The newly developed hyperspherical color space (HCS) fusion method is a cost-effective panchromatic (pan) sharpening technique, but it still suffers from local spectral distortion and insufficient overall spatial details. This paper attempts to improve the HCS fusion method by combining the spectrum gain modulation and wavelet transformation and presents a novel fusion method called enhanced HCS (EHCS) to mitigate the mentioned issue. This method mainly contains three steps. (1) The proposed method extracts the angular component from multispectral images through hyperspherical color transformation (HCT). (2) A modulated multispectral image is generated by using the spectrum gain modulation. Then, the spatial details of the modulated multispectral image are substituted with the spatial details of the pan image by multiscale wavelet decomposition. (3) Finally, the angular component and modulated detail image are integrated to achieve the fused image by the inverse HCT. To validate the proposed method, three high spatial resolution images, i.e., two WorldView-2 and one QuickBird datasets, acquired at Fujian, China, were used. The comparison with HCS method was also done. Both the visual and quantitative analysis demonstrate that the EHCS method can significantly preserve the spectral characteristics and enhance the spatial detail as well.
A robust method to conduct land use change detection between multi-temporal images using projection pursuit learning
network architecture (PPLNA) is proposed. The method uses a parallel approach that includes three different PPLNs:
two of them are used to generate the change map using the multi-spectral information, while the third produces a change
mask exploiting multi-temporality. The distinctive feature and major merit of PPLNA from traditional neural network for
land use change detection are the proposed method simultaneously exploits both the post classification of multi-spectral
and multi-temporal information that is associated with the changes values of the pixel spectral reflectance, and hence
improve the change detection accuracies. To validate the performance of the proposed method, the experiments using the
ETM+ images for the area of Calgary have been carried out. The accuracies of the final classification and change
detection maps have been evaluated with ground truth comparisons. The experimental result demonstrates that the
proposed method achieves better accuracies.
A hybrid method integrated wavelet spectral feature with total least square algorithm for improving abundance estimation of hyper-spectral mixture pixels is proposed. The method uses the wavelet transform as a pre-processing step for the spectral feature extraction to decrease the within end-member variability, and then utilizes total least square (TLS) algorithm to capture the spectral variations between end-members. The hybrid method can take both technique advantages to reduce the impact of spectral variations with different format. Consequently, the approach provides a potential ability to reduce and tackle within end-member variation inherent in real mixture pixels, and hence to improve abundance estimation. Experiment of simulating mixture spectral data is conducted to validate the procedures, and the results demonstrate that the proposed method can reduce the abundance estimation deviation over 20% on average in the case of spectral end-member variations, as compared to that of the original hyper-spectral signals with least square estimation approach does. Comparisons with the decomposition of wavelet based features (DWT) and total least square have also been implemented, and the experiment shows the hybrid method can also improve the abundance estimation by 5%-10% than those of DWT and TLS do in terms of average RMSE.
Texture analysis has received great attention in the interpretation of high-resolution satellite images. This paper aims to
find optimal filters for discriminating between residential areas and other land cover types in high spatial resolution
satellite imagery. Moreover, in order to reduce the blurring border effect, inherent in texture analysis and which
introduces important errors in the transition areas between different texture units, a classification procedure is designed
for such high spatial resolution satellite images as follows. Firstly, residential areas are detected using Gabor texture
features, and two clusters, one a residential area and the other not, are detected using the fuzzy C-Means algorithm, in the
frequency space based on Gabor filters. Sequentially, a mask is generated to eliminate residential areas so that other
land-cover types would be classified accurately, and not interfered with the spectrally heterogeneous residential areas.
Afterwards, other objects are classified using spectral features by the MAP (maximum a posterior) - ICM (iterated
conditional mode) classification algorithm designed to enforce the spatial constraints into classification. Experimental
results on high spatial resolution remote sensing data confirm that the proposed algorithm provide remarkably better
detection accuracy than conventional approaches in terms of both objective measurements and visual evaluation.