Abstract:

Fractional vegetation coverage is a key parameter for monitoring ecosystems and its functions. How to improve the retrieval accuracy of fractional vegetation coverage in large areas is very important for the sustainable development of the environment in ecologically fragile areas. Based on machine learning methods such as back propagation neural network （BP-ANN）， support vector regression （SVR） and random forest （RF）， this study uses UAV， Worldview-2 and Landsat 8 OLI remote sensing data to carry out multi-scale inversion of fractional vegetation coverage in Horqin Sandy Land. The results show that： （1） The RF model performs better than the BP-ANN and SVR model. It can invert the vegetation coverage of sandy land with higher accuracy on the scale of unit （experimental area） and regional （research area）. There was a significant correlation between the inversion value and the measured value of UAV at the linear level （P is less than 0.01）. （2） On the unit and regional scales， the test set of the constructed vegetation coverage inversion model are R² of 0.84 and 0.80， MSE of 0.0145 and 0.0370， and the consistency index d of 0.9576 and 0.8991 respectively. The method of gradually realizing the large area fractional vegetation coverage inversion of low spatial resolution remote sensing images by using multi-source remote sensing data and machine learning method， which can not only effectively improve the inversion accuracy of fraction vegetation coverage in sandy land （R²=0.78， less than 0.63）， but also provide support for regional ecological environment monitoring and ecosystem health assessment.

Key words: fractional vegetation cover, multi-source remote sensing, machine learning method, Horqin Sandy Land