Development of Agrometeorological Models for Estimation of Cotton Yield

DOI: 10.5958/2349-4433.2015.00006.9 Development of Agrometeorological Models for Estimation of Cotton Yield K K Gill and Kavita Bhatt School of Climate Change and Agricultural Meteorology Punjab Agricultural University, Ludhiana 141004 ( Punjab) ABSTRACT An attempt was made to predict American cotton (Gossypium barbadense) yield over Punjab region (Ludhiana, Bathinda and Ferozepur) by regression models. Three statistical models were developed for forecasting the yield of the American cotton using the yield data and weekly weather variable viz. maximum and minimum temperature, morning and evening relative humidity, sunshine hours, rainfall and number of rainy days. In the first model (Basic model) sensitive period for American cotton yield with respect to weather parameters were identified for different weather parameters by using correlation and selected windows were taken for further regression analysis. The second model is Modified model, where composite index was taken as one of the extra variable in multiple regressions. In the third model multiple regression analysis was done by using SPSS software. Regression equations were developed separately for all the three models and were used to predict the yield of American cotton. The historical weather data for the period of 1971-2009, 1978-2009 and 2001-2009 were used to develop forecast models for Ludhiana, Bathinda and Ferozepur, respectively. The recent three year meteorological data (2010-2012) was used to validate the models. For Ludhiana, among all the three models, Basic model explained up to 69 per cent variation, modified model explained 75 per cent and the highest i.e. 89 per cent variation was explained by SPSS model. For Bathinda district, basic model, modified model and SPSS model explained 50, 57 and 68 per cent variation, respectively. For Ferozepur, basic model explained up to 67 per cent, modified model explained 68 per cent and SPSS model explained 93 per cent variation in cotton yield. The results revealed that SPSS model fits better for all the three districts as far as American cotton yield is concerned. Key Words: American cotton, Correlation, Multiple regression, SPSS, Composite index, Yield forecasting INTRODUCTION Cotton is the world s most important fiber crop and the second most important oil seed crop. In Punjab, 4.80 lakh hectare area is used under cotton crop and its production is 21.0 lakh bales. The primary product of the cotton plant is lint that covers the seeds within the boll. Adequate soil temperature and moisture conditions at planting are necessary to ensure proper seed germination and crop emergence. The recommended soil temperature at seed depth should be above 18 C, to ensure healthy and uniform stands (Oosterhuis, 2001). Many factors, such as length of the growing season, climate (including solar radiation, temperature, light, wind, rainfall, dew), cultivar, availability of nutrients and soil moisture, pests and cultural practices affect cotton growth (Sawan, 2012). There are various management practices that should be followed to help mitigate some of the environmental risks associated with growing of cotton. Linear regression models (Wanjura and Barker, 1985; Muhidong, 1996; Pan, 2003) were built to explain effects of climatic factors on mature cotton fiber length and strength. In India, research in the area of crop-weather relations has been relatively very little. Some studies (Parthasarathy et al 1992; Kumar et al 2011) have used the simplest way of linear model wherein the meteorological factors are directly included * Corresponding Author s Email: kgill2002@gmail.com 27

