Problem Description
Distribution Network Operators (DNO) manage thousands of buildings across the UK. Each building houses critical equipment essential for reliable energy transmission and distribution in the region. Poorly maintained roofs can lead to water leaks, clogged gutters, pooling water, vegetation-related damage, etc, leading to energy service interruptions and costly repairs.
A challenge for DNOs, having thousands of buildings, is that sending a crew to inspect each of these buildings is expensive and time-consuming. Surveying roofs requires specialised equipment and on some occasions roofs are inaccessible requiring drone flights.
Specific Challenges a DNO will face:
Regulations
Using drones for inspections in urban areas is complicated due to licensing issues
Logistics
Surveying roofs requires heavy ladders and specialised equipment, potentially posing health and safety risks
Cost
Inspecting thousands of buildings is expensive
Our Solution
To remotely assess substation roof condition for DNOs, Keen AI developed and tested various computer vision models using high-resolution aerial imagery.
We specifically focused on determining four roof characteristics as shown below:
- Roof type
- Roof material
- Roof vegetation analysis, and
- Roof condition
Input Aerial Imagery, AI Model and Performance
We tested a number of data sources including high resolution airborne RGB imagery at 10 and 12.5cm spatial resolution, solar fluxes and Digital Surface Model (DSM) information obtained from Google. We also tested very high resolution optical satellite data from the Pleiades-Neo satellite, which after pansharpening provides 30cm RGB imagery plus the deep blue, red edge and Near-infrared (NIR) spectral bands.
Given the nature of this problem where an AI model needs to derive multiple insights from the same building (i.e. roof type, material, condition and presence of vegetation), we concluded that a single AI model able to derive multiple roof characteristics performed better than multiple individual models focusing on deriving a specific roof characteristic.
Figure 4: Schematic representation of a multitask deep learning model
This agrees with existing literature for similar problems and is known as multi-task deep learning. Some advantages of multi-task learning include models with better generalisation (i.e. better performance on previously unseen data) and consistency and coherence in predictions across tasks since the same model produces all outputs. See more in-depth aspects of multitask learning here [1].
The trained model shows good performance for each of the roof characteristics assessed when using airborne imagery. While we also tested the use of Solar Fluxes and DSM imagery as model inputs, the obtained results did not show notable improvements compared to RGB only. Using satellite imagery we found that the NIR bands may help to identify the presence of vegetation. However, the off-Nadir angle (angle between satellite nadir and the roof of interest) in the archived satellite data we obtained was not appropriate in some cases as it was too high. This caused some of the roofs of interest to be obscured by nearby buildings.
Solution & Business Benefits
Our roof monitoring solution provides significant benefits for DNOs:
- Cost Savings:
- Reduced Inspection Costs: By using remote sensing technology and AI analytics, DNOs can avoid the high costs associated with sending crews to inspect every building.
- Eliminated Drone Flight Expenses: The need for costly and logistically challenging drone flights in urban areas is minimised.
- Enhanced Safety:
- Lowered Risk of Accidents: Reduced need for physical inspections, thereby decreasing the risk of accidents that can occur during traditional roof inspections involving ladders and other equipment.
- Improved Efficiency:
- Priority-Based Maintenance: It is possible to conduct periodic surveys and generate alerts, helping DNOs to prioritise roofs that require urgent attention over those that are in good condition. This ensures efficient allocation of maintenance resources.
- Proactive Maintenance Planning: With continuous remote monitoring, DNOs can move from reactive to proactive maintenance, addressing potential issues before they escalate.
- Comprehensive Reporting: Regular updates and comprehensive reports help DNOs stay informed about the state of their infrastructure, ensuring they can plan and budget effectively.
- Operational Continuity:
- Minimised Service Interruptions: By maintaining roofs in good condition, the risk of equipment damage and service interruptions due to water leaks and other issues is significantly reduced.
- Extended Equipment Lifespan: Properly maintained roofs protect the critical equipment housed within the buildings, potentially extending the lifespan of this expensive infrastructure.
By leveraging the combination of airborne imagery and AI, DNOs benefit from a safer, more cost-effective, and efficient approach to maintaining the roofs of their buildings. This solution not only supports operational continuity but also enhances their ability to manage and protect their critical assets.
Reduce Inspection Costs with Satellite Data and AI.
References & Credits
[1] Crawshaw, M. (2020). Multi-task learning with deep neural networks: A survey. arXiv preprint arXiv:2009.09796.
Featured image courtesy of Airbus Defense and Space. Pleaides-Neo, 30cm.
