Roof confusion removal for accurate vegetation extraction in the urban environment

Abstract

We put forward the spectral confusion phenomenon between vegetation and artificial objects - mostly roofs painted with "cool" blue/purple/green pigments in the urban environment. Both of them have the feature of low red and high near-infrared reflectance. For accurate vegetation extraction using high spatial resolution imagery (HSRI), we have developed a two-step threshold segmentation (TSTS) method to solve this confusion. The first step is extracting vegetation and confusing roofs together through threshold segmentation of the NDVI image, and the second step is removing roof confusion through threshold segmentation of an image generated by vegetation and achromatic objects indifferent transformation (VAOIT). VAOIT is derived from the fitting straight line of random trained vegetation and achromatic objects at either highly correlated band combinations: band1/band2 and band1/band3. Efficiency of the method is tested through producer accuracy assessment, and it is demonstrated that VAOIT using either band1/band2 or band1/band3 can remove blue and purple roofs perfectly (producer accuracy=at least 95%), while the former is powerless and the latter is goodish (producer accu- racy=approximately 90%) in removing green roofs. Since too few green roofs exist in our case, more green-roof samples are needed for further test in other cities. Our case study in Shenyang, China demonstrates that TSTS can correct overestimate of vegetation coverage by 2.14%, mostly in industrial blocks, which shows the necessity of roof confusion removal, especially for industrial cities.