EARLY PREDICTION OF CROP PRODUCTIONS USING DROUGHT INDICES AT DIFFERENT TIME SCALES AND REMOTE SENSING DATA: APPLICATION IN THE EBRO VALLEY (NORTH-EAST SPAIN)
Sergio M. Vicente-Serrano1,2*, José M. Cuadrat-Prats2 and Alfredo Romo3
1 Centre d'Etudes Spatiales de la Biosphère (CESBIO), 18 avenue. Edouard Belin, bpi 2801, 31401 Toulouse cedex 9. France.
2Departamento de Geografía. Universidad de Zaragoza. C/ Pedro Cerbuna 12. 50009. Zaragoza. Spain.
3Laboratorio de Teledetección. Departamento de Física Aplicada I. Facultad de Ciencias. Universidad de Valladolid. 47071. Valladolid. España.
* e-mail:
Abstract. In this letter are shown the results of early crop prediction from combined use of AVHRR-NDVI data and drought indices at different time scales. The study was carried out in an agricultural municipality located in the Middle Ebro valley, one of the most arid regions in Europe. The methodology proposed here has allowed the prediction of wheat and barley productions in February, four months before harvest. Moreover, the predictive models created have explained 88% and 82% of the temporal variability of wheat and barley productions, respectively. This procedure could be very useful to manage crop productions at a municipal level. Moreover, insurance companies could take advantage of the early prediction of crop loses, which are very frequent in this area affected by droughts.
Key-words: Crop prediction, NDVI, NOAA-AVHRR, Standardized Precipitation Index, Mediterranean region, Ebro valley
1. Introduction
Remote sensing has been widely used for early prediction of crop productions (Curran 1987) on the basis of predictive empirical models. Those models have been derived from satellite sensor imagery and crop production data (Tucker and Sellers 1986; Basnyat et al. 2004). Different authors have shown the high reliability of early crop predictions obtained by means of NDVI-AVHRR data (i.e., Rasmussen 1997; Kalubarme et al. 2003) or other satellite sensor imagery such as Landsat-TM and ETM+ (Thenkabain, 2003; Baez-González et al. 2005), IRS (Ray and Pokharna, 1999) and SPOT (Moulin et al. 1998).
Crop prediction models have been applied in large and homogeneous agricultural regions of Africa, USA and South America (i.e., Seiler et al. 2000; Wannebo and Rosenzweig 2003; Mkhabela et al. 2005). Few studies have analysed the usefulness of NDVI-AVHRR data to predict crop productions in small, non-homogeneous areas (i.e. at municipal level). Moreover, few works have predicted crop productions by means of remotely sensed data in semi-arid Mediterranean regions (i.e., Quarmby et al. 1993; Royo et al. 2003). Drought indices have been also used for crop prediction with good results (Karl and Koscielny 1982; Yamoah et al. 2000; Quiring and Papakryiakou 2003).
These two approaches (remote sensing data and drought indices) have been useful for crop prediction. However, until now both approaches have been used independently. In this letter, the combined use of NDVI data and drought indices has been analysed to predict crop productions at fine spatial resolution. The analysis has been done in the northernmost semi-arid region of Europe, the Centre of the Ebro valley. In this area droughts are very frequent and severe (Vicente-Serrano and Beguería 2003; Vicente-Serrano et al. 2004), and crop productions have an important inter-annual variability (Austin et al. 1998). Therefore, early prediction of crop production would be very useful in this region.
2. Methodology
Data on wheat and barley production (proportion between sowed and harvested), in Monegrillo municipality, between 1987 and 2000 have been used (Figure 1). In this area, cereal is usually sowed in November and harvested during the last two weeks of June. For details about the production data refer to Austin et al. (1998).
[Insert Figure 1 about here]
Full resolution (1 km2) NOAA-AVHRR images (bands 1 and 2) have also been used during the same period. The images were processed by the LATUV (Remote Sensing Laboratory of Valladolid University). Processing consisted on: a) atmospheric, radiometric and geometric corrections (Illera et al. 1996 and 1997), b) NDVI calculation, and c) creation of ten-day composites in order to reduce possible cloud and remaining atmospheric disturbances (Holben 1986). A temporal filter was applied to the ten-day composites NDVI in order to reduce even more residual atmospheric effects. The temporal filter developed by Quarmby et al. (1993) was applied. This substitutes abnormally low NDVI values according to equation (1):
NDVInew (n)=MAX{NDVI(n), (NDVI(n-1)+NDVI(n+1))/2} (1)
where NDVInew (n) is the definitive NDVI value for a ten-day composite of each pixel. High NDVI values are not changed because they are less affected by atmospheric effects and usually are collected on optimal conditions (near nadir).
The pixels corresponding to cereal areas in the municipality were identified using the CORINE land cover map (CLC, 1990). Every ten days, a mean value of NDVI was calculated as the average NDVI over these pixels (88 in total).
