Optimal classification methods for mapping agricultural tillage practices

Abstract

The classification of agricultural tillage systems has proven challenging in the past using traditional classification methods due to the similarity of spectral reflectance signatures of soils and senescent crop residues. In this study, five classification methods were examined to determine the most suitable classification algorithm for the identification of no-till (NT) and traditional tillage (TT) cropping methods: minimum distance (MD), Mahalanobis distance, Maximum Likelihood (ML), spectral angle mapping (SAM), and the cosine of the angle concept (CAC). A Landsat ETM+ image acquired over southern Michigan and northern Indiana was used to test these classification methods. Each classification method was validated with 293 ground truth sampling locations collected commensurate with the satellite overpass. Classification accuracy was then assessed using error matrix analysis, Kappa statistics, and tests for statistical significance. The results indicate that of the classification routines examined, the two spectral angle methods were superior to the others. The cosine of the angle concept algorithm outperformed all the other classification routines for tillage practice identification and mapping, yielding an overall accuracy of 97.2% (Kappa=0.959).