Increasing the Informativity of Multispectral Satellite Images Using Texture Analysis Data

Abstract

The article considers the problem of incresing the informativity of multispectral images of medium (10–30 m) and high (1–4 m) spatial resolution obtained from foreign and national satellite remote sensing systems by involving additional textural information from panchromatic satellite images of very high spatial resolution (≲(1–0.4) m). The images of test sites on the territory of Savvat’yevo forestry (Tver region) from Landsat 8, Sentinel 2 and WorldView 2 satellites equipped with multispectral instruments were an object of this research. Geo-referenced ground survey data were used to validate the calculation results. We used the values of the spectral reflectance in the visible and near-infrared channels normalized to the appropriate integral characteristic as spectral features. Statistical characteristics were calculated in order to extract texture features based on the distribution of the co-occurrence of gray levels (Haralick texture features) within a moving window running the image with a given spatial step. A correlation analysis of textural features was carried out considering changes in distance and angle of adjacency. It was shown that for the selected leading features (autocorrelation, asymmetry, contrast and correlation) the first three can be used with an arbitrary direction of adjacency, while the latter needs to be considered in two different directions. Also we have found that all the considered classification algorithms provide a significant increase of accuracy when both spectral and textural features are used, in comparison with the traditional spectral classification. This result was shown for all images of test sites obtained by different satellites. It is possible to make a preliminary conclusion that the proposed integrated approach of thematic processing can improve the quality of object recognition in the case of using images of both medium and high spatial resolution. Estimates obtained during the thematic mapping of dominant and subdominant forest species showed close classification accuracies for different initial multispectral images (with a scatter of no more than 5 % around the average value of 85 %). Mostly this is due to the presence of specific errors in the ground-based forest inventory data and indicates the necessity of their updating with the use of satellite remote sensing images.