基于遗传算法优化BP神经网络的土壤盐渍化反演

应用于土壤盐分含量(Soil Salinity Content,SSC)反演的BP神经网络(Back Propagation Neural Network,BPNN)较少关注对模型精度影响较大的结构参数和初始权重的优化。该文利用Landsat-8 OLI、Sentinel-1 SAR影像数据及SRTM高程数据,基于谷歌地球引擎(GEE)平台构建反演参数,并建立3种反演模型:先利用遗传算法(Genetic Algorithm,GA)同步优化输入层反演参数子集和隐含层神经元数量,再优化初始权重的BPNN(GA-BP)模型;将变量投影重要性(Variable Importance in Projection,VIP)算法分割阈值分别设为1和0.5,优化出两组输入层反演参数子集并将其分别代入GA优化隐含层神经元数量,再优化初始权重的BPNN(VIP1-GA-BP、VIP2-GA-BP)模型。在玛纳斯流域和三工河流域各选一靶区进行SSC反演,对比分析GA-BP、VIP1-GA-BP、VIP2-GA-BP模型的反演精度,并统计各类盐渍土的面积比例,结果表明:1)两靶区3组模型反演精度由高到低排序均为GA-BP、VIP1-GA-BP、VIP2-GA-BP;2)盐分指数和植被指数在SSC反演中起到重要作用,同一模型筛选的反演参数存在空间分异性,但高程适用于不同的筛选模型,具有较强的鲁棒性;3)两靶区3组模型反演的SSC值域范围与实际采样点SSC值域范围的差异均较小,各子区GA-BP反演的SSC空间分布地物轮廓最清晰,且地物内SSC的均质性最好;4)玛纳斯靶区和三工河靶区面积占比最大的盐渍土类型分别为盐渍土和中度盐渍土。研究结果为构建具有一定推广性的干旱区土壤盐分含量反演模型奠定了基础。

Retrieval of Soil Salinity Content Based on BP Neural Network Optimized by Genetic Algorithm

Less attention is paid to optimization of structural parameters and initial weights in back propagation neural network(BPNN)applied to inversion of soil salinity content(SSC).Selecting a target area in Manas River Basin and Sangong River Basin respectively,using Landsat-8 OLI,Sentinel-1 SAR and SRTM digital elevation data(DEM),the inversion parameters were extracted based on Google earth engine(GEE).In each target area,using the genetic algorithm(GA),the subset of input layer inversion parameters and the number of hidden layer neurons were synchronously optimized,and then the initial weights were optimized to build BPNN(GA-BP)model;two subsets of input layer inversion parameters optimized by variable importance in projection(VIP)whose segmentation thresholds were set as 1 and 0.5 were put into BPNN models(VIP1-GA-BP,VIP2-GA-BP)whose number of hidden layer neurons and initial weights were optimized by the genetic algorithm.The area ratio of various types of saline soils in each target area was counted and the precision of GA-BP,VIP1-GA-BP and VIP2-GA-BP were compared for the SSC inversion.The results are as follows.1)In each target area,the ranking of prediction accuracy of models from good to bad is GA-BP,VIP1-GA-BP,VIP2-GA-BP.2)Salinity index and vegetation index play an important role in the inversion of SSC;the inversion parameters screened by the same model have spatial differentiation,and elevation of the inversion parameter has strong robustness.3)In each target area,the difference between the SSC range predicted by each model and that of the actual sampling points is relatively small,and the spatial distribution of SSC inversed by GA-BP in each sub-area is relatively clear,and the homogeneity of SSC within the same surface feature is good.4)The types of saline soil with the largest proportion in Manas target area and Sangong target area are salinized soil and moderately salinized soil respectively.This work has laid a foundation for building a generalized inversion model of SSC for arid regions.