Spatio-temporal Autocorrelation Analysis for Regional Land-cover Change Detection from Remote Sensing Data

Abstract

Of the various applications of remote sensing data, characterizing the land-cover dynamics is of utmost significance, providing insights into science, management policy, and several regulatory actions. Recent research works indicate that there is a need to understand and monitor land-cover dynamics at regional scale rather than local scale. However, the regional change is a more generalized concept and therefore, the use of pixel based analysis alone may not be sufficient to get proper insights regarding the land-cover change in remotely sensed imagery. Moreover, higher spectral variation and mixed pixels are two key challenges imposed by satellite imagery, resulting into poor performance of existing pixel-based methods for regional land-cover change detection. In this work, we have proposed a novel approach for detecting regional land-cover changes in satellite imagery using spatio-temporal autocorrelation analysis. Autocorrelation among the neighborhood pixels at various spatio-temporal lags has been utilized here to address the problem of mixed pixel and spectral variation. An index (γ), based on the estimated autocorrelations, has been proposed to classify the regions as 'change' and 'no-change' regions. Moreover, a parameter (σ) has been introduced to provide the measure of regional change significance. The method has been evaluated with Landsat ETM+ imagery (30m resolution) of four zones in and around Kolkata (India), comprising a total of 430 sq. km area (ã 4.8 × 105 pixels). The experimental results are encouraging, with an overall accuracy of 90.66%.