Deep Learning Frameworks for Geospatial Object Detection
These researchers are using MSI for seven projects related to spatial data:
- Foundation models for vehicle data. Foundation models, defined as AI models trained on broad data that can be adapted (e.g., domain-specifically fine-tuned) to a wide range of downstream tasks, have the potential to acquire knowledge from real-world data and provide results more efficiently. Trained by a surrogate task (e.g., drive cycle simulation), foundation models' advanced pattern recognition capabilities can be leveraged to predict traffic flows, vehicle emissions, etc. with higher accuracy, drawing from both real-world and synthetic data that encompasses a wide range of driving conditions and behaviors. The synthetic data for a small set of vehicles and roads to train foundation models may be generated from physics-based simulators, such as Simulation of Urban MObility (SUMO), which simulates vehicle drive cycles based on physics models, taking into account road features, vehicle features, and traffic demand. These simulated drive cycles can then be used with FASTSim to generate vehicle parameters such as energy use or emissions. The trained foundation models are likely to be faster than traditional physics simulation (e.g., SUMO, FASTSim) for generating forecasts for a richer collection of roads and vehicle types. MSI resources are needed to train these foundation models due to their large nature, considering both the extensive training data required (e.g., data from different driving conditions) and the size of the models themselves (e.g., models with as many as billions of parameters).
- Reducing uncertainty in sea-level rise prediction. Given multi-model ensemble climate projections and other explanatory variables (e.g. ice sheet melt from glaciers, from Greenland and Antarctica, land water storage, and stereo dynamic ocean), the goal of the project is to accurately and reliably predict future sea-level rise while lowering the uncertainty. This problem is important because sea-level rise affects millions of people in coastal communities. This problem is challenging due to spatial variability and unknowns such as possible tipping points (e.g., collapse of Greenland or West Antarctic ice-shelf), climate feedback loops (e.g., clouds, permafrost thawing), future policy decisions, and human actions. Most existing sea-level rise modeling approaches use the same set of weights globally, during either regression or deep learning, to combine different explanatory variables to project future sea-level. Such approaches are inadequate when different regions require different weighting schemes (e.g., the weights for polar ice sheets depend on the distance from the ice sheets) for accurate and reliable sea-level rise predictions. This project aims at building spatial-variability aware models by utilizing the projections from multiple simulation models and explanatory variables to accurately predict the future sea level rise and reduce the uncertainty. This is a computationally intensive task, requiring access to high-performance computing resources such as GPUs provided by MSI.
- Co-traveler detection for coordinated trip generation based on spatiotemporal co-occurrence pattern mining, human trajectories, and time-expanded graphs. Coordinated trips refer to a collection of journeys by a set of travelers with common destinations, routes, or events. Coordinated trips can be further categorized into normal coordinated trips and anomalous coordinated trips. Normal coordinated trips are coordinated trips that are not harmful to society, such as family outings, get-togethers, and carpools. Normal coordinated trips can save energy, such as carpooling, or make events more exciting and appealing, such as flash mob dancing. Anomalous coordinated trips are coordinated trips that cause societal damages, such as the group of unruly teenagers who harassed people in Dinkytown in May 2023 and flash mob robberies in Los Angeles. Promoting normal coordinated trips and preventing anomalous coordinated trips are essential for society’s safety and welfare. Generating coordinated trips requires us to infer social relationships and detect co-travelers with common destinations, routes, etc. This project will tackle these tasks using points of interest (POIs) taxonomy and human trajectories. POIs are geographic locations that are of particular interest for specific purposes. In this project, POIs include restaurants, residences, schools, etc. The human trajectory data contains a varying number of locations and timestamps of each agent’s movements. This project has used nearly a million agents with daily trajectories for a year. The size of the data covering trajectories in the US in 2019 is 5TB. To infer relationships between agents, we need to mine spatiotemporal co-occurrence patterns, which requires pairwise comparison and matching. A hotspot detection to infer POIs is also necessary to determine the trajectory semantics. The extensive data size and high computation intensity of this project require the use of MSI resources.
