A method of estimating imperviousness for the catchment modelling of urban environments

Abstract

Urban impervious surfaces, a symbol of urbanisation, have permanently changed urban hydrology behaviour and play a critical role in modelling rainfall-runoff process. The distribution pattern of impervious surfaces is intrinsically connected with functional land zoning schemes. However, estimating impervious fractions for catchment modelling is becoming increasingly difficult due to intricate land zoning categories and heterogeneous land use land cover (LULC) during urbanisation. This study demonstrates an integrated approach of deep learning (DL) and grid sampling method to overcome the challenges of LULC classification, sample standardisation and statistical sample extraction. The classified impervious features were extracted within the land zoning scope and translated into polynomial functions using a probability-fitting approach to measure the occurrence likelihood distribution of samples' impervious fraction. Then, we use the information entropy (IE) to evaluate prediction stability by quantifying the condition entropy and information gain (IG) from each functional land zones to the occurrence likelihood of different impervious fraction intervals. The DL model shows robust LULC prediction, while probability-fitting study of impervious samples reflects the distribution differential of impervious fractions under the land zoning categories. The IE stability test shows a robust approach that clarifies different confident ranges of imperviousness estimation based on land zoning information. HIGHLIGHTS Using DL techniques to classify and segment land use land cover (LULC) from the remote sensing imagery of a large urban catchment (total area: 4,348 ha) and determine impervious LULC classes.; The urban functional land zoning concept was involved to define the impervious feature extraction and analysis scope.; The IE concept was introduced to assess the stability of results and to quantify confidence ranges for different land zoning categories.;

Keywords

• catchment modelling
• deep learning
• information entropy
• land use land cover (lulc)
• probability-fitting