Superresolution Land Cover Mapping Using Spatial Regularization

Abstract

Superresolution mapping (SRM) is a method of predicting the spatial locations of land cover classes within mixed pixels in remotely sensed images. This paper proposes a novel SRM framework that is operated from the perspective of spatial regularization. Within the proposed framework, SRM aims to generate final superresolution land cover maps that conform to inputted fraction images, with spatial regularization intended for exploiting a priori knowledge about the land cover maps. Two SRM models are constructed by using maximal spatial dependence as the spatial regularization term and the L1 or L2 norm as the data fidelity term. The proposed models are evaluated by using synthetic Landsat, real IKONOS, and real Airborne Visible/Infrared Imaging Spectrometer images and compared with hard classification technologies, as well as pixel-swapping, Hopfield neural network, and Markov random field SRM models. We perform linear spectral mixture analysis (LSMA) and multiple endmember spectral mixture analysis (MESMA) to estimate fraction images. Results show that the accuracy of inputted fraction images plays an important role in the final superresolution land cover maps and that using MESMA fraction images results in higher accuracy than using LSMA fraction images. Moreover, the L-curve criterion is suitable for choosing the optimal regularization parameter in both SRM models. Compared with hard classification technologies and other SRM models, the proposed model derives the highest Kappa coefficients and lowest class area proportion errors when MESMA fraction images are used as input.