Executive Summary : | Availability of diverse sources of spatial data have enhanced the decision support capabilities. However, lack of a proper dynamic data-centric system limits the potential utilization of the available data. In this regard, this study proposes a spatial data warehousing system that can dynamically assimilate the multimodal data and transform to an interpretable representation in real-time. The proposed research explores graph-based representations to provide a uniform format for handling the spatial information of different resolutions, scales and modalities. Raster and vector data will be transformed to a graph-based representation automatically so that advanced learning strategies can be employed to derive intuitions. The novel encoding schemes address the lack of interpretability prevalent in the state-of-the-art deep learning models. It will also enable the real-time prediction, interpolation and analyses, even in the absence of certain modalities/sources. It may be noted that the proposed framework integrates raster and vector data, and the adoption of graph-based strategy makes the processing and analyses domain independent without losing the domain specific characteristics. The proposed generic frameworks will be illustrated for real-time protected area monitoring and also for dynamic landslide susceptibility mapping. Multi-sensor and multimodalities of data, along with the analysis of spectral, spatial and temporal dimensions, are essential in monitoring protected areas and for landslide susceptibility mapping. In this regard, the proposed latent graph generation and graph-based uniform representation strategies ensure real-time integration of multimodal data irrespective of their resolution or formats (being vector or raster). The transformation and graph-based analysis facilitate not only real-time computationally efficient analyses but also make the models interpretable. The explainability and interpretability enhancement strategies will ensure the reliability of the frameworks. Additionally, the novel approach of using a graph-based representation eliminates domain bias and makes the approach generic, ensuring optimal performance even in study areas having significantly different geographies. The main highlights of this research are: i) Generic spatial data warehousing framework; ii) Graph based uniform representation of multimodal data; iii) Real-time transformation through latent graph generation; iv) Domain independent representation of domain knowledge; v) Adaptive real-time interpolation, data assimilation and decision support; vi) Deep learning of graphs for different analyses; vii) Interpretability and explainability based strategies to ensure reliability and explainability of the decisions; viii) Investigations on different novel deep learning based regularizations, architectures, neural units, convolutions, etc. |