in a linear fashion. In the present paper, we have used composite index comprised of different technological factors in the model to measure the impact of weather on crops. The present attempt is based on the correlation regression technique. The study involves seven weather parameters (maximum and minimum temperature, morning and evening relative humidity, sunshine hours, precipitation and number of rainy days on weekly basis) for developing the three different statistical models for predicting American cotton yield for different districts of Punjab. MATERIALS AND METHODS In the present study, the yearly production (q) and area (ha) under American cotton crop for the period 1972 to 2012 in respect of Ludhiana, Bathinda and Ferozepur was collected from Statistical Abstract of Punjab. For each year, the total production of American cotton for different districts was divided by the total area to calculate the cotton productivity. Long series weekly data of different weather elements (maximum temperature, minimum temperature, rainfall, sunshine hour, number of rainy days, morning and evening relative humidity) for the years 1972 to 2012 were collected from the meteorological observatory installed at School of Climate Change and Agricultural Meteorology, PAU, Ludhiana for Ludhiana district. The other two districts representing central plain zone i.e. Bathinda and Ferozepur, the weather data has been collected from IMD (India Meteorological Department), Chandigarh for the months covering the life cycle of the crop (14 th week to 42 nd standard meteorological week) except the harvesting period, since the forecast is to be given before harvesting. The correlation analysis was carried out by using Pearson correlation technique and the statistical model was developed using multiregression method. All the three models were developed from a data series of 34-39 years (1974 to 2012 for Ludhiana, 1978-2012 for Bathinda and 2001-2012 for Ferozepur district) and the models were verified with independent data for Gill and Bhatt the years from 2010 to 2012, outside their sampling series. The performance of the model was examined critically by computing percentage deviations of estimates and forecast yield figures. The basic model was developed by using weather parameters and taking into account the sensitive period window for cotton crop. The correlations were worked out for sensitive periods and then multiple regression equations were developed by using G-stat. The average reported crop-yield was taken as dependent variable with weather parameters as independent variables. Out of all the periods, the sensitive periods of statistical and phenological significance were selected in terms of standard meteorological weeks (SMWs) for regression analysis. Secondly, a modified model was developed by introducing composite index in the basic model keeping other independent variables constant. The development of modified model was intended to improve the accuracy of forecast of cotton yield, by superimposing the impact of agricultural technology in the form of linear time scale. The third model is based on analyzing regression using SPSS (Statistical Product and Service Solutions) software. Pearson s correlations between observed cotton yield and weather parameters and with combinations of weather parameters were computed. Sum of weather parameter and sum product of different weather parameter and correlation coefficient has been derived. Multiple regressions between dependent variable (Yield) and independent variables (time, sum and sum products for different weather parameters) were done using SPSS software. Regression equation was written using the regression formula. Regression equation for all the three models is given by: Y e = a 0 + Ó a i x i + Ó a j x j i = 1 Where, Y e = Estimated yield, kg/ha a 0 = Regression constant a i = Regression coefficients for meteorological predictor variables 28

