Development and Evaluation of Advanced Classification Systems using Remotely Sensed Data for Accurate Land-Use/Land-Cover Mapping

Abstract

Our general objective in this research was to explore, design, develop, implement, and evaluate advanced classification approaches for more accurate land-use/land-cover mapping using remotely sensed data. The overall research consists of three interrelated studies.

Simulated Annealing (SA) has been shown to be able to overcome the local minimum problem in many optimization methods. Our hypothesis in the first study was that SA-based classification systems could help overcome the local minimum problem in one of such methods, K-means, and thus improve the classification performance. Our experimental results using Landsat Thematic Mapper (TM) images have shown that SA-based classification systems significantly improved the classification accuracy over that of the K-means algorithm when appropriate parameters were chosen.

In the second study, we developed an automated Artificial Neural Network (ANN) classification system and performed a classification of a Landsat TM image. Two hypotheses were tested: 1) the ANN system was suitable for land cover mapping, and 2) the incorporation of SA network could overcome the local minima problem in ANN approaches and improve the resulting classification accuracy. Our study demonstrated that the ANN classification system was a robust and suitable classification system for land cover mapping. Experimental results indicated that the incorporation of SA improved the classification accuracy of an unsupervised ANN network.

ANNs have been shown to have great potential to fuse multiple source data sets. Our hypothesis in the third study was that a proposed two-stage ANN-based multisource classification of Landsat TM and SPOT images could increase the classification accuracy of the derived land categories. Our experimental results demonstrated that the two-stage multisource classification had the best classification performance when proper training sets were used. The adequate and reliable training sets were successfully selected by an Automated Data Selector (ADS) developed in this study.

In summary, our application studies have demonstrated that the automated SA and ANN classification systems are feasible and robust for land-use/land-cover mapping using remotely sensed imageries.