BRICK KILN DETECTION IN NORTH INDIA WITH SENTINEL IMAGERY USING DEEP LEARNING OF SMALL DATASETS

摘要

Urban air pollution has been rising across several developing countries in Asia. In cities lying in Indo-Gangetic plains of South Asia, air pollution is a result of urbanization growth e.g. construction and industrialization. Brick kilns, factories that make construction bricks, are significant pollutant emitters due to excessive coal usage. Being an unregulated industry, brick kilns are often overlooked in emission inventories due to non-availability of brick kiln locations. We propose a remote sensing-based approach to prepare brick-kiln location map over large-area that can be used in emission inventories in North India. In this paper we used unique features of brick kilns of the bull trench kiln (BK) type to identify them in open remote sensing dataset - Sentinel 2. This is an 'automatic target recognition' problem in which deep learning-based approaches have been very promising. To overcome requirement of huge training dataset for deep learning-based target recognition, we used transfer learning and fine-tuning approach to achieve recognition with a small dataset of 200 samples per class. Apart from the brick-kiln, urban, vegetation and fallow land samples were also prepared for training. Publicly available data: 10 meter resolution Sentinel-2 optical imagery was used for deep learning followed by use of PALSAR for post-processing. Over the region lying of east of New Delhi a hotspot of brick kilns exist that serve to the construction activities in surrounding regions. The brick kiln's producer accuracy with the use of publicly available satellite data was 72.3% and user accuracy was 99.1%. We find the brick kilns are located in those where horizontal or vertical urban growth is taking place. By analyzing the NDVI over the brick kiln locations, their age of operation and impact on operations due to governmental policies could also be estimated. These brick kiln locations can be further used lor land-use emission inventories to assess PM2.5 and black carbon emissions as well as their concentration by chemical transport modeling.