Trustworthy Multimodal AI for Agricultural and Environmental Modelling and Forecasting

This project will develop trustworthy and explainable multimodal artificial intelligence methods for spatio-temporal modelling and forecasting of agricultural productivity and environmental risks. It will integrate Earth observation imagery, climate records, soil characteristics, land-use information and agricultural statistics to forecast crop yields, drought conditions and vegetation stress. The aim is to deliver reliable, interpretable forecasts that support agricultural planning, environmental management and climate adaptation.

Agricultural and environmental processes involve complex interactions across locations and seasons. Satellite observations and ground-based records provide complementary information, but their integration is challenging because of differences in resolution, coverage, frequency and quality. AI can capture these relationships, while practical use requires credible uncertainty estimates, understandable explanations and reliable performance under changing conditions. The project will address these challenges through collaborative methodological research and field-relevant validation.

The central research question is how heterogeneous data can be integrated into AI models that produce accurate, explainable and transferable forecasts across agricultural regions and climatic conditions.

The project has five objectives: establish a reproducible framework for integrating environmental and agricultural data; develop statistical, machine learning and hybrid models that capture spatial dependence and seasonal dynamics; incorporate explainability, uncertainty quantification and robustness assessment; validate forecasts across locations, seasons and extreme conditions; and develop a prototype decision-support tool for agricultural forecasting and early warning.

The methodology will begin with data quality assessment covering provenance, missing observations, measurement consistency, licensing and spatial and temporal alignment. Satellite-derived vegetation and land-surface indicators will be combined with rainfall, temperature, soil properties and available crop records. Preprocessing and feature selection will use training data only to prevent information leakage.

Model development will compare seasonal and statistical benchmarks with tree-based learning, convolutional networks, recurrent networks, attention-based architectures and hybrid approaches. Multimodal fusion will assess the contribution of each data source and model reliability when some inputs are unavailable. Mathematical constraints and agronomic knowledge will guide model design where scientifically appropriate.

Explainable machine learning methods will identify the variables, locations and historical periods influencing predictions. Explanations will be evaluated for stability and consistency with domain knowledge, without interpreting predictive associations as causal evidence. Probabilistic methods will generate prediction intervals and risk estimates, with calibration assessed on independent data. Robustness tests will examine sensitivity to missing observations, noisy measurements, extreme weather and geographic or climatic shifts. Performance comparisons across regions and production systems will identify differences in reliability.

Validation will combine chronological testing, rolling-origin evaluation and spatial holdout experiments. Assessment will cover forecast accuracy, uncertainty calibration, performance during extreme events and useful warning lead times. Benchmark comparisons and ablation studies will establish whether multimodal integration and additional model complexity provide measurable benefits. Initial pilots will focus on selected Kenyan agricultural areas, with comparative locations agreed by partners according to data access and research priorities.

The cooperation will combine expertise in mathematical modelling, statistics, trustworthy AI, explainable machine learning, Earth observation and agriculture. South Eastern Kenya University could contribute statistical and hybrid forecasting methods, Kenyan case studies, contextual interpretation and stakeholder engagement. Collaborating institutions could contribute multimodal learning, geospatial processing, uncertainty-aware modelling and comparative validation. Responsibilities, data access and pilot locations will be agreed jointly.

Expected outputs include reproducible data integration workflows, validated forecasting models, a trustworthiness evaluation framework, joint publications and a prototype decision-support tool. The cooperation will also support postgraduate supervision, researcher exchanges and training in geospatial analytics and responsible AI. The project aims to strengthen agricultural planning, early warning and evidence-based climate resilience. contact me: rmutwiri@seku.ac.ke