Improved method for discriminating agricultural crops using geostatistics and remote sensing

Reliable land cover mapping of agricultural areas require high resolution remote sensing and robust classification techniques. In this paper, we propose the integration of spectral information with spatial information using the traditional statistical supervised classifier "Maximum Likelihood" and a geostatistical tool, "Indicator Kriging" algorithm, for the development of land cover maps by supervised classification from remotely sensed data at medium and high spatial resolution. The proposed method showed better results in classes' discrimination with smoother resulting maps than the ones produced using only spectral information. Two different satellites imagery were analyzed: a Landsat TM5 image at medium spatial resolution acquired during 2006 and an Ikonos II image at higher spatial resolution acquired during 2008. The better performance of the "combined" approach compared to the traditional Maximum Likelihood technique was confirmed by confusion matrix. The overall accuracy increases from 76.16% to 85.96% for LandsatTM image and from 71.56% to 80.25% for the IKONOS image.