Digital Soil Mapping Using Machine Learning Algorithms in a Tropical Mountainous Area

ABSTRACT: Increasingly, applications of machine learning techniques for digital soil mapping (DSM) are being used for different soil mapping purposes. Considering the variety of models available, it is important to know their performance in relation to soil data and environmental variables involved in soil mapping. This paper investigated the performance of eight machine learning algorithms for soil mapping in a tropical mountainous area of an official rural settlement in the Zona da Mata region in Brazil. Morphometric maps generated from a digital elevation model, together with Landsat-8 satellite imagery, and climatic maps, were among the set of covariates to be selected by the Recursive Feature Elimination algorithm to predict soil types using machine learning algorithms. Mapping performance was assessed using the confusion matrix, and the Z-test among the Kappa indexes of the matrices. In a conventional soil survey, the soils described and classified in the Brazilian System of Soil Classification [Argissolos Vermelho-Amarelos Distróficos – PVAd (Acrisols), Cambissolos Háplicos Tb Distróficos - CXbd (Cambisols), Gleissolos Háplicos Háplicos Tb Distróficos - GXbd (Gleysols), Latossolos Amarelos Distróficos - LAd (Xanthic Ferralsos), Latossolos Vermelho-Amarelos Distróficos - LVAd (Rhodic Ferralsols), and Neossolos Litólicos Distróficos - RLd (Neossols)] were grouped into composite mapping units (MU) using the conventional method. The eight algorithms showed similar performance without statistical difference (Kappa 0.42-0.48). The mapping of soils with varying slopes (LAd, LVAd, CXbd) showed lower accuracy, whereas soils on hydromorphic lowlands (GXbd) were classified more accurately. In map algebra, the result was rather satisfactory, with 63-67 % agreement between the conventional soil map and maps produced by machine learning. The areas with the largest disagreement in the DSM occurred in the LAd unit due to subtle color variation in the Latossolos mantle without a clear relation to any environmental variable, highlighting difficulties in DSM regarding hill slope landforms. Model performance was satisfactory, and good agreement with the conventional soil map demonstrates the importance of the DSM as a potential complementary tool for assisting soil mapping in mountainous areas in Brazil for the purpose of land use planning.

摘要：机器学习技术在数字土壤制图（digital soil mapping, DSM）领域的应用正日益广泛地服务于各类土壤制图相关需求。鉴于当前可用模型种类繁多，了解不同模型针对土壤制图相关土壤数据与环境变量的表现至关重要。本研究针对巴西祖纳达马塔（Zona da Mata）区域一处官方农村定居点的热带山地，探究了8种机器学习算法在土壤制图中的表现。本研究采用的协变量集由数字高程模型生成的形态计量图、Landsat-8卫星影像以及气候图构成，通过递归特征消除法（Recursive Feature Elimination）从中筛选最优变量，以借助机器学习算法实现土壤类型预测。制图表现通过混淆矩阵（confusion matrix）以及各矩阵Kappa指数（Kappa indexes）间的Z检验（Z-test）进行评估。本研究依托常规方法，将按照巴西土壤分类系统（Brazilian System of Soil Classification）开展描述与分类的下述土壤划分为复合制图单元（composite mapping units，MU）：瘠薄红黄铁质土（Argissolos Vermelho-Amarelos Distróficos，缩写PVAd，对应Acrisols）、瘠薄腐殖质薄层土（Cambissolos Háplicos Tb Distróficos，缩写CXbd，对应Cambisols）、瘠薄腐殖质滞水黏磐土（Gleissolos Háplicos Háplicos Tb Distróficos，缩写GXbd，对应Gleysols）、瘠薄黄色铁质土（Latossolos Amarelos Distróficos，缩写LAd，对应Xanthic Ferralsols）、瘠薄红黄铁质土（Latossolos Vermelho-Amarelos Distróficos，缩写LVAd，对应Rhodic Ferralsols）以及瘠薄岩性新成土（Neossolos Litólicos Distróficos，缩写RLd，对应Neossols）。8种算法的表现相近，无统计学差异（Kappa指数区间为0.42~0.48）。坡度各异的土壤（LAd、LVAd、CXbd）的制图精度较低，而水成低地土壤（GXbd）的分类精度更高。在地图代数分析中，研究结果较为理想：常规土壤图与机器学习生成的土壤图的吻合度达63%~67%。数字土壤制图中分歧最大的区域出现在LAd单元，这是因为铁质土（Latossolos）覆盖层存在细微的颜色变化，且与任何环境变量均无明确关联，凸显了山地地形下数字土壤制图所面临的挑战。模型表现良好，且与常规土壤图的高吻合度证明，数字土壤制图可作为潜在辅助工具，助力巴西山地地区的土壤制图工作，服务于土地利用规划目标。