Deep Learning Approach for Monitoring Urban Land Cover Changes

Krishna Kumar Perikamana, Krishnachandran Balakrishnan, Pratyush Tripathy  | 2024


Time series data of urban environment are extremely useful for analyzing urban growth trends, changes in impervious surface and vegetation distribution, and the associated consequences on urban microclimate. While medium-resolution Landsat data are excellent for such analysis, the existence of mixed pixels, which reflect a composite spectral response of the component classes, limits pixel-based categorization. Hence, to yield the full potential of the Landsat data, the proportions of classes (subpixels) in these mixed pixels need to be estimated which gives individual class proportions. This study proposes a subpixel classification method that leverages the temporal overlap of Landsat-5 TM and Resourcesat-1 LISS-IV sensors. The 30-m resolution Landsat-5 TM imagery is available from 1985 to 2011 and the 5-m resolution Resourcesat-1 LISS-IV imagery is available from 2003 to 2013. Hence, a direct mapping from Landsat-5 TM to Resourcesat-1 LISS-IV is highly feasible to estimate the percentage of built-up and vegetation fraction in each 30-m resolution cell. In the chapter, a convolutional neural network was trained to predict fractional land cover maps from 30 m Landsat-5 TM satellite data. The reference land cover fractions were estimated from a hard-classified LISS-IV image for Bengaluru from 2011. Further, the generalizability of the proposed model was demonstrated using satellite data for Mumbai (2009). For both Bengaluru (2011) and Mumbai (2009) datasets, the mean absolute error falls within the range of 7.2%–11.3% for both built-up and vegetation fraction prediction at the 30 m cell scale. The chapter highlights the subpixel classification by successfully simulating the sensor behavior of Landsat-5 TM, which can effectively be used for time series analysis to monitor the changes within the urban landscape.