人工神经网络及其与地统计的混合模型在小面积丘陵区土壤有机碳预测制图上的应用研究

Mapping of Soil Organic Carbon Using Neural Network and Its Mixed Model with Geostatistics in a Small Area of Typical Hilly Region

当前,随着大数据和人工智能的快速发展,深度学习也逐渐被引入数字土壤制图(DSM)研究中。作为深度学习的重要基础之一,人工神经网络(ANN)在DSM中的应用已比较广泛。然而,ANN在40 km2以下的小区域上应用较少,且与其它常用模型的比较研究还不多。因此,在不同的情况下,还没有充分的依据来选用复杂的ANN及其与地统计的混合模型,还是简单的多元线性回归(MLR)、随机森林(RF)、普通克里格(OK)、回归克里格(RK)等方法进行土壤制图。为此,本文以广西高峰林场内一地形变异明显的小区域(面积约3.03 km2)为研究区,以13个地形因子和1个植被因子作为辅助变量,用ANN中常用的径向基神经网络(RBFNN)及其与OK相结合的模型(RBFNN-OK),对土壤有机碳(SOC)含量的空间分布进行预测,并与MLR、RF、OK、RK方法所得结果相比较。结果表明,与其他方法相比,RBFNN-OK和RBFNN在独立随机样本的验证集上预测准确性明显偏低;RBFNN-OK及RBFNN模型预测值的均方根误差值(RMSE)分别为6.57 g kg^(-1)和6.26 g kg^(-1),比MLR、RF、OK、RK高26.54%~31.17%。这可能是因为在小区域上,基于小样本的RBFNN模型泛化能力降低,以至于对训练集以外的样点预测准确性较差。因此,可以认为以RBFNN为典型的ANN及其与地统计的混合模型在小样本、小区域的DSM中适用性可能较差。

At present, with the rapid development of big data and artificial intelligence, deep learning is gradually introduced into the field of digital soil mapping(DSM). As an important origin of deep learning, artificial neural network(ANN) has already been widely used in the DSM field. However, ANN is rarely applied in a small area below 40 km2 and few studies ever compared ANN with other commonly used models. Consequently, for soil mapping under different conditions, there is no sufficient criterion for selecting complex ANN and its mixed model with geostatistics, or simple multiple linear regression(MLR), random forest(RF), ordinary kriging(OK),regression kriging(RK) and other methods. For this reason, this paper took a small area(about 3.03 km2) with obvious topographical variation in GaofengForestin, Guangxi as a study area and selected 13 topographic factors and 1 vegetation factor as auxiliary variables, the radial basis function neural network(RBFNN) and its model combined with OK(RBFNN-OK) were established to predict the spatial distribution of soil organic carbon(SOC)content, and compared with MLR, RF, OK and RK. Results showed that RBFNN-OK and RBFNN had a significantly lower prediction accuracy, based on independent random samples. For instance, the root mean square error(RMSE) values in the RBFNN-OK and RBFNN models were 6.57 g kg^(-1) and 6.26 g kg^(-1), respectively, which were larger than the RMSE values of MLR, RF, OK and RK by 26.54%-31.17%. This is probably because in a small region the generalization ability of RBFNN model based on a small dataset is reduced. The prediction accuracy of samples outside the training dataset was poor. Therefore, this paper indicated that ANN and its mixed model with geostatistics might have poor applicability in DSM with a small number of samples and within a small region.