Statistical learning algorithms for identifying contrasting tillage practices with Landsat Thematic Mapper data

Abstract

Tillage management practices have a direct impact on water-holding capacity, evaporation, carbon sequestration and water quality. This study examines the feasibility of two statistical learning algorithms, namely the least square support vector machine (LSSVM) and relevance vector machine (RVM), for identifying two contrasting tillage management practices using remote-sensing data. LSSVM is firmly based on statistical learning theory, whereas RVM is a probabilistic model where the training takes place in a Bayesian framework. Input to the LSSVM and RVM algorithms were reflectance values at different bandwidths and indices derived from Landsat Thematic Mapper (TM) data. Ground-truth data for this study were collected from 72 commercial production fields in two counties located in the Texas High Plains of the south-central USA. Numerous LSSVM- and RVM-based tillage models were developed and evaluated for tillage classification accuracy. The percentage correct and kappa statistic were used for the evaluation. The results showed that the best LSSVM and RVM models included the use of TM band 5 or vegetation indices that involved TM band 5, indicating sensitivity of near-infrared reflectance of crop residue cover on the surface. This is consistent with other remote-sensing models reported in the literature. Overall classification accuracies of the best LSSVM and RVM models were 87.8 and 90.2%, respectively. The corresponding kappa statistics for those models were 0.75 and 0.80, respectively. Furthermore, comparison of the best LSSVM and RVM models with the published logistic regression-based tillage models developed with the same data indicated the superiority of the RVM model over LSSVM and logistic regression models in determining contrasting tillage practices with Landsat TM data.