Drought was evaluated by means of the Standardized Precipitation Index (SPI, McKee et al. 1993). This index has several advantages over others. Calculation of the SPI is easier than on more complex indices. Moreover, the SPI is comparable in both time and space, and is not affected by geographical or topographical differences. The SPI allows for the determination of duration, magnitude and intensity of droughts and only requires precipitation data (Wu et al., 2005). Its main advantages are that it can be calculated for several time scales (McKee et al., 1995) and identifies various drought types: hydrological, agricultural or environmental.
The SPI has been extensively used for drought analysis at country and continental level in Europe (Lloyd-Hughes and Saunders, 2002), South Africa (Rouault and Richard, 2003), Greece (Tsakiris and Vangelis, 2004) and Spain (Vicente-Serrano et al., 2004b).
Data from the Monegrillo weather-station were used for the calculation of the SPI. SPI was obtained each ten days at four time scales (1, 3, 6 and 12 months) (Hayes et al. 1999) using the Pearson III distribution and the L-moment method, as recommended by Vicente-Serrano (2005) for the same region. Negative values of SPI indicate dry conditions. Moderate, severe and extreme droughts correspond to SPI values under -0.84, -1.28 and –1.65, respectively. The complete procedure using the Pearson III distribution and the L-moment method can be consulted in detail in Vicente-Serrano (2005).
The correlations among cereal production, NDVI and SPI were calculated at the four time scales every ten days between January and April. Linear multiple regression analysis was used to predict barley and wheat productions. The accuracy of the predictions was evaluated by means of the following error statistics: Root Mean Square Error (RMSE), Mean Absolute Error (MAE) and Willmot’s D (Willmot 1982).
3. Results
Table 1 shows correlations among wheat and barley productions, NDVI and SPI at the four time scales, for each ten days period, from January to April. Strong correlations have been found considering the 1-months and 3-months SPI in the different ten-day periods of February, both for barley and wheat production. Moreover, for some ten-day periods of April, correlations higher than 0.8 have been found with the 6-months SPI. Correlations with the NDVI were also significant the first two ten-day periods of February, the second and third ten days of March and the first ten days of April.
[Insert Table 1 about here]
In view of the high correlations found among the wheat and barley productions, the 1-month SPI and the NDVI in the first ten days of February, a predictive model for the cereal production 4 months before harvest was calculated. Stepwise regression analysis was used for this purpose to avoid colinearity between variables (Hair et al. 1998). The highest correlations between NDVI and crop yield were obtained in March, corresponding to flowering and maximum leaf area index (LAI). Nevertheless, also high correlations were obtained in February. Therefore, only variables previous to the second week of February were included in the models. Table 2 shows the results for barley and wheat production. Included variables were the 1-month SPI and the NDVI in the first ten days of February, being both of them significant at a=0.05. The models have predicted 88% and 82% of the temporal variability of wheat and barley production, respectively. SPI has had a higher explanatory role than NDVI (higher Beta standardised coefficients) but the inclusion of NDVI has increased significantly the accuracy of the final prediction. Although water availability is the main constraint for crop development, the NDVI can summarise other factors related to the crop management and development that also affect inter-annual variability of crop productions.
[Insert Table 2 about here]
Figure 2 shows the evolution of observed and predicted productions of barley and wheat between 1987 and 2000 in Monegrillo. Little differences have been found between both series.
[Insert Figure 2 about here]
Table 3 shows the error statistics obtained from the comparison between observed and predicted production. Mean absolute error was very small in comparison with the variable range (Barley = 0-16.6, Wheat = 0-13.1). The high D values also have indicated little differences between observed and predicted data, which has demonstrated the predictive capabilities of the statistical models. The error statistics obtained from a simple regression model, which only considers the SPI in the first ten-days of February, have also been reported in table 3. It can be seen that the performance is lower compared to the model including also the NDVI.
[Insert Table 3 about here]
4. Conclusions
Cereals are very important for the economy of large rural areas in Spain. The inter-annual variability of crop production is very high due to aridity and frequent droughts (Austin et al. 1998). This letter has shown the high potential of the combined use of high temporal resolution remote sensing images and drought indices to predict cereal productions four months prior harvest in a semi-arid region of North-East Spain. The results obtained here only correspond to a municipality and further analysis are needed in other semiarid regions to confirm the potential of this method. Nevertheless, the combined use of remote sensing and drought indices seems to be a very useful approach for early crop prediction, having a good potential to reduce crop uncertainty to farmers and insurance companies.
Acknowledgements
This work has been supported by the projects: BSO2002-02743 and REN2003-07453 financed by the Spanish Comission of Science and Technology (CICYT) and FEDER, and “Programa de grupos de investigación consolidados” (grupo Clima, Cambio Global y Sistemas Naturales, BOA 48 of 20-04-2005), financed by the Aragón Government. We want to thank Pascual Cano for the readiness of the crop data used in this work. Research of the first author was supported by a postdoctoral fellowship by the Ministerio de Educación y Ciencia (Spain).
