Building simplification using backpropagation neural networks: a combination of cartographers' expertise and raster-based local perception

Abstract

As a result of the complexity of building shapes and spatial distribution of buildings in urban and suburban areas, building generalization has been one of the most challenging tasks in automated map generalization. The key operator of the generalization is to resolve spatial conflicts inside building polygons caused by a reduction in map scale. Therefore, building simplification eliminates unnecessary details of the building, without the distortion of its original shape. Understanding that map generalization is an intelligent process, we proposed a novel approach for building simplification with raster-based local perception using a backpropagation (BP) neural network (BPNN) model for learning cartographers' knowledge. The model was structured in three layers, an input layer with a 24 × 24 retina, a hidden layer with four nodes, and an output layer with three nodes. Cartographers' expertise, coupled with a square detector designed for the perception of local contexts was presented to the model. In the test, a total of 468 building polygons were simplified, and the mean value of the similarity of building polygons before and after the simplification was 0.9796. The approach not only made the building simplification feasible but also produced satisfactory results.