Integrating MODIS and Landsat Data for Land Cover Classification by Multilevel Decision Rule

Xudong Guan,Rui Zhang,Chong Huang

doi:10.3390/land10020208

Abstract

In some cloudy and rainy regions, the cloud cover is high in moderate-high resolution remote sensing images collected by satellites with a low revisit cycle, such as Landsat. This presents an obstacle for classifying land cover in cloud-covered parts of the image. A decision fusion scheme is proposed for improving land cover classification accuracy by integrating the complementary information of MODIS (Moderate-resolution Imaging Spectroradiometer) time series data with Landsat moderate-high spatial resolution data. The multilevel decision fusion method includes two processes. First, MODIS and Landsat data are pre-classified by fuzzy classifiers. Second, the pre-classified results are assembled according to their assessed performance. Thus, better pre-classified results are retained and worse pre-classified results are restrained. For the purpose of solving the resolution difference between MODIS and Landsat data, the proposed fusion scheme employs an object-oriented weight assignment method. A decision rule based on a compromise operator is applied to assemble pre-classified results. Three levels of data containing different types of information are combined, namely the MODIS pixel-level and object-level data, and the Landsat pixel-level data. The multilevel decision fusion scheme was tested on a site in northeast Thailand. The fusion results were compared with the single data source classification results, showing that the multilevel decision fusion results had a higher overall accuracy. The overall accuracy is improved by more than 5 percent. The method was also compared to the two-level combination results and a weighted sum decision rule-based approach. A comparison experiment showed that the multilevel decision fusion rule had a higher overall accuracy than the weighted sum decision rule-based approach and the low-level combination approach. A major limitation of the method is that the accuracy of some of the land covers, where areas are small, are not as improved as the overall accuracy.

Highlights

IntroductionRemote sensing has developed rapidly, efficiently producing Land cover (LC) maps in data and technology [1]
Fuzzy classification was performed on the MODIS data based on a time series similarity measure method, while the Landsat data fuzzy classification were based on a nearest neighbor classifier
When the land cover within a MODIS pixel is homogenous, the information extracted with MODIS 8-day Normalized Difference Vegetation Index (NDVI) data are more reliable

Summary

Introduction

Remote sensing has developed rapidly, efficiently producing LC maps in data and technology [1]. Numerous classification methods have been developed to satisfy requirements and achieve higher accuracy of LC mapping, including graphics technology, such as computer vision and geoscientific knowledge (e.g., multilayer analysis based on object-oriented methods). The types of remote sensing data vary widely, from multi-resolution optical data to synthetic aperture radar (SAR) data. Neither the single classification methods nor data are universally optimal for all situations [2]. Multi-source data classification can show discrepancies in the results. It is thought that the disagreement between the classification results of different remote sensing data sources is a reflection of the complementariness between data [3]

Methods

Findings

Discussion

Conclusion