Development of Agrometeorological Models x i = i th meteorological predictor variable i = 1, 2, n a j = Regression coefficients for technological trend variables. x j = j th technological trend variable RESULTS AND DISCUSSION For different districts of Punjab, where American cotton is grown, three different yield forecast equation have been developed using the three different models by using weekly data of weather variables. Out of all the periods, the sensitive periods of statistical and phenological significance were selected for different districts in terms of standard meteorological weeks (SMWs) for regression analysis. For Ludhiana district, effect of sensitive period on cotton yield was analysed and represented in table 1. The sensitive periods for American cotton crop represent square formation, flowering, boll formation and boll opening. Analysis of sensitive period for different districts (Table 1) showed maximum temperature and rainfall had negative effect on the cotton yield in all the districts. Model 1: In the basic model the weather data at critical periods were correlated with the yield and these correlations were used for regression analysis and the regression equations were developed. For Ludhiana district the regression expression is as follows Yield=(-703.90)-0.45Tmax(14-16)- 2.33Tmin(18-20)+62.57Tmin(29-30)- 4.64Rmax(15-16)+2.94Rmax(25-26)- 8.27Rmin(14-15)+6.38Rmin(21-22)- 3.21Rain(15-16)+50.68Rdays(15-17)- 26.51SSH(20-21) R 2 = 69 per cent The above analysis showed 69 per cent of variation in cotton yield due to weather parameters. The per cent error ranged between minimum -0.9 per cent to 15.2 per cent. The forecasted yield and per cent error of 34 years, based on above regression equation is given in Table 2 Regression equation developed for Bathinda district is as follows Yield=2191.57-17.93*Tmax(25-26)- 49.51*Tmin(17)+0.84*Rain(24-25)+5.00*Rdays(25)-89.09*Rdays(41) R 2 = 50 per cent The basic model explained 50 per cent variation in cotton yield due to weather parameters. The per cent error ranged between minimum -1.5 to 24.3 per cent. The forecasted yield and per cent error of 35 years, based on above regression equation is given in Table 3. For Ferozepur district, regression expression is as follows Yield=1720.04+6.19*Tmax(32-35)- 45.73*Tmin(32-35)-40.30*Rain(15- *16)+30.77*Rdays(25-26) R 2 = 66 per cent This model explained 66 per cent of variation in cotton yield. The percent error ranged between minimum -4.5 per cent to 15.7 per cent. The forecast yield and percent error of 12 years, based on above regression equation is given in Table 4. Model 2: In second model (Modified model), the weather parameters in critical periods along with technology trend variable were used through multiple regression analysis to obtain forecast model. Improved agricultural technology necessitated the need to modify the basic model Table 1. Sensitive periods and effect of weather variables on American cotton yield in Punjab. Districts Sensitive Period (SMWs) Stage of cotton crop Effect on cotton yield Maximum Temperature Ludhiana 21 and 22 Square formation -ve Bathinda 25 and 26 Flowering stage -ve Ferozepur 32 and 35 Ball formation -ve Rainfall Ludhiana 15 and 16 Vegetative stage -ve Bathinda 34 Ball formation -ve Ferozepur 22 and 23 Square formation -ve 29

by introducing composite index as an independent linear time scale. The regression expression for Ludhiana is as follows: Yield=(-926.42)-32.18*Tmax(1-3)- 31.61*Tmax(9-12)-16.19*Tmin(46-47)+19.89*Tmin(49-51)+85.20*Tmin(6-12)+5.94*Rain(9-10)-145.96*Rdays(8-9)+20.03*Rmax(51-2)-6.08*Rmin(1-3)+94.87*SSH(9-11)+0.28*Composite Index R 2 = 78 per cent The value of R 2 has increased to 78 per cent in the modified model. The error per cent for modified model ranged from -3.8 per cent to 15.5 per cent for the last 34 years. Regression expression with modified model For Bathinda is as follows: Yield=1886.3-11.02*Tmax(25-26)- 50.87*Tmin(17)+0.91*Rain(24-25)+0.44*Rdays(25)-98.10*Rdays(41)+4.57E- 02*Composite Index R 2 = 53 per cent The modified model showed increase in the value of R 2 (54%). The error per cent for modified model ranged from -1.0 to 21.4 per cent for the last 35 years. Regression equation developed for Ferozepur district is as follows: Yield=1741.89+5.99*Tmax(32-35)- 28.89*Tmin(32-35)-79.07*Rain(15-16)+0.46*Rdays(25-26)-4.30*Composite Index R 2 = 75 per cent The modified model showed increase in