• en
  • pt-br

Multinomial Logistic Regression and Random Forest Classifiers in Digital Mapping of Soil Classes in Western Haiti

Wesly Jeune, Márcio Rocha Francelino, Eliana de Souza, Elpídio Inácio Fernandes Filho, Genelício Crusoé Rocha

18/Jun/2018

DOI: 10.1590/18069657rbcs20170133

ABSTRACT

Digital soil mapping (DSM) has been increasingly used to provide quick and accurate spatial information to support decision-makers in agricultural and environmental planning programs. In this study, we used a DSM approach to map soils in western Haiti and compare the performance of the Multinomial Logistic Regression (MLR) with Random Forest (RF) to classify the soils. The study area of 4,300 km2 is mostly composed of diverse limestone rocks, alluvial deposits, and, to a lesser extent, basalt. A soil survey was conducted whereby soils were described and classified at 258 sites. Soil samples were collected and subjected to physical and chemical analyses. Recursive Feature Elimination (RFE) was used to select the most important covariates from auxiliary data, such as climate, lithology, and morphometric properties to describe the soil-landscape relationship. Mapping performance was assessed by the Kappa index and overall accuracy derived from a confusion matrix generated using a 5-fold cross validation process. In addition, an external mapping validation was carried out using an independent soil dataset. Accordingly, the soil dataset was split into 80 % and 20 % for training and validation of the models, respectively. No significant statistical difference (Z = 0.56< |1.96|) was found between maps generated with both classifiers (Kappa index 0.45 for MLR and 0.42 for RF). Based on the Kappa values, the classification performance can be characterized as moderate for both algorithms. Surprisingly, the RF classifier outperformed MLR in the validation process (Kappa values of 0.55 and 0.33, respectively). These results suggest a higher generalization ability of RF. However, no significant statistical difference (Z = 1.83< |1.96|) was observed. The soil map derived from RF indicated the occurrence of Leptosols (48.5 %), Gleysols (19.6 %), Chernozems (8 %), and Fluvisols (6.6 %) in most of the study area. The DSM approaches proved suitable for mapping soils in western Haiti and could be used in other parts of the country, thereby closing information gaps with regard to Haitian soils.

Multinomial Logistic Regression and Random Forest Classifiers in Digital Mapping of Soil Classes in Western Haiti

Comments