Application of the Wavelet Neural Network to Land Use Efficiency Analysis:A Case Study on Lanzhou City
小波神经网络在土地利用效益分析中的应用——以兰州市为例

As an indicator of land use sustainability, land use efficiency is a key index to measure how land is used and reflect the condition of regional development. In general, land use efficiency is an integrated concept consisting of three sub-systems: economy benefit, society benefit, and ecology benefit. Many models have been used to study land use efficiency during the past several decades. Empirical studies suggest that the traditional methods suffer somewhat subjectivity, which could lead to unreliable evaluation results. In the present work, the authors proposed an improved neural network model to analyze land use efficiency in Lanzhou City, Gansu Province. The neural network model is a predictive model combined with wavelet analysis and artificial neural network, named wavelet neural network (WNN). Wavelet neural network is a special feedforward neural network, which is of great capabilities of both self-study of artificial neural network and multi-resolution analysis power of wavelet. Lanzhou City is being faced with serious problems associated with land use allocation in the context of accelerating industrialization and urbanization. Societal, economic and ecologic data involving 8 indicators, i.e., fixed assets investment, per capita net income, per capita GDP, growth rate of GDP, per capita grain yield, per capita arable land, population density, and forest coverage, in Lanzhou City across the period 1996 -2003 were used as inputs. The land use efficiency model was subsequently constructed using the wavelet neural network in combination with these data. We applied the model to comprehensively investigate the land use efficiency of Lanzhou City in years 2004 and 2005. Some parameters with regard to the WNN model were specified as follows: network structure 8-17-1, best training times 500 until the network is converged, learning rate 0.02, and error only 0.001. Finally, comparison among the WNN, entropy weight method, and back-propagation (BP) neural network was performed. Results show that the WNN model is relatively simple, suggesting a higher approximation capability and a faster convergent speed. The forecasting accuracy seems to be better than other methods being investigation. In addition, the activation function of WNN for nonlinear function approximation tends to be more effective and accurate than the Sigmoid function involved in the conventional neural network for the same step. It is indicated that the WNN model can be useful in eliminating the effects of human interventions and open a new opportunity to investigate land use efficiency. It can also provide a basis for optimal allocation of land resources.