the value of R 2 to 75 per cent. The error per cent for modified model ranged between -2.5 to 15.8 for the last 12 years. Model 3: Regression using SPSS software: The multi-regression analysis using SPSS was employed for the estimation of American cotton yield in different districts of Punjab. The regression expression for Ludhiana is as follows: Yield =856.78+6.12Time+27.13Z11-0.04Z150+0.01Z350+0.60Z471 R 2 = 89 per cent Here, Z11 is the sum product of maximum temperature Z150 is the sum product of maximum temperature* rainfall 30 Gill and Bhatt Z350 is the sum product of morning relative humidity*rainfall Z471 is evening relative humidity*sunshine hours The regression equation showed that time, maximum temperature, combination of maximum temperature and rainfall, combination of morning relative humidity and rainfall and combination of evening relative humidity and sunshine hours plays an important role on American cotton yield in Ludhiana area. The per cent error ranged between minimum -1.3 per cent in year 1998 and maximum 10.8 per cent in year 1975. The forecasted yield and per cent error based on above regression equation is given in Table 4. The value of R 2 is 89 per cent indicated that weather variables were able to explain 89 per cent of variation in the American cotton yield at Ludhiana region. The regression expression is as follows for Bathinda district: Regression equation =367.36+6.17 Time+ 3.10Z31 R 2 =68 per cent Here, Z31 is the sum product of rainfall The regression equation showed that time and rainfall plays an important role on American cotton yield in Bathinda area. The per cent error ranged between -15.5 per cent in year 1994 and maximum 15.6 per cent in year 2000. The forecasted yield and per cent error based on above regression equation is given in Table 5. The value of R 2 is 68 per cent indicates that weather variables are able to explain 68 per cent of variation in the American cotton yield at Bathinda region. For Ferozepur, regression expression is as follows: Regression equation = 554.51+ 0.79Z141+ 0.69Z341 R 2 = 93 per cent Here, Z141 is maximum temperature* rainy days Z341 is rainfall * rainy days The regression equation showed that time and combination of maximum temperature and rainy days and the combination of rainfall and rainy

days plays an important role on American cotton yield in Ferozepur area. The per cent error ranged between -1.7 per cent in the year 2002 and maximum of 9.6 per cent in the year 2011. The forecasted yield and per cent error based on above regression equation is given in Table 6. The value of R 2 is 93 per cent indicating that weather variables are able to explain 93 per cent of variation in the American cotton yield in Ferozepur region. Development of Agrometeorological Models Validation of the models The recent three years i.e. 2010, 2011 and 2012 were taken for validation of the three models. For Ludhiana district, the forecast yield of three years was compared with actual yield obtained that year to calculate the error percentage. The results showed (Table 2) that in the year 2010, all the three models (model 1, 2 and 3) showed less forecasted yield than the actual yield, but in year 2011, model 2 and 3 showed higher yield than the actual. In the year 2012, model 1 and 3 showed less forecasted yield than the actual while model 2 showed high forecasted yield. The comparisons between the actual and forecasted yield for the year 2010, 2011 and 2012 by the three different models for American cotton crop is summarized as figure 1. The error per cent for model 1 in year 2010, 2011 and 2012 were 3.0, 1.9 and 2.7 per cent, respectively. The error percent of model 2 was more than the basic model for all the three years and is 6.1, -4.5 and -4.0 per cent for the years 2010, 2011 and 2012, respectively. The error for model 3 was 8.3 per cent (2010), -10.0 per cent Figure 2. Comparison between the actual and forecasted yield (by three different models) of American cotton for the year 2010, 2011 and 2012 for