Abstract:
Multimodal dataset shifts consisting of both concept and covariate shifts are addressed in this study to improve texture-based land-use classification accuracy for optical panchromatic and multispectral remote sensing. Multitemporal and multisensor variances between train and test data are caused by atmospheric, phenological, sensor, illumination and viewing geometry differences, which cause supervised classification inaccuracies. The first dataset shift reduction strategy involves input modification through shadow removal before feature extraction with gray-level co-occurrence matrix and local binary pattern features.
Components of a Rayleigh quotient-based manifold alignment framework is investigated to reduce multimodal dataset shift at the input level of the classifier through unsupervised classification, followed by manifold matching to transfer classification labels by finding across-domain cluster correspondences. The ability of weighted hierarchical agglomerative clustering to partition poorly separated feature spaces is explored and weight-generalized internal validation is used for unsupervised cardinality determination. Manifold matching solves the Hungarian algorithm with a cost matrix featuring geometric similarity measurements that assume the preservation of intrinsic structure across the dataset shift. Local neighborhood geometric co-occurrence frequency information is recovered and a novel integration thereof is shown to improve matching accuracy.
A final strategy for addressing multimodal dataset shift is multiscale feature learning, which is used within a convolutional neural network to obtain optimal hierarchical feature representations instead of engineered texture features that may be sub-optimal. Feature learning is shown to produce features that are robust against multimodal acquisition differences in a benchmark land-use classification dataset. A novel multiscale input strategy is proposed for an optimized convolutional neural network that improves classification accuracy to a competitive level for the UC Merced benchmark dataset and outperforms single-scale input methods. All the proposed strategies for addressing multimodal dataset shift in land-use image classification have resulted in significant accuracy improvements for various multitemporal and multimodal datasets.