- Identification of aberration patterns (i.e., behavior showing significant deviation from the normal) using a large number of multi-attribute trajectories. Identifying such aberration patterns can help improve maritime security and prevent illicit activities (e.g., illegal fishing, illegal oil transfer to violate United Nations sanctions) where the involved objects may hide their movement by deliberately not reporting their locations. The project uses multi-attribute trajectory data (MTD), which consists of various attributes (e.g., drought, rate of turn, emission). The challenges of this problem arise from the complexity of modeling gaps and large amounts of data. The project includes two computationally intensive tasks that require MSI resources:
- Detecting potential rendezvous patterns via Spatiotemporal (ST) joins based on two or more modeled gaps, resulting in high exponential gap enumeration costs for millions of trajectory gaps.
- Conducting data slicing and refinement of the resultant shape derived from the intersection of ST joins at a finer granular level providing a relatively tighter ST bound and, thereby, reducing the manual post processing effort done by human analysts.
- Identification of spatial patterns (e.g., colocation) of cell interactions (e.g., immune and tumor) to help distinguish between responder and non-responder tissue samples (e.g., clinical outcome to immunotherapy). Most of the related works to identify similar spatial patterns are limited to hand-constructed features using traditional spatial association measures (e.g., Ripley’s cross-k, G-cross, etc). However, these may not be sufficient in capturing the relevant measures of spatial interactions (e.g., directional spatial relationships) among tumor and immune cells such as surrounded. To overcome these limitations, these researchers have been exploring the effectiveness of AI-constructed measures with the help of novel GeoAI deep neural network techniques, namely, spatial-interaction aware multi-category deep neural network (SAMCNet), to go beyond the hand-constructed features. Cellular maps derived from multiplex immunofluorescence (MxIF) imagery contain over 60 different cell types, resulting in a million trillion (260) potential spatial colocation patterns to explore. Thus, MSI resources are used to help to construct input to SAMCNet, which itself contains millions of parameters required to be trained. MSI resources are needed to refine SAMCNet by considering spatial variability and training the model across different regions (e.g., tumor-core and tumor-interface) that result in learning million to billion parameters.
- Mining engine data and trajectories to find environmentally-friendly paths. Given a road network, an origin, a destination, and onboard diagnostics (OBD) data, the energy-efficient path selection problem aims to find the path with the least expected energy consumption. Taking energy-efficient routes instead of the fastest route can help avoid over one million metric tons of carbon emissions every year. To find the energy-efficient path, the researchers will first leverage MSI resources to perform the data preprocessing task to process large volumes of truck OBD data provided by Volvo (e.g. map matching, data cleaning, etc.), and to train a novel physics-informed neural network for energy estimation on road segments in subpaths to take the contextual information into consideration. A city road network (e.g. 70 miles around Minneapolis) could contain hundreds of thousands of road segments and tens of millions of subpaths, so both the time cost and storage cost of calculating the energy-efficient paths in a city road network are high, necessitating the use of MSI resources. This work will be leveraged further to find environmentally-friendly paths for electric vehicles (EVs) to reduce the carbon footprint. The carbon emission data set will be retrieved from WattTime API, which collects and aggregates historic nationwide emission information. This problem is computationally intensive and storage intensive because of the spatial-temporal variability within the carbon emission rate of EV charging stations and the large number of candidate paths in a road network. For example, for the worst-case scenario where the network graph is complete, each permutation of intersections except the origin and the destination is a potential path (e.g. given a complete network graph with only 20 intersections, there are 6*1015 candidate paths between an origin-destination pair).
- Detecting statistically significant regional colocation patterns. This problem has applications in ecology, economics, sociology, etc. In any relevant spatial dataset, there can be an exponential number of candidate patterns and an exponential number of candidate regions which makes algorithms such as colocation pattern mining very expensive. For example, the Safegraph POI dataset has locations of 1,473 different retail brands which can result in 21,473 potential colocation patterns. Adding statistical significance to the colocation detection process ensures that the identified patterns did not occur by chance. Significance testing adds extra computational cost to the algorithm since it is necessary to generate hundreds of null hypothesis datasets for each participating feature to model their distribution under complete spatial randomness, making it necessary to use resources that can handle such intensive computations.
Mr. Jayant Gupta
Professor Shashi Shekhar