| Recent advances in technology has allowed geospatial professionals the capability to remotely collect multiple types of sensor data over a specific target.On one hand,the heterogeneity of these data has brought about challenges in their efficient and effective processing,but on the other hand,the influx in and availability of such ancillary datasets has also opened up new opportunities to combinatorically utilize multi-source and/or multi-modal to improve the performances of geospatial applications.The adequacy of LiDAR data –which provides rich spatial/geometric information – for(complex)urban scene understanding tasks like classification is constrained by the lack of or limited availability of spectral information.Hyperspectral imaging on the other hand provides rich spectral information but comes with issues such as high dimensionality,lack of elevation information,limited availability of training samples and some level of noise impact from imaging sensors and the environment,thus making hyperspectral image processing tasks like classification a challenge.Given the limitations and capabilities of each sensor,and the fact that the basis of a good classification scheme stems from robust and discriminative feature extraction/representation,this study aims at obtaining optimal feature representations of objects in an urban scene for raw LiDAR point cloud labelling,via a deep learning based spatio-spectral feature representation approach.In this approach,a class-aware band selection and reduction technique(CBSR)is developed for initial hyperspectral feature extraction;followed by a double-branched convolutional Gaussian-Bernoulli deep belief network(CGBDBN)for deep,hierarchical,spatial feature extraction from both LiDAR and hyperspectral image data.Using stacked ensemble modeling,spatio-spectral features are generated from the two feature streams using a fusion rule and then classified,the result of which is then used in labelling a raw 3D LiDAR point cloud through projection.Experiments were conducted on the IEEE geoscience and remote sensing society(GRSS)2018 data fusion contest dataset for urban scene classification and the MUUFL Gulfport dataset for urban scene classification.The results indicate that the developed CBSR technique – which has the capacity to maximize inter-class variability,and also minimize the decay of minor but significant features that can help distinguish spectrally similar classes – attained very competitive results compared with state-of-the-art approaches,thus indicating its ability to generate highly discriminative features and making it a robust spectral feature representation technique.Also,the weight sharing property,probabilistic modelling,deep and hierarchical nature of the CGBDBN gives it the ability to capture highly discriminative,hierarchical and semantic features.Furthermore,the stacking of ensembles to complement each other in a classification task is more efficient and robust compared to weighting the ensembles.Moreover,compared to the spatial or spectral features,the generated spatio-spectral features are more discriminative and significantly aided in improving the proposed model's efficacy.In addition,the incorporation of hyperspectral features as spectral feature in multi-source data classification tasks is much more effective in boosting said performance compared to discrete multi-spectral ones like those obtained from RGB or IRRG.Also,the strategy of labelling raw point clouds based on existing HSI ground truth not only reduces time and labor costs,but also opens up an avenue to label point data into informative sub-classes.Overall,the proposed approach,based on the evaluation metrics and computational time,is a robust and effective approach for both coarse-and fine-grained raw point cloud labelling. |
PDF Full Text Request