Monitoring UK national freshwater with deep learning


UK freshwater bodies are under increasing pressures from growing populations, increased usages and urbanisation. This has become apparent in the summer of 2018 with hosepipe bans in North West England and Northern Ireland being imposed as a result of very low levels in freshwater reservoirs. Monitoring and managing these reservoirs at a UK national scale is a considerable challenge that can be very labour intensive. Even for basic parameters such as water body size, the establishment of a comprehensive, time-dependent, database for all UK water bodies of any meaningful size remains an unattainable goal. However, given that water body dimensions can easily be established from aerial and high resolution satellite imagery, the raw data required for this task exists in very large volumes and is very often freely available. The crucial limiting factor lies in the identification of the water bodies from large data sets of imagery which fall within the remit of what is now termed ‘Big Data’. This project proposes to apply AI technology to the problem of freshwater body identification (lakes and rivers) in the UK. The project has 3 objectives:
1- Mine the DIGIMAP database for colour imagery from the years 1996 to 2018 and train an AI image classifier to identify freshwater bodies over the entire UK.
2- Mine the freely available (for research) high resolution imagery made available by the startup (funded by Google) for recent imagery at very high temporal resolutions (~weekly images)
3- Ground truth and validate recent imagery with high resolution drone surveys.
4- Acquire declassified USGS (formerely from the CIA) greyscale image data of the UK,  from the 1960s, 1970s and 1980s and additional imagery from the RAF archives spanning the period of 1945-1965. Train a second AI to identify surface water features in historic imagery.
5- Use the outputs of objectives 1 to 4 in order to establish quantitative metrics for the evolution of  UK lakes and rivers over the last 50 years.

Click on an image to expand

Image Captions

Figure 1. Imagery of Fort Williams from 1945 ad 2018. Examination of water boundaries shows change over time, but identification of these boundaries is a complex task.


The project will rely on Artificial Intelligence in the form of image recognition. Both Machine learning and Deep learning will be investigated in order to develop a reliable image classifier that can approach human performance in the identification of water in national-scale time series of imagery despite changing colours and shades as a function of time. It is likely that this will require the usage of an advanced convolutional neural network trained with the specific features that can be seen from the air such as houses, fields, and, crucially, lakes and rivers. Several advanced AI image classifiers do exist, but these tend to be trained by the images now prevalent on social media. As a result, these networks are not well tuned to the classification of features from an aerial perspective.
Data will be sourced from the DIGIMAP national databases for the period 1996-2018. The project will also use contemporary high temporal frequency data from the startup. Greyscale archival imagery from the RAF archive, freely available on Google Earth, and the USGS for the period of 1962 – 1984 will procide data for the past 7 decades. Ground calibration and local validation of classification results will be done with high resolution drone surveys over selected sites in England and Scotland. Theses low-altitude surveys will focus on water body edges and cases where there is an established seasonal change of colour, and on the question of detection limits in terms of size (i.e. the smallest detectable water body, or narrowest detectable river)

Project Timeline

Year 1

• Construction of the data mining architecture for Digimap and data
• Preliminary tests with Machine and deep learning image classifiers for colour imagery.
• Drone piloting and safety course
• Acquisition of archival spy satellite imagery from the USGS.

Year 2

• Drone operations for validation
• Preliminary tests with machine and deep learning image classifiers for greyscale archival imagery.
• Deployment of automated data mining algorithms for ‘Big Data’ image acquisition at national scale.

Year 3

• Large scale processing using Durham University Conputer Science Department GPU stack facility.
• Develpment of a regional indices that summarise the evolution of waterbody area at a high temporal resolution, for the last 50 years.

Year 3.5

• Thesis Production

& Skills

The student will be trained for a full CAA-approved drone safety and piloting course in order to be eligible for Durham University insurance covering drone operations. Student will then be trained in advanced methods of machine learning and deep learning, including Convolutional Neural Networks. This will include funded attendance at the Deep Learning summer school organised by the British Machine Vision Association between year 1 and year 2.

References & further reading

From On-Road to Off: Transfer Learning within a Deep Convolutional Neural Network for Segmentation and Classification of Off-Road Scenes (C.J. Holder, T.P. Breckon, X. Wei), In Proc. European Conference on Computer Vision Workshops, Springer, pp. 149-162, 2016

Zhang, L., Zhang, L., & Du, B. (2016). Deep learning for remote sensing data: A technical tutorial on the state of the art. IEEE Geoscience and Remote Sensing Magazine, 4(2), 22-40.

Jean, N., Burke, M., Xie, M., Davis, W. M., Lobell, D. B., & Ermon, S. (2016). Combining satellite imagery and machine learning to predict poverty. Science, 353(6301), 790-794.

Zou, Qin, et al. “Deep Learning Based Feature Selection for Remote Sensing Scene Classification.” IEEE Geosci. Remote Sensing Lett. 12.11 (2015): 2321-2325.

LeCun, Yann, Yoshua Bengio, and Geoffrey Hinton. “Deep learning.” nature 521.7553 (2015): 436.

Further Information

Dr Patrice Carbonneau:

Apply Now