Explainable AI for UK agricultural land use decision-making

点击翻译按钮获取中文摘要

Agricultural land use dynamics and their associated driving factors represent highly complex systems of flows that are subject to non-linearities, sensitivities, and uncertainties across spatial and temporal scales. They are therefore challenging to represent using traditional statistical modelling approaches. Existing process-based modelling has enabled advances in understanding of individual biophysical processes underpinning agricultural land use systems (e.g. crop, livestock and biogeochemical models). However, these tend to focus on individual processes in detail or link a limited number of processes at large scales, thereby mostly ignoring the complex interdependencies between the multiple interacting biophysical and socio-economic components of land use systems. Artificial intelligence (AI) techniques offer great potential to complement such modelling approaches by mining the deep knowledge (e.g. farming patterns and behaviours) encapsulated in 'big' data from ground-based sensors (such as frequently used for precision farming) and Earth Observation satellites. This will deliver enhanced insight on the past and current state and spatio-temporal dynamics of agricultural land use system flows and how they can be influenced by decisions on agricultural policies and related farm management practices. Our proposal aims to develop a novel explainable AI framework that is transparent, data-driven and spatially-explicit by using probabilistic inference and explicit "if-then" rules. We will demonstrate proof-of-concept for two pilot regions of the UK (Oxfordshire and Lincolnshire), and the framework will be set up in a way that can be readily expanded to the whole UK. Specifically, we will draw on time-series of agricultural land use and production datasets (in-kind support from industry project partner SOYL) to identify the key socio-economic and environmental driving factors that have led to historic agricultural land use changes in the pilot regions. We will then establish explainable AI-rules for the characterisation of these agricultural land use changes and refine them within the framework through machine learning and parameter optimisation.We will demonstrate and test the potential of the explainable AI framework for providing a new and robust method for predicting changing patterns of agricultural land use in the two pilot regions. This will include testing the ability of the AI framework for improving understanding of past and present agricultural land use dynamics across multiple temporal and spatial scales from 'big' data. It will also assess the potential for continually updating the predictions of land use dynamics in real-time using data from sensors. This could provide early warning when certain driving conditions are triggered or used to repeatedly refine short-term projections of land use change and their estimates of uncertainty.