Accurate and site-specific information on tillage practice is vital to understand the impacts of crop management on water quality, soil conservation, and soil carbon sequestration. Remote sensing is a cost-effective technique for surveillance and rapid assessment of tillage practice over large areas. A new empirical approach for accurately predicting tillage class using discriminant analysis (DA) on historical multi-temporal Landsat-TM 5 imagery has been developed. Ground truth data were obtained from the USDA-NRCS at 48 locations (20 conventional till [CT] and 28 conservation tillage or no-till [NT]). Classification accuracies were obtained for the DA models using reflectance values of Landsat-5 TM bands and Normalized Difference Tillage Index (NDTI) values. The performance of the DA models was compared with Logistic Regression (LR) models. On the basis of classification accuracy and kappa (κ) value, our results showed that the DA models performed better in tillage classification than the LR models. However, using NDTI values, both the DA and LR models performed similarly in tillage class discrimination. Model performance improved when a subset of locations rather than years was used. The results indicated broad-scale mapping of tillage practices is feasible using historical Landsat-5 TM imagery and DA-based classification.
ASJC Scopus subject areas
- Plant Science
- Soil Science
- Agricultural and Biological Sciences (miscellaneous)