Study On A Support Vector Machine Based Open Crop Model (Sbocm)

Crop model is a series of methods to describe crop's growth by mathematical ways. Two distinct modeling ways-statistical modeling and simulation modeling could be identified, and the latter was thought to be the primary way in the past years. In this article, we firstly reviewed domestic and abroad evolution of crop models research and application since1960s, with great honors to the success during those years. Then the article discussed the used models'deficiencies, which were not suitable to the large areal prediction, such as the simulation results and real practice not consistent, too complex operation and so on. The article tried to analyze the reasons and advise that statistical modeling had its own advantages in areal simulation. Because of such a standpoint, this article used support vector machine (SVM) to build a new rice's development and yield prediction system-SBOCM, which was thought to be applied for large areal prediction.Based on the principle that data used for modeling should be applied by multiple scaling, this article designed an'open reading frame'to build input vectors, by using those data such as1:1,000,000Chinese soil database released by Chinese Academy of Sciences, daily meteorological data and basic information of observation stations both of which were released by China Meteorological Administration (CMA). While those record of different rice development stages and rice yield from each observation station were used to play as the output vectors. By using the'open reading frame' principle component analysis was chosen to play as a variable filter and the length of 'open reading frame'was7-day (development modeling) and11-day (yield modeling) restricted by the sample size. Because of3different rice cultivations—middle season rice, early rice and late rice, models should be build by different ways.Confirming the kernel function with its hyperparameters and optimizing the penalty coefficient were the two major contents during SVM training. This article tried to search the optimal parameters for each5development stages,3development points and3cultivations, by using K-fold crossvalidation. When modeling, SVM classification (SVC) was used to build each development models and2different samples were designed for SVC training. K-fold crossvalidation showed that models had high sensitivities trained by the first sample while those had a balance between sensitivity and false positivity, with Fl value among0.7to0.8. Main errors of daily simulation focused in5days around the development event point. SVM regression was used to build each yield prediction models at each tillering stage, heading date and milky stage and results showed that the'Heading date+milky stage' sample was the optimal sample scenario. The relative errors were18.3%,8.5%and10.2%for the3different cultivations. In the end, this article analyzed the defects during the SMV modeling, and advised a SBOCM system which was thought to be suitable to large areal simulation research. This system could be further optimized by combing more social and economic factors in future.