Bhatinda district (2011) and 7.8 per cent (2012). For Bathinda district, the forecast yield for three different years (2010, 2011 and 2012) was compared with actual yield to calculate the error percentage. The results showed (Table 3) that in the year 2010 and 2012, all the three models showed lesser yield than the actual yield. The comparisons between the actual and forecast yield for the year 2010, 2011 and 2012 by the three different models for mustard crop is summarized as figure 2. The error percent for model 1 in year 2010 was 18.6 per cent, in 2011 was 21.2 per cent and in 2012 was 12.5 per cent in the year 2012. The error percent of model 2 for all the three years and was 9.6, 13.1 and 5.4 per cent for the year 2010, 2011 and 2012, respectively. The error percent for model 3 was 8.8, 10.4 and 3.9 per cent for the year 2010, 2011 and 2012, respectively. For Ferozepur district, the forecast yield of three different years (2010, 2011 and 2012) was compared with actual yield to calculate the error percentage. The results showed (Table 4) that in the year 2010, model 1 and 3 showed higher yield than the actual yield while in 2011, all the three models showed less forecast yield than the actual. In the year 2012, model 1 and 2 indicated lesser yield than the actual while model 3 was predicting higher yield. The comparisons between the actual and forecast yield for the year 2010, 2011 and 2012 by the three different models for cotton crop is summarized as figure 3. Figure 1. Comparison between the actual and forecasted yield (by three different models) of American cotton for the year 2010, 2011 and 2012 for Ludhiana district 31

Gill and Bhatt Table 2. Forecasted yield and error per cent of three different models for Ludhiana. Year Actual yield Model 1 Model 2 Model 3 (kg /ha) Forecasted % Error Forecasted % Error Forecasted % Error yield(kg /ha) yield(kg /ha) yield(kg /ha) 1970-71 333 295 11.3 370-11.0 372-11.8 1972-73 349 354-1.3 364-4.2 382-9.6 1973-74 383 430-12.3 401-4.7 405-5.7 1974-75 360 367-1.9 342 5.1 321 10.8 1975-76 340 382-12.3 315 7.4 351-3.1 1976-77 297 303-2.2 314-5.6 288 3.0 1977-78 187 208-11.2 184 1.9 182 2.6 1978-79 252 267-6.1 274-8.7 229 9.3 1979-80 306 268 12.6 291 5.0 298 2.7 1980-81 345 400-16.0 337 2.2 342 0.8 1981-82 262 229 12.8 276-5.4 282-7.7 1982-83 152 146 4.0 146 3.7 142 6.8 1983-84 108 111-2.9 97 10.2 100 7.1 1984-85 315 326-3.6 347-10.0 305 3.0 1985-86 452 485-7.3 431 4.6 448 0.8 1986-87 540 504 6.8 524 3.0 619-14.6 1987-88 456 465-2.0 455 0.2 536-17.6 1988-89 378 366 3.3 354 6.2 397-5.0 1989-90 425 466-9.8 424 0.2 424 0.1 1090-91 245 208 15.2 256-4.3 262-6.7 1991-92 429 394 8.2 362 15.5 404 5.8 1992-93 484 452 6.6 426 12.1 552-14.1 1994-95 380 417-9.7 408-7.3 418-10.1 1995-96 416 434-4.4 432-3.8 451-8.4 1996-97 467 411 11.9 454 2.9 456 2.4 1997-98 213 188 11.9 208 2.2 216-1.3 1998-99 179 161 10.2 207-15.6 208-16.2 2005-06 748 755-0.9 715 4.4 771-3.1 2006-07 763 751 1.6 752 1.4 854-12.0 2007-08 668 592 11.4 595 11.0 755-13.0 2008-09 743 737 0.8 711 4.3 764-2.8 2009-10 673 653 3.0 632 6.1 617 8.3 2010-11 646 634 1.9 675-4.5 711-10.0 2011-12 537 523 2.7 559-4.0 495 7.8 32

Development of Agrometeorological Models Table 3. Forecasted yield and error per cent of three different models for Bathinda. Year Actual yield Model 1 Model 2 Model 3 (kg /ha) Forecasted % Error Forecasted % Error Forecasted % Error yield(kg /ha) yield(kg /ha) yield(kg /ha) 1977-78 392 460-17.3 386 1.6 388 1.1 1978-79 403 409-1.5 448-11.1 388 3.7 1979-80 337 307 9.0 279 17.3 325 3.5 1980-81 344 341 0.8 319 7.3 295 14.2 1981-82 359 344 4.2 342 4.8 407-13.4 1982-83 315 252 19.9 324-2.7 357-13.4 1983-84 209 174 16.8 252-20.4 205 1.7 1984-85 453 468-3.2 458-1.0 420 7.4 1985-86 435 350 19.5 435 0.0 392 9.8 1986-87 536 523 2.4 498 7.1 607-13.3 1987-88 503 494 1.8 491 2.3 544-8.2 1988-89 462 353 23.7 433 6.4 405 12.3 1989-90 588 550 6.5 548 6.9 512 12.9 1090-91 489 468 4.4 463 5.3 486 0.6 1991-92 647 604 6.6 598 7.6 555 14.3 1992-93 577 465 19.5 471 18.4 497 13.8 