- Increasing spatial resolution of satellite imagery makes it possible to monitor environmental changes from space automatically using artificial intelligence
- Google’s TensorFlow AI software will be applied to Copernicus Sentinel-1, Sentinel-2 and Planet satellite imagery
- Expected outcomes include the automated detection of land use changes such as creation of wind energy parks, golf courses and footpaths in national parks.
Recent advances in computing technology, cloud computing and high-performance computing are paralleled with those in advanced artificial intelligence (AI) algorithms and significant investment in the European Copernicus Earth Observation programme and its Sentinel satellite missions. AI enables automatic detection of spatial patterns in environmental data such as satellite images based on training data. The paradigm of looking for spatial patterns instead of the historic focus on spectral information in satellite imagery allows the detection of new types of land cover and land use (Comber et al. 2016) such as golf courses, whose spectral signatures resemble other land covers such as pastures but whose spatial patterns are very characteristic.
Machine learning / AI (Le Cun et al. 2015) have previously been applied to hyperspectral image classification (Hu et al. 2015), CORINE land cover mapping from Sentinel-1 SAR images (Balzter et al. 2015), providing evidence of slavery from WorldView satellite images (Boyd et al. 2018).
This interdisciplinary studentship (Department of Mathematics and School of Geography, Geology and the Environment) aims to explore the application of AI software (AIS) to operational automated land use monitoring, focusing on UK national scale. Time-series stacks of 10 m spatial resolution Sentinel-2 images and 3 m Planet images will be input into the AIS. The AIS will be trained based on visual interpretation-based land cover maps of the UK produced at Leicester (CORINE land cover maps 2012 and 2018).
- How accurately can an AIS be trained on the CORINE land cover map 2012 identify land cover / land use classes from new satellite imagery for 2018 (validated with visually interpreted CORINE 2018 map)?
- Are there particular types of land use change that AIS can accurately predict that otherwise require labour-intensive visual interpretation?
- Can future land cover map updates be automated aided fully or in part by AIS?
Methods will be drawn from Mathematical Modelling, AI and Earth Observation and Geography. The TensorFlow AIS (Abadi et al. 2016) will be implemented on the High-Performance Computing facility SPECTRE-2 at Leicester and linked with an existing 10 m resolution Sentinel-2 image processing chain developed in Python (https://github.com/clcr/pyeo). It will also be adapted to ingest 3 m resolution Planet imagery. Training data for the AIS is available from the CORINE 2012 map (Cole et al. 2018). Once trained up, the AIS will identify land use patterns from new satellite data from 2018 and assess land use changes in the UK. The outcomes of this analysis will be validated against the recently completed CORINE 2018 map produced at the University of Leicester (PI Balzter). Experiments with different spatial filters and configurations of TensorFlow will be undertaken to optimise the detection of particularly difficult land use types (e.g. wind farms and golf courses).
Training and Skills
The student will be trained in Sentinel data processing on the HPC facility SPECTRE-2 at Leicester, developed in the School of Geography, Geology and the Environment. The student will take the new MSc module GY7709 (Satellite Data Analysis in Python), available in 2019-2020, and any other modules deemed suitable, dependent on the background of the student. Complementary individual training in using AI, especially TensorFlow, will be available from the Department of Mathematics. Further training will take place ‘on-the-job’ as part of the research team in the Centre for Landscape and Climate Research.
Year 1: Literature review, refinement of research questions and work plan, liaison with and consultation of project partners, installation of AIS on HPC, training in satellite data processing in Python
Year 2: Training the AIS, completing test runs with Sentinel-2 data and evaluating outcomes, iteratively refining data flow and accuracy, running AI ‘experiments’
Year 3: Implementing Planet imagery AIS runs, evaluating and comparing results, submitting two papers for publication
Partners and collaboration (including CASE)
- Planet – satellite assembly, operations and data analytics company
- Specto Natura – consultant company in Cambridgeshire specialising in land cover mapping
- Centre for Ecology and Hydrology, Lancaster – NERC research centre with UK land cover expertise
- Google – London office has expressed interest in projects around AI and satellite data
Informal enquiries are welcome to Prof. Heiko Balzter, firstname.lastname@example.org