Explainable AI and Foundational Models for large-area crop type and yield mapping in data-scarce environments
Programming skills are required for this topic.
Note: The image used in the topic description was generated using AI.
In many parts of the world, agricultural monitoring is hindered by limited field data, inconsistent surveys, and highly heterogeneous landscapes. Traditional crop mapping and yield estimation methods often rely on extensive, high-quality ground truth, which is largely unavailable in data-scarce regions. This MSc research proposes a novel framework that leverages foundational Earth observation models together with Explainable AI (XAI) to map crop types or estimate yields across large areas while minimizing dependence on labelled datasets. The study will investigate how pretrained, general-purpose EO models can be adapted to extract meaningful temporal and environmental patterns from satellite time-series and spatial covariates when only sparse ground data are available. A key contribution is the integration of XAI tools to interpret model ‘behaviour’, revealing the spectral, temporal, and environmental features driving predictions and enabling transparency in a domain where trust, accountability, and verifiability are essential for policy and food-security planning.
Objectives:
- To develop a data-efficient AI workflow, supported by a foundational EO model, capable of mapping crop types or estimating yields in regions with limited ground truth.
- To evaluate techniques such as semi-supervised learning, transfer learning, and weak supervision to reduce dependence on labelled samples.
- To integrate Explainable AI tools to interpret model predictions and identify the most influential environmental, spectral, and temporal features.
- To scale the workflow for regional or national-level crop monitoring in heterogeneous and data-scarce agricultural landscapes.
Methodology:
The study collects multi-temporal satellite imagery and environmental datasets and preprocesses them into consistent temporal and spatial inputs suited for large-area analysis in data-scarce regions. A foundational Earth observation model is adapted to the selected study areas using data-efficient techniques such as semi-supervised learning or weak supervision, enabling it to learn crop or yield-related patterns from minimal labelled data. The adapted model is then applied across the full region to generate large-scale crop type or yield maps, with performance evaluated against the limited available ground observations. Finally, Explainable AI methods are used to interpret model behaviour, revealing the spectral, temporal, and environmental drivers of predictions and identifying uncertainties in areas lacking ground truth.
Note: the topic will focus either on crop or yield mapping
Mai, G., Xie, Y., Jia, X., Lao, N., Rao, J., Zhu, Q., Liu, Z., Chiang, Y.-Y., Jiao, J., 2025. Towards the next generation of Geospatial Artificial Intelligence. International Journal of Applied Earth Observation and Geoinformation 136, 104368.
Zhang, C., Kerner, H., Wang, S., Hao, P., Li, Z., Hunt, K.A., Abernethy, J., Zhao, H., Gao, F., Di, L., Guo, C., Liu, Z., Yang, Z., Mueller, R., Boryan, C., Chen, Q., Beeson, P.C., Zhang, H.K., Shen, Y., 2025. Remote sensing for crop mapping: A perspective on current and future crop-specific land cover data products. Remote Sensing of Environment 330, 114995.
This topic aligns with SIS by requiring the collection, processing, and uncertainty-aware analysis of multi-sensor satellite data and spatial covariates, as well as the validation of large-area crop or yield maps derived from foundational models. It contributes to TE because it involves understanding and modelling environmental and biophysical processes that influence crop distribution and productivity. The interpretability and policy relevance of the output maps connect to SPG, supporting evidence-based decision-making for agricultural agencies, governments, and food-security stakeholders operating in heterogeneous and data-scarce landscapes.