1993-94 495 532-7.4 520-5.0 572-15.5 1994-95 521 541-3.8 589-13.1 470 9.9 1995-96 464 517-11.4 541-16.7 527-13.5 1996-97 471 578-22.7 570-21.1 535-13.6 1997-98 236 246-4.3 265-12.5 208 12.0 1998-99 173 176-1.6 136 21.4 168 3.0 1999-00 302 340-12.5 280 7.2 255 15.6 2000-01 382 289 24.3 334 12.6 366 4.2 2001-02 415 450-8.4 430-3.6 403 3.0 2002-03 480 539-12.3 492-2.5 551-14.8 2003-04 560 437 22.0 461 17.7 528 5.7 2004-05 727 609 16.3 617 15.1 646 11.2 2005-06 755 606 19.7 665 12.0 692 8.3 2006-07 790 696 11.9 733 7.2 706 10.6 2007-08 685 586 14.5 616 10.1 630 8.0 2008-09 800 611 23.6 701 12.4 715 10.7 2009-10 750 610 18.6 678 9.6 684 8.8 2010-11 705 555 21.2 613 13.1 632 10.4 2011-12 629 550 12.5 595 5.4 605 3.9 33

Gill and Bhatt Table 4. Forecasted yield and error per cent of three different models for Ferozepur. Year Actual yield Model 1 Model 2 Model 3 (kg /ha) Forecasted % Error Forecasted % Error Forecasted % Error yield(kg /ha) yield(kg /ha) yield(kg /ha) 2000-01 505 528-4.5 518-2.5 504 0.1 2001-02 389 361 7.3 363 6.6 395-1.7 2002-03 425 448-5.5 382 10.1 434-2.1 2003-04 644 612 5.0 604 6.2 667-3.5 2004-05 704 659 6.4 658 6.5 669 4.9 2005-06 750 632 15.7 656 12.6 733 2.3 2006-07 733 676 7.8 662 9.7 670 8.6 2007-08 577 681-18.0 680-17.8 587-1.8 2008-09 688 721-4.8 726-5.5 755-9.8 2009-10 564 620-9.9 508 10.0 607-7.7 2010-11 560 524 6.4 471 15.8 506 9.6 2011-12 464 425 8.5 407 12.2 509-9.6 Figure 3. Comparison between the actual and forecasted yield (by three different models) of American cotton for the year 2010, 2011 and 2012 for Ferozepur district The error per cent for model 1 during 2010, 2011 and 2012 was -9.9, 6.4 and 8.5, respectively whereas for model 3,it was -7.7, 9.6 and -9.6 per cent during the year 2010, 2011 and 2012, respectively. The error per cent of model 2 was higher than the basic model which was found to be 10.0, 15.8 and 12.2 per cent during all the three years. cotton harvest. The district government authorities also can make use of the forecast model developed using weather indices, in this study, for obtaining accurate pre-harvest estimates of American cotton crop for the upcoming years. Till the final production of crops becomes known, decisions have to be made on the basis of informed predictions or scientific forecasts. The main beneficiaries are farmers (decide their procurement prices), traders, exporters and importers. REFERENCES Kumar S, Raju B M K, Rao C A R, Kareemulla K and Venkateswarlu B (2011). Sensitivity of yields of major rainfed crops to climate change. Indian Journal Agricultural Economics 66 (3): 340-352. Muhidong J (1996). A Cotton Fiber Quality Model. Agricultural and Biological Engineering. Mississippi State University, Mississippi. Oosterhuis D M (2001). Development of a Cotton Plant. In: Seagull, R. and P. Alspaugh (eds) Cotton Fiber Development and Processing, an illustrated overview. International Textile Center, Texas Tech University, Lubbock, TX. Pan X B (2003). Crop model Principle. China Meteorological Press, Beijing (in Chinese). Parthasarathy B, Kumar, K R and Munot A A (1992). Forecasting of rainy-season food grain production based on monsoon rainfall. Indian Journal AgriculturalScience, 62(1): 1-8. Sawan Z M (2012). Cotton (Gossypium barbadense) production as affected by climatic factors and soil moisture status., Acta Ecologica Sinica.32 (3): 123 137. Wanjura D F and Barker G L (1985). Cotton lint yield accumulation rate and quality development. Field Crops Res 10: 205 218. CONCLUSION Using the forecast models, pre-harvest estimates of American cotton yield for Ludhiana, Bathinda and Ferozepur districts can be computed successfully in advance before the actual harvest. The error per cent of these models remained up to 25 per cent. As the data used for developing these models is of high degree of accuracy, its reliability is also high. The modified model 2 is having close proximity with the actual yield and it can be used by district government prior to Received on 25/11/2014 Accepted on 16/12/2014 34