Assessing food production systems using machine learning and remote sensing

Understanding and monitoring global agricultural activity, especially in the face of a rapidly changing climate, has applications ranging from food supply predictions to subsidy allocation and environmental monitoring. Rapid advancements in computer vision and satellite imagery acquisition present opportunities to extract planetary knowledge automatically and at large scale. Satellites provide global coverage at relatively low marginal cost. However, many regions lack ground truth labels essential for training machine learning models, which remains a significant challenge, particularly in developing countries. Bridging this label gap is critical for effective agricultural monitoring and decision-making.

This thesis first explores self- and semi-supervised learning strategies for crop type mapping:

The first approach involves using generative joint energy-based models to improve feature extraction and model performance with limited labelled data. These models can efficiently handle limited labelled data by learning better feature representations from the abundant unlabelled data. This is achieved through a domain-agnostic approach, meaning the techniques are not tailored to specific types of data and can be applied broadly across various remote sensing datasets.

The second approach employs multi-task learning, incorporating physical and biological characteristics of crops derived from the PROSAIL radiative transfer model. By leveraging this domain-specific information, the model gains a deeper understanding of the crop growth processes, leading to improved generalisation across different crop types iand environmental conditions. This allows the model to simultaneously learn multiple related tasks, enhancing its ability to capture complex interactions related to crop types and traits within the data.

The final chapter extends these methodologies to practical applications, focussing on cover crop mapping. Cover crops play a crucial role in sustainable agriculture by improving soil health, reducing erosion, and enhancing crop yields. By creating detailed cover crop maps, this research assesses the effectiveness and impacts of cover crops on agricultural productivity. In this chapter we use causal machine learning to estimate impacts on net primary productivity (NPP). This research provides valuable insights into sustainable agricultural practices. This integration of machine learning, remote sensing, and causal inference techniques offers a comprehensive toolset for assessing and optimising future food production systems.

This research also presents EngScotCrop, a new benchmark crop classification dataset, a large-scale open-access dataset of multimodal satellite image time series alongside agricultural parcel boundaries (UKFields). This thesis aims to inspire further research into increasingly accurate agricultural maps at larger spatial scales, supporting sustainable development.