The session recording will be archived on the site until June 25th, 2023

This session was streamed but not recorded

Curated Track: 

Description:

Advances in machine learning (ML), specifically specialized deep learning architectures of convolutional neural networks (CNNs), Recurrent Neural Networks (RNNs) and Graph Neural Networks (GNNs), revolutionized the methodology in geospatial sciences and remote sensing. Building on the knowledge developed by ML researchers and practitioners, Geospatial Artificial Intelligence (GeoAI) incorporates spatial and temporal dependence in observations for improved predictive performance in various application domains including, for example, damage assessment, urbanization characterization, population and epidemiologic dynamics, transit forecasting, change detection, maritime safety, land cover and land use, geological mapping, and large scale modeling of climate variables.
While artificial intelligence (AI) and GeoAI are revolutionizing predictive modeling in the field, there are several challenges in incorporating these models in the geospatial domain. While semi-supervised learning and transfer learning mechanisms allow training on smaller labeled samples, still, lack of reliable and AI-ready labels in most application areas creates a barrier in adopting these methods. The size, dimension and volume of geospatial data layers may also create more burdens in terms of data pre-processing, model complexity, size, and memory requirements. Additionally, the trustworthiness of ML algorithms is further stressed by the necessity to rely on interpretative and subjective labels, difference in acquisition systems, and sparsity of the data and of ground truth. Moreover, the deployment of GeoAI models tends to be more challenging than traditional ML models because of the necessity to have matching spatial coordinate systems, and the necessity of users to access large datasets in remote locations and unreliable connections. Understanding how to further extend traditional ML model deployment strategy to accommodate GeoAI requirements creates several important research questions for the spatial data science community.
In this session, we welcome submissions broadly contributing to the research on GeoAI deployed for geospatial datasets and remotely sensed data, algorithm development, and strategies for model deployment, as well as the consideration of geospatial properties for model output analysis.

Topics include (but are not limited to):
-Label ambiguity
Training with non-conventional, non-AI ready labels
Training with ambiguous labels
Training with sparse or noisy labels
-Scale challenges
Model architectures adapted to multi-scale data
Multi-scale spatial processes
Moving objects spanning multiple training samples
-Algorithm and data processing for GeoAI
Geographically inspired model architectures
Raster and vector use strategies for ML systems
Data access, pre-processing, and processing pipelines
Data probability distribution and transformations for better performance
Use of geospatial data with generic pre-trained models
-Temporal drift and geographical biases
Temporal model drift and update strategies
Algorithm development for large areas
Geospatial and temporal generalizability
-Interpretation and geospatial evaluation systems for end users
Human-in-the-loop training and evaluation
Uncertainty modeling of ML outputs based on input uncertainty
Geospatially aware performance evaluation metrics
Confidence interval analysis for GeoAI products

Non-Presenting Participants