Land Cover Classification and Analysis Using Radar and Landsat Data in North Central Ethiopia

Radar and Landsat data were used to classify land cover in north central Ethiopia. Both images were registered and resampled to 12.5 m spatial resolution. Maximum Likelihood Classifier (MLC) and C4.5 algorithm were applied. The original radar data produced low overall classification accuracy (66%). To improve this classification accuracy, de-speckling and texture measures were used for image enhancements. The de-speckling methods used in this study are Median, Lee-Sigma and Gamma-MAP. Lee-sigma, Gamma-MAP and Median de-speckling improved the overall accuracy by 15, 18 and 20% respectively. The maximum overall accuracy achieved in this study by despeckling method is 86.4% using Median at 27*27. Urban producer accuracy improved by 58% by using Median despeckling. All de-speckling techniques improved urban user accuracy to more than 90%. In most de-speckling cases, MLC outperformed C4.5 classifier in the overall classification accuracy. The highest overall accuracies achieved by texture are 88.8 and 90.5% when MLC and C4.5 algorithms at window size 51*51 were used respectively. This shows 22% improvement compared to the original radar data. Urban and forest producer accuracy improved by 58 and 26% respectively at window size 43*43. The overall classification accuracy of Landsat data is 93.7%. Combining Landsat and derived radar data measures improved land cover accuracy by about 5%. This study showed the importance of texture and de-speckling techniques to improve a land cover classification in radar data. Therefore, radar data can be used as an alternative to optical data in the tropics and Ethiopia for land cover classification.