices method to retrieve must be reliable. Tao et al. (2005) did not either assess the reliability of the statistical method they used to derive fAPAR from Landsat TM.
Fig. 3. The scattergram of field mean reflectance for MODIS band 1 (Red) and band 2 (NIR) for corn crop. The blue data points represent the data form the LUT and the red data pints the ground‐based measurements (Doraiswamy et al., 2005).
6
3.1.2. Methods used to estimate crop yield
Lobell et al. (2003)’s efficiency model based on fAPAR derived from Landsat reflectance were compared with field‐based estimates and errors were of less than 4% predicting wheat yield for 2 dates. Doraiswmay et al. (2005) who estimated crop yield from a climate‐based model adjusted with MODIS‐LAI model reported their results were within 10% of the County yields reported by the USDA NASS. Moving from regional to national scale, Tao et al. (2004) compared two different efficiency models (CASA and GLO‐PME2) and observed yields. Both models tend to underestimate maize yield and the fractional difference between estimated and observed lies between ‐0.5 and 0.5 which is not reliable enough. The authors concluded that both the two models were potentially useful for NPP monitoring or yield forecasting at regional or national scale, but better result could be obtained with high resolution image. Makhabella et al (2005) also used an AVHRR‐based crop yield simulation model. They tested it within 3 regions and obtained quite low results (R2 were 0.61, 0.68, 0.51).
3.2. Limitations 3.2.1. Sensor characteristics ‐ Spatial resolution The spatial resolution of the instrument is a critical limitation. Reliable crop yield estimates are possible with MODIS and Landsat TM at to regional scale, but not with AVHRR (resolution=1.1km2). ‐ Optical sensors cloud sensitivity In Europe, the cloud cover makes it difficult to obtain more than 2 repetitive scenes using visible sensors. Thus, radar could be an alternative solution. Additionally, as AVHRR imagery covers the whole Earth twice a day, frequency to get cloud free images increases but at the cost of decreasing spatial resolution 3.2.2. Methods ‐ Statistical methods vs. simulation models The regression based approach is applicable only for a given region and the same range of weather conditions where it was developed. Therefore, this method cannot be simply transferred. The relationship has to be calculated each time for different regions. ‐ Simplistic models Statistical and efficiency models are simple models because they only use one physiological process whereas weather conditions are important as well, therefore this method is not consistent under varying conditions (Barrett et al, 1999). Hence, simple models cannot compete with crop growth models and climate‐based physiological model used by Doraiswamy (2005). However, they are suitable for assessing yield at regional scale where minimum input data is available. ‐ Ground data requirement Some models require acquiring ground variable like SAIL (to retrieve biophysical parameters) or the climate‐ based physiological model. 3.3. Alternatives remote sensing methods to crop yield estimation as regional scale 3.3.1. To improve crop yield estimation reliability at national scale
7
Using a meteorological‐based crop growth model instead of biophysical‐based yield predictions can improve crop yield estimation considerably at country level. Traditionally, crop growth models are computed by programs that integrate information on daily weather, crop management, soil characteristics, genetics or varieties, and pest stress to determine daily plant growth and subsequent yield (Liang, 2004 and Jagtap et al, 2002). Roebeling and al. (2004) derived meteorological information from Meteosat and integrated it in the EARS‐GCS model. This model is entirely driven by daily information on meteorological indicators derived from the visible and the infrared bands (global solar radiation and evapotranspiration). When comparing it to forecasts of the European Statistical Office (EUROSTAT) and Monitoring Agriculture by Remote Sensing‐ Crop Growth Monitoring System (MARS‐CGMS), they obtained with predictions errors of 4%. 3.3.2. Radar technology utilization Shao et al (2001) showed that RADAR’s backscatter provide confidence that multitemporal RADARSAT data is capable of rive mapping and can provide input to production estimate but the results is far for perfect. They used HH polarization and C‐Band only, thus they encourage further research using multifrequency and multipolarization data to be able to estimate more than one target. However, Li et (2003) obtained a good correlation between statistical data and their prediction from Radarsat backscatter with miltitemporal images. The regression model they used is below (Eq. 12). The procedure used is very low‐cost and operate in cloudy areas. Yd = ‐27.212 X1 + 34,848 X2 + 13.84 X3 + 481.992 R=0.908
(12)
Where X1 is the backscatter coefficient of the first temporal image; X2 of the second temporal image, X3 of the third image.
Conclusion Accurate crop yield estimation at regional scale (such as agricultural valleys or counties) can be achieved by retrieving biophysical parameters using them as input parameters in crop yield simulation models or efficiency models. Satellites with medium resolution such as LANDSAT and MODIS are required. If low resolution images such as AVHRR imagery are used, then the crop yield model is no longer reliable. At country level, sun‐synchronized satellites cannot be utilized because either swath dimension is too small or spatial resolution is too coarse. However, crop yield forecast can be done regularly and in a reliable manner using a geostationary satellites (Roebeling et al., 2004). METEOSAT‐derived meteorological indicators can serve as input parameters to a crop growth algorithm and results obtained are very reliable. SAR technology is very promising but little research has been carried out for the moment.
