| 摘 要： |
Optimization of the sampling design is significant for generating accurate soil maps based on samples. Simulated annealing (SA) with the mean squared prediction error (MSE) as objective function has been proven as an effective method for sampling optimization with prior samples for the MSE calculation. In this study, an improved artificial neural network (the improved ANN) was developed firstly and then utilized to calculate the objective function of SA for sampling optimization (named as the improved ANN_SA). The proposed approach was evaluated to generate optimized samples with a series of samples sizes (from 10 to 500 points) for mapping soil organic matter (SOM) content from the existing 5054 samples in a study area with 1067 km(2) in Jiangsu Province, China. A commonly-used sampling optimization using SA with the objective function calculated by regression kriging (named as RK_SA) was taken as a reference method. The two sampling optimization approaches were compared in mapping SOM using three prediction methods, the improved ANN, regression kriging (RK) and random forest. The results showed that the proposed sampling optimization approach generally achieved more accurate prediction over different sampling sizes. The maximum improvement of prediction accuracy by using the proposed sampling strategy vs. the reference strategy was 12.5%, 53.6%, and 15.5% when using the improved ANN, RK, and random forest as soil mapping methods, respectively. Moreover, the improved ANN and random forest generated more accurate soil predictions than RK with the proposed sampling optimization approach. The superiority of the proposed sampling optimization was more obvious in low sampling densities (smaller than 0.3 points/km(2)). We conclude that the improved ANN_SA sampling is a potential effective sampling optimization approach, and machine learning is a promising method for soil mapping when using this sampling approach. |
在线留言
联系我们
所长信箱