JULY 9, 2018BY MARK ALTAWEELGIS INDUSTRY
The field of geospatial artificial intelligence, or geoAI, has used many of the same techniques within general artificial intelligence (AI). However, there are both challenges and opportunities that AI has to face in applying geospatial knowledge that also addresses issues of time and spatial bias.
Recently, the GeoAI conference held its first annual meeting in 2017. In 2018, the second GeoAI conference will be held. Key issues of focus include urban and land use dynamics, where the explosion of data has led to a call for better techniques to address challenges in the field.[1]
Large-scale projects that are specifically geospatial and applying AI methods are still relatively few; however, there is now an increased awareness that geospatial diversity, in scale and types of locations, and the role that temporal change and scale have in dealing with applications that help decision-makers and research. One project that has faced such challenges has focused on predicting particulate matter air pollution in Los Angeles using open source GIS data. The effort uses the Pediatric Research Integrated Sensor Monitoring Systems (PRIMS) dataset and OpenStreetMap (OSM) to better predict the relationship between land use and roads in relation to where particle concentration would be at a given time and day. Traffic patterns and urban human activities, over time and based on past behaviour, helped to predict pollution levels before they reach dangerous levels, which can allow the city of Los Angles to issue relevant warnings as needed.[2]
Mapping, Map Interpretation, and Deep Learning Techniques
One challenge has been to develop automated map readers using deep learning techniques that can separate textual information, such as names of places, from map features, including contours. The development of optical character recognition (OCR) allows map readers to understand variation between types of information, such as textual and graphic-based, so that they can be separated and interpreted together, such as naming of features.[3]
DEEP LEARNING TRAINING MAP TEXT SAMPLES. FROM: LI, LIU, & ZHOU, 2018.
Challenges To Using Artificial Intelligences in GIS
There are though a number of challenges that remain despite some successful efforts. One challenge is varying temporal resolution. For instance, modelling and predicting chronic disease patterns that have long latency periods has not been successfully done. In this case, various factors, including long development periods, and multiple physical, environmental factors could mean that existing spatial datasets may not be diverse enough, or even go back far enough in time, to allow forecasts to be easily made.[4] Another challenge is most image data are biased towards English-speaking Western states, limiting AI in application to some countries, as learning datasets are less available.[5]
Other areas of focus have been on enhancing low resolution imagery to improve knowledge awareness for given areas or even historical data purposes. The use of convolutional neural networks (CNNs) has been extended to low resolution satellite imagery and it has been show to improve feature identification as low resolution data could be enhanced with basic input in different spectral bands. This allows such approaches to possibly address the limitation of earlier satellite systems, such as the early Landsat systems, to be enhanced and better utilised for long-term land use change.[6]
While overall, we see progress in feature recognition, enhancement of data, better ways to integrate big data into forecasting applications, among the biggest challenges is dealing with cases of diverse temporal resolution and biases in data. This includes cases where topics that involve latency and different triggers to be evident are a focus for existing geoAI or better understanding spatial patterns in non-English speaking states. Nevertheless, the rapid growth of geoAI suggests such problems could be on the horizon as areas to be addressed in the near future.
References
[1] For more on the GeoAI conference, see: https://udi.ornl.gov/geoai.
[2] For more on the Los Angles effort, see: Lin Y, Chiang Y-Y, Pan F, Stripelis D, Ambite JL, Eckel SP, Habre R.(2017); Mining public datasets for modeling intra-city PM2.5 concentrations at a fine spatial resolution. In: Proceedings of the 25th ACM SIGSPATIAL international conference on advances in geographic information systems. Los Angeles area, CA. p. 1–10.
[3] For more on optical character recognition, see: Li, H., Liu, J., & Zhou, X. (2018). Intelligent Map Reader: A Framework for Topographic Map Understanding With Deep Learning and Gazetteer. IEEE Access, 6, 25363–25376. https://doi.org/10.1109/ACCESS.2018.2823501.
[4] For more on the challenge of temporal and spatial resolution in relation to health science and forecasting using AI, see: VoPham, T., Hart, J. E., Laden, F., & Chiang, Y.-Y. (2018). Emerging trends in geospatial artificial intelligence (geoAI): potential applications for environmental epidemiology. Environmental Health, 17(1). https://doi.org/10.1186/s12940-018-0386-x.
[5] For more on image bias and its effect on geospatial AI, see: https://arxiv.org/pdf/1711.08536.pdf
[6] For more on enhancement techniques using CNN for low resolution satellite imagery, see: Collins, C. B., Beck, J. M., Bridges, S. M., Rushing, J. A., & Graves, S. J. (2017). Deep learning for multisensor image resolution enhancement (pp. 37–44). ACM Press. https://doi.org/10.1145/3149808.3149815.
See Also
Transforming Satellite Imagery Classification with Deep Learning
Orbital Insight Exemplifies How Artificial Intelligence (AI) Can Be Used to Study Images from Space
Future of Mobile Spatial Applications and Image Recognition