基于BP神经网络的宁夏耕地资源动态变化及预测

Dynamic Change and Prediction of Cultivated Land Resource in Ningxia Based on BP Neural Network

利用2000~2014年统计数据,分析了宁夏近15 a来的耕地动态变化特征,采用主成分分析法提取影响耕地变化的主要驱动因子,将驱动因子作为输入数据,构建了预测耕地资源变化趋势的BP神经网络,并开展预测研究。结果表明:宁夏耕地资源变化经历了增加—迅速减少—波动增长—迅速增长4个阶段,人均耕地经历了迅速减少—缓慢减少—迅速增加3个阶段;其中,2002~2004年和2010~2012年的土地利用动态度K<0,2003年动态度最小(K=-7.673%),2005~2010年和2012~2014年的土地利用动态度K>0,2014年动态度最大(K=11.424%);宁夏耕地资源主要驱动因子概括为农业发展、经济发展及退耕还林,经济发展对耕地产生巨大压力,而农业科技进步某种程度减小了人口对耕地的压力,退耕还林也是宁夏耕地减少的重要原因,三者相互作用共同对宁夏耕地产生影响;预测宁夏耕地资源的BP神经网络模型为3层(3×16×1),对2011~2014年的耕地资源预测取得较好效果,最小误差仅为376 hm^2。

Using the statistical data from 2000 to 2014, the paper analyzed the characteristics of the dynamic change of cultivated land in Ningxia in recent 15 years, extracted the main driving factors which affected cultivated land change using principal component analysis, built a BP neural network which is used to forecast cultivated land resources and carry out the prediction research. The results showed that: Cultivated land resource in Ningxia experienced four phases including increasing, rapidly reduced, fluctuations in growth,rapid growth, and per capita cultivated land experienced rapidly reduce - slowly - rapidly increasing stage. Dynamic analysis of land use,dynamic attitude was less than zero from 2002 to 2004 and from 2010 to 2012, minimum value was 7.673% in 2003, dynamic attitude wasgreater than zero from 2005 to 2010 and from 2012 to 2014, maximum value was 11.424% in 2014. Major driving factors of cultivatedland resources in Ningxia are summarized as agricultural development, economic development and returning farmland to forest, economicdevelopment has a huge pressure on arable land, agricultural science and technology progress to some extent reduce the pressure, andreturning farmland to forest is also the reason for decrease of cultivated land in Ningxia. BP neural network model which is used for prediction of cultivated land resource in Ningxia is three layers(3×16×1), the cultivated land resource forecasting from 2011 to 2014 using the model achieve better effect, least error is only 376 hm2.