Analysis Of The Relationship Between The Rice Yield And Soil And Weather In Fujian Province Based On Neural Network

Rice production is affected by sets of varietals and environmental parameters, includinggenetic characteristics, soil, weather and cultivational management. Rice grain yield for agiven cultivar is mainly dependent upon local weather conditions such as sunshine time,solar radiation and temperature, when plants are grown on the conditions of amplenutrients and water. Environment factors affect plant growth and development, and thefluctuations of environment and occurrences of climatic extremes particularly at criticalcrop growth stages may reduce yield significantly. The concern on past, present and futureweather aberrations, climate trends, soil reaction and their effects on agriculture hascontinued to stimulate research on the analysis of environment factors in rice growth.The variation in rice production along spatial and temporal gradients would beattributable to different climates and soils because they are the mainly factors of theenvironment factors. Also the different climatic and soil conditions are mostly associatedwith either cropping season or crop yield in the same year for a particular area. There isdistinct difference in climate between north and south, coastal and inland regions, andvalleys and mountains in Fujian province. Fujian is a mountainous region, about 50% of theland is hill, 70% of the rice was planted in mountain farm. However, a clear understandingof the vulnerability of food crops as well as the agronomic impacts of climate and soilvariability in mountainous area such as Fujian province enable one to implement adaptivestrategies to mitigate its negative effects and make a better yield prediction.In recent years, crop growth models have become increasingly important as majorcomponents of agriculture-related decision-support systems. Regression or correlationanalyses are generally used to characterize the statistical relationship between controlledvariables and crop yield. Technologically, empirical crop growth models are relativelysimple to build or develop, but these models cannot take account of temporal changes incrop yields without long-term field experiments. Empirical crop growth models attempt todetermine functional relationships between crop yield and other factors using either anexisting or a specially designed agronomic experiment. Furthermore, the derived functionalequation is locally specific, and it is thus difficult to extrapolate to other areas unlessenvironmental conditions are similar. The main limitation of traditional regression-basedempirical models is the lack of non-linear modeling ability, which is apparent in cropresponses to agro-ecological conditions. This may be the case particularly when variousland management practices are applied under different scenarios. Some adaptive andnon-parametric models have been recently introduced in environmental science forpredictive purposes. Artificial neural network (ANN) models are a powerful empiricalmodeling approach and yet relatively simple compared to mechanistic models. It is feltANN models offer a more versatile empirical modeling approach in comparison to the linearregression methods used in rice yield since the rice yield is non-linear and autoregressive innature. Because ANN allow an illustration of complex and non-linear relationships withoutrigorous assumptions regarding the distribution of samples. The method is gainingpopularity for research areas where there is little or incomplete understanding of theproblem to be solved, but where training data are available.Historical (1993-1999, 2000-2003) Fujian rice yield data from the Hybrid VarietyPerformance trials Fujian Agricultural Administration were accessed. The rice dataincluded 16 locations. Each location-specific soil properties (soil water capacity,organic-matter content, hydrolytic-N, quick-acting P, quick-acting K, bulk density, porosity,pH and CaCO₃ content) and the weather factors (total sunshine time, total solar radiation,total wind speed, temperature summation and rainfall data from February to November and total sunshine time from July to August) were obtained from whether station in eachlocation. The correlations between environment factors (climate, soil) and yield were studyin order to analyze the interaction between genotype and environment. The main factorsinfluencing yield were studied by principal component analysis. With stepwise regressionanalysis, functions about the environment factors and yield were derived and thecoefficient of determination (r) were reached. The yield predicted model of rice in Fujianwere developed with BP, CBP and GBP in Matlab 7.0 environment. The effectiveness ofmultiple linear regression models to ANN models were also compare in this paper. Themajor achievements are as follows:1. The structure of GEI in rice trails was stable in general. The GEI caused by a singletest site also maintains a relatively stable value. But the GEI was different among sites.The GEI of the Zhangping, Youxi, Jiangou, Wuyishan, Zhangpu, Fuzhou and Fuding wasmore than 10% respectively, and Yongding, Liancheng, Ninghua, Jiangle, Shaowu,Pucheng, Quanzhou, Putian and Ningde was less than 10% respectively.2. There was a positive correlation between soil properties (organic-matter content,hydrolytic-N, quick-acting P, quick-acting K, bulk density, porosity, and CaCO₃ content)and mean yield of test site, but there was an inverse relationship between soil watercapacity, pH and mean yield. The main factors influencing rice yield were soil fertility,yield characters, soil management and soil acidity. The coefficient of determination (r) ofthe regression function reached 0.9080. Comparing the BP, CBP, GBP and regressionmodel, the best predicted model is GBP, next is CBP and BP, the regression model is theworst.3. There was a positive correlation between climate factors (total sunshine time, totalsolar radiation and temperature summation from February to November and total sunshinetime from July to August) and mean yield of test site, but there was an inverse relationshipbetween total wind speed and total rainfall from February to November and mean yield.The main factors influencing rice yield were heat, kernel, soil management and wind andrain. The coefficient of determination (r) of the regression function reached 0.9574.Comparing the BP, CBP, GBP and regression model, the best predicted model is CBP,next is GBP and BP, the regression model is the worst.4. There was a positive correlation between soil properties (organic-matter content,hydrolytic-N, quick-acting P, quick-acting K, bulk density, porosity, CaCO₃ content, totalsunshine time, total solar radiation and temperature summation from February toNovember and total sunshine time from July to August) and mean yield of test site, butthere was an inverse relationship between total wind speed and total rainfall from Februaryto November, pH and mean yield. The main factors influencing rice yield were soilcharacters, heat, kernel, wind and rain, and P and K. The coefficient of determination (r) ofthe regression function reached 0.9756. Comparing the BP, CBP, GBP and regressionmodel, the best predicted model is BP, next is GBP and CBP, the regression model is theworst.5. Adjustment of ANN parameters included the number of hidden nodes, initializenetwork weight and optimizational network weight. The adjusted adjusted hidden nodeswas 9. The Genetic Algorithms (GA) and Chaos method also applied to optimize weighttraining of the Back Propagation neural network(BP). The better BP(CGBP) wasdeveloped to predict the rice yield in Fujian. Comparing regression model and BP, theerror of CGBP was the least. CGBP predicted yield more accurately than BP andregression model.Rice yield shows more complex, non-linear dynamics among yield responses andsoil-management inputs. The models reported here are appropriate for predicting rice yieldsin Fujian province in China for average conditions. With additional information of thecropping system, crop management should broaden the usefulness, and possibly increasethe predictive accuracy of ANN-based yield prediction in mountains area.