Timely and accurate detection of land cover/use changes is one of the most important issues in land planning and management. Remote sensing (RS) images have become an important data source for change detection (CD) in recent decades. Thresholding of difference image (DI) is a prevalent approach for RS-based CD. It can be shown that the changes in an environment are occurred in such a way that the different spectral changes of phenomenon can be detected in different parts of electromagnetic spectrum. Hence, utilization of several spectral bands can offer a higher accuracy in CD process. However, prevalent thresholding techniques are developed for one-dimensional space and they are not appropriate for CD in multi-dimentional space of RS images. The common approach to overcome this deficiency is to fuse data at feature and/or decision level. Some methods have already been developed for this purpose. Whereas, it is enigmatic to decide which of data fusion technique is the most appropriate one, a common particularity in all these approaches (except: voting and Bayesian) is their supervised nature, as the analyst must determine some parameters which can be the best fit to a certain application and dataset. On the other hand, unsupervised approaches, generally have low accuracy in CD process.
In order to develop the thresholding technique to support multi-spectral images, a simple yet effective data fusion approach is proposed in this paper. The developed method is a linear combination of multi-spectral change image based on fusion. Applied weights in linear combination are optimized using Particle Swarm Optimization (PSO) algorithm.
The proposed approach consists of the following two major steps. In the first step a multi-spectral change image is generated. Several methods can be used for that purpose. In this research, we chose difference image operation as it is simple to implement and easy to interpret. It includes a simple and straightforward arithmetic difference between the digital values of the two images obtained on different dates. In the next step, PSO is initialized with arbitrary weights and the weighted image fusion is then carried out as follows: . Where denotes the weight associated to ith band of multi-spectral difference image ( ), such that Afterwards, the OTSU thresholding technique is applied to produce binary change mask (BCM) and evaluate the fitness of the fused change index (FCI). If any of the termination conditions (optimum fitness or maximum number of iteration) is satisfied, the current weights are saved as optimum weights of a weighted linear combination or else they are updated with PSO algorithm to reach the optimum values.
The performance of the developed technique is evaluated on a bi-temporal multispectral images acquired by the Landsat-5 Thematic Mapper (TM) sensor in July 2000 and 2009. This data set is characterized by a spatial resolution of 30m×30m and 7 spectral bands ranging from blue light to shortwave infrared (0.45~2.35 µm). It is worth noting that the 6th band of these images (thermal infrared band), is not utilized due to low spatial resolution. The selected area is co-registered subsets of size (470×830 pixels) of two full scenes, including Khodafarin Dam (an earth-fill embankment dam on the Aras River straddling the border between Iran and Azerbaijan).
Moreover to visual assessment of CD results, quantitative analysis has been carried out by selecting 2799 samples of changed regions and 5168 samples of unchanged regions, according to field work and image interpretation. The proposed linear combination of multispectral difference images based on fusion which is the development of the thresholding technique to support the multi-spectral images, has better accuracy in CD in comparison with individual spectral bands of DI and the other state-of-the-art image fusion algorithms at feature and/or decision level. Overall accuracy of 90.68% using the proposed method in comparison to an overall accuracy of 79.06% and 70.81% related to the prevalent voting algorithms (data fusion at decision level) and 80.77% related to the Bayesian algorithm (data fusion at feature level), confirms the effectiveness of the proposed method for unsupervised CD in multi-spectral and multi-temporal RS images.