8
ANNEX
9
Spatial resolution
Spectral resolution
Utilized wavelengths
Doraiswamy et al, 2005 Scale = US county
Band 1 (Red) Band 2 (NIR)
Method to derive the biophysical parameters LUT (SAIL) dependant on input ground data MODIS‐derived LAI by a statistic method
Biophysical parameter yield simulation input Seasonal LAI
Crop yield simulation method Climate‐based crop yield model adjusted with MODIS derived LAI
MODIS on TERRA satellite
250m
36 bands between 0.405‐ 14.386 μm
Landsat TM
30m
1.1km
Yield simulation model ancillary inputs
Band 1 (Blue) Band 2 (Green) Band 3 (Red) Band 4 (NIR) Band 5 (IR)
Band 3 (Red) Band 4 (NIR)
Statistical methods from SR and NDVI
fAPAR (daily to biweekly intervals)
Simple Efficiency model
N/A
Band 1 (Red) Band 2 (NIR) Band 3 (MIR) Band 4 &5 (FIR)
Band 1(Red) Band2 (NIR)
Statistical methods from SR and NDVI
fAPAR
Efficiency model named CASA
N/A
Efficiency model named GLO‐ PEM2
N/A
From weather stations distributed around the study area: - Daily Tmax - Daily Tmin - Amount of rainfall
NOAA‐ AVHRR
NOAA‐ AVHRR
1.1km
Band 1 (Red) Band 2 (NIR) Band 3 (MIR) Band 4 &5 (FIR)
Band 1(Red) Band2 (NIR)
Simply NDVI
Satistical method (cumulative average NDVI)
N/A
METEOSAT first generation geostationary centered of the equator
‐ VIS: 2,25km ‐WV & IR: 5km
Band 1 (VIS: 0.4‐ 1.1) Band 2 (IR) Band 3 (TIR)
Band 1 (VIS) Band 2 (IR)
No biophysical parameters model
N/A
Metorological‐ based‐Crop growth algorithm (EARS‐CGS)
‐ Evapotranspiration (calculated from the surface energy balance) ‐ Maintenance respiration (% of standing biomass)
Roebeling et al, 2004 Scale = EU countries
Tao et la, 2005 Scale = wide country
Makhabella et al, 2005
Sensors
Lobell et al, 2003 Scale = country region
Table 1: Summarize of some relevant current methods used to estimate crop yield at regional scale.
10
References Barrett E.C., Curtis L.F. 1999. Chapter 16: Landuse and crop production. In: Introduction to environmental Remote Sensing. 3rd. New York: Chapman & Hall. p 1347 – 368. Doraiswamy P. C., Sinclair T. R., Hollinger, S., Akhmedov B., Stern A., Prueger P. 2005. Application of MODIS derived parameters for regional crop yield assessment. Remote Sensing of Environment. 97: 192 – 202. Jagtap S., Jones J. 2002. Adaptation and evaluation of the CROPGRO‐soybean model to predict regional yield and production. Agriculture, Ecosystems and Environment. 93: 73 ‐ 85. Jensen J.R. 2000. Chapter 10: Remote Sensing of vegetation. In: Remote Sensing of the Environment: and Earth Resource Perspective. Jersey: Prentice Hall. p 333 ‐ 377. Liang S. 2004. Chapter 8: Estimation of land surface variables. In: Jin Au Kong Editors. Quantitative remote sensing of land surfaces. New Jersey: John Wiley & Sons. p. 246 – 309. Liang S. 2004. Chapter 13: Estimation Applications. In: Jin Au Kong Editors. Quantitative remote sensing of land surfaces. New Jersey: John Wiley & Sons. p 472– 524. Lobell D., Asner P., Ortiz‐Monasterio J. I., Benning T. 2003. Remote Sensing of regional crop production in the Yaqui Valley, Mexico: estimates and uncertainties. Agriculture, Ecosystems and Environment. 94: 205‐220. Mkhabela M. S. Mkhabela M. S. and Nkosazana N. M. 2005. Early maize yield forecasting in the four agro‐ecological regions of Swaziland using NDVI data derivfed from NOAA’s – AVHRR. Agricultural and forest meteorology. 129: 1‐9. Roebeling R. A., Van Putten E., Genovese, G., Rosema A. 2004. Application of Meteosat derived meteorological information for crop yield predictions in Europe. International Journal of Remote Sensing. 25 (23): 5389‐5401. Running S.W., Thornton P. E., Nemani R., Glassy J. M. 2000. Chapter 3: Global Terrestrial gross and net primary productivity from the Earth observing system. In: Sala O. E., Jackson R.B., Mooney H. A., Howarth R.W., editors. Methods in Ecosystem Science. New York: Springer Verlag. p 44 ‐ 57. Shao Y., Fan X., Liu H., Xiao, J., Ross, S., Brisco B., Brown R., Staples G. 2001. Rice monitoring and production estimation using multitemporal RADARSAT. Remote Sensing of Environment. 76: 310 – 325. Tao F., Yokozawa M., Zhqng Z., Xu Y., Hasashi Y. 2004. Remote sensing of crop production in China by production efficiency models: models comparisons, estimates and uncertainties. Ecological m odeling. 183: 385‐396. Yan L., Qifang L., Xia L., Shengding L., Guobin C., Sholin P. 2003. Towards an operational system for regional‐scale rice yield estimation using a time‐series of Radarsat ScanSAR images. International Journal of Remote Sensing. 24 (21) : 4207 – 4220.
11