f 
T 
— (0 A 
    
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B7, 2012 
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia 
POST-CLASSIFICATION APPROACH BASED ON GEOSTATISTICS TO REMOTE 
SENSING IMAGES : SPECTRAL AND SPATIAL INFORMATION FUSION 
N. Yao**, J. X. Zhang*, Z. J. Lin®, C. F. Ren? 
* Remote Sensing information Engineering School, Wuhan University, 129 Luoyu Road, Wuhan, China 
nayao@foxmail.com; jxzhang@whu.edu.cn; ren_cf@163.com 
® Chinese Academy of Surveying and Mapping, 16 Beitaiping Road, Beijing, China - lincasm@casm.ac.cn 
KEY WORDS: Classification, Information, Fusion, Spectral, Spatial, Geostatistics, Kriging 
ABSTRACT: 
Classification of remote sensing imagery provides an inexpensive yet efficient approach to land cover mapping. In supervised image 
classification, training samples are collected through certain sampling schemes, which are used to derive classification rules, aiming 
for adequate accuracy for the applications at hand. However, in conventional classification methods, the potential of training 
samples in terms of locational information is not tapped further, confounding the classification accuracy to the limited separability 
inherent to the given input feature vector. This paper explores two methods pertaining to geostatistics, i.e., simple kriging with local 
mean and cokriging, to predict class occurrences based on training samples’ indicator transforms (location and classes) and 
spectrally derived class probabilities, thus calibrating the a posterior class probability vectors derived from initial spectral 
classification. The results showed that classification accuracy is significantly increased by these two methods for utilizing spatial 
information contained in training samples and initial spectral classification, compared with those obtainable with spectral 
classification. Moreover, the proposed methods constitute a valuable strategy for making fuller use of information residing in 
training data for improving spectrally derived classification, which is independent of the specific classifiers initially adopted for 
image classification. 
1. INTRODUCTION 
Land cover, as a spatial factor impacting and linking human life 
and natural environment, refers to the observed (bio)physical 
cover on the earth's surface (FAO, 2000). Remote sensing is an 
attractive data source for land cover mapping. Although remote 
sensing has been used successfully in mapping a range of land 
covers at a variety of spatial and temporal scales, the land cover 
maps derived are often judged to be of insufficient quality for 
operational applications (Foody, 2002). Therefore, how to 
improve the quality of land cover maps has been a hot issue all 
the time. 
Thematic mapping, exemplified by land cover mapping, from 
remotely sensed data is typically based on an image 
classification (Foody, 2002). Hence, the performance of a 
certain classifier would become a key factor which impacts on 
the classification accuracy and further impacts on the quality of 
the derived maps. An operating classifier can be considered as a 
system that reduces the initial uncertainty by consuming the 
information contained in the input vector (Battiti, 1995). Battiti 
further indicates that the final uncertainty will be zero in the 
ideal case (i.e., the class will be certain), while it can be higher 
in the actual applications for at least two different reasons, i.e., 
insufficient input information or suboptimal operation. 
Hence, the performance of a classifier is one possibility that 
relates to suboptimal operation. That is, even if sufficient input 
information is given, classification accuracy quantified by the 
confusion matrix may be lower than its potential value due to 
the insufficient training of the classifier. On the other hand, the 
information loss during the training period manifests a 
conceivable promotion space of classification accuracy. 
  
* Corresponding author. 
Therefore, if a strategy is capable of compensating or reducing 
the information loss, an accuracy promotion of classification 
would be foreseeable. 
Traditional spectral classification of remotely sensed images 
applied on a pixel-by-pixel basis ignores the potentially useful 
spatial information between the values of proximate pixels 
(Atkinson, 2000; Zhang, 2009). Geostatistical approaches, 
adopted in this paper, indeed aim to employ the spatial 
information inherent in remotely sensed images to enhance the 
spectral classification. Spatial information mainly serves two 
purposes that derive texture “wavebands” for subsequent use in 
classification or smooth the imagery prior to or after 
classification (Atkinson, 2000). 
This paper utilizes two kriging approaches, i.e., simple kriging 
with local and cokriging, to fusion the input (e.g., spectral ) and 
spatial information. In the following sections, first the 
principles of the two kriging paradigms are revisited, then an 
index which assesses the potential separability of a data set is 
introduced, and finally the experimental results are presented 
and analyzed. 
2. METHOD 
Spectral response in each waveband of a remotely sensed image 
may be treated as a continuous variable, which is also defined 
as a regional variable in geostatistics. The variogram (or 
covariance function) is a quantitative model which reflects the 
relationship and spatial structure of the regional variables.