References
Austin, R.B., Cantero-Martínez, C., Arrúe, J.L., Playán, E. and Cano-Marcellán, P., 1998, Yield-rainfall relationships in cereal cropping systems in the Ebro river valley of Spain. European Journal of Agronomy, 8, 239-248.
Baez-Gonzalez, A.D., Kiniry, J.R., Maas, S.J., Tiscareno, M., Macias, J., Mendoza, J.L., Richardson, C.W., Salinas, J. and Manjarrez, J.R., (2005): Large-area maize yield forecasting using leaf area index based yield model. Agronomy Journal, 97, 418-425.
Basnyat, P., McConkey, B., Lafond, G.R., Moulin, A. and Pelcat, Y., 2004, Optimal time for remote sensing to relate to crop grain yield on the Canadian prairies. Canadian Journal of Plant Science, 84, 97-103.
CLC, 1990, Proyecto Ocupación del suelo del Programa CORINE: Definiciones de la clasificación Española. Comisión Técnica del Proyecto Ocupación del Suelo de España. Instituto Geográfico Nacional.
Curran, P.J., 1987, Remote sensing in agriculture: an introductory review. Journal of Geography, 86, 147-156.
Hair, J.F., Anderson, R.E., Tatham, R.L. and Black, W.C., 1998, Multivariate data analysis. Prentice Hall International, Inc. 799 pp.
Hayes, M., Wilhite, D.A., Svoboda, M. and Vanyarkho, O., 1999, Monitoring the 1996 drought using the Standardized Precipitation Index. Bulletin of the American Meteorological Society, 80, 429-438.
Holben, B., 1986, Characteristics of maximum value composite images from temporal AVHRR data. International Journal of Remote Sensing, 6, 1271-1328.
Illera, P., Delgado, J.A. and Calle, A., 1996, A navigation algorithm for satellite images. International Journal of Remote Sensing, 17, 577-588.
Illera, P., El-Kaouri, A., and Vegas, A., 1997, Influencia de las condiciones atmosféricas sobre el NDVI-AVHRR. In Teledetección. Usos y aplicaciones. Universidad de Valladolid: 415-422. Valladolid. Spain.
Karl, T.R. and Koscielny, A.J., 1982, Drought in the United States: 1895-1981. Journal of Climatology, 2, 313-329.
Kalubarme, M.H., Potdar, M.B., Manjunath, K.R., Mahey, R.K. and Siddhu, S.S., 2003, Growth profile based crop yield models: a case of large area wheat yield modelling and its extendibility using atmospheric corrected NOAA AVHRR data. International Journal of Remote Sensing, 24, 2037-2054.
Lloyd-Hughes, B. and Saunders, M.A., 2002, A drought climatology for Europe, International Journal of Climatology 22, 1571- 1592.
McKee, T.B.N., Doesken, J. and Kleist, J., 1993, The relationship of drought frecuency and duration to time scales. Eight Conf. On Applied Climatology. Anaheim, CA, Amer. Meteor. Soc. 179-184.
McKee, T.B.N., Doesken, J. and Kleist, J., 1995, Drought monitoring with multiple time scales, Ninth. Conf. On Applied Climatology, Dallas, TX, Amer. Meteor. Soc., 233-236.
Moulin, S., Bondeau, A. and Delecolle, R., 1998, Combining agricultural crop models and satellite observations: from field to regional scales. International Journal of Remote Sensing, 19, 1021-1036.
Mkhabela, M.S., Mkhabela, M.S. and Mashinini, N.N., 2005, Early maize yield forecasting in the four agro-ecological regions of Swaziland using NDVI data derived from NOAAs-AVHRR. Agricultural and Forest Meteorology, 129, 1-9.
Quarmby, N.A., Milnes, M., Hindle, T.L. and Silleos, N., 1993, The use of multitemporal NDVI measurements from AVHRR data for crop yield estimation and prediction. International Journal of Remote Sensing, 14, 199-210.
Quiring, S.M. and Papakryiakou, T.N., 2003, An evaluation of agricultural drought indices for the Canadian praires. Agricultural and Forest Meteorology, 118, 49-62.
Rasmussen, M.S., 1997, Operational yield forecast using AVHRR NDVI data: reduction of environmental and inter-annual variability. International Journal of Remote Sensing, 18, 1059-1077.
Ray, S.S. and Pokharna, S.S., 1999, Cotton yield estimation using agrometeorological model and satellite-derived spectral profile. International Journal of Remote Sensing, 20, 2693-2702.
Royo, C., Aparicio, N., Villegas, D., Casadesus, J., Monneveux, P. and Araus, J.L., 2003, Usefulness of spectral reflectance indices as durum wheat yield predictors under contrasting Mediterranean conditions. International Journal of Remote Sensing, 24, 4403-4419.