THE CURRENT SITUATION
In the era of data-driven decision-making and artificial intelligence, organizations face significant challenges in acquiring, managing, and utilizing high-quality datasets. One critical issue in certain use cases is for sure the acquisition method and the maximum resolution reachable of images to be analyzed by such algorithms.
Traditionally, the monitoring of large infrastructures has relied on manual visual inspections, involving both staff and suppliers, or occasionally through drone surveys or high-resolution satellite data. However, the current process requires significant manpower to collect, verify, process, and share information, leading to many hours of work for all involved parties, generally including the need of physical presence at site.
Additionally, many large infrastructure projects span over extensive areas (often exceeding 300 hectares), which limits the feasibility of comprehensive on-site inspections. This often results in reduced frequency and coverage of spot surveys or necessitates increased work hours, making it almost impractical to survey the entire site using standard methods within a reasonable timeframe.
THE CHALLENGE SOLUTION’S REQUIRED FEATURES AND DETAILS
The goal of the challenge is to prove a super-resolution algorithm capable of maximizing the resolution of images for a specific area of interest, starting from freely available satellite data.
This technology aims to prove high-definition images from low-resolution data, enhancing the capacity for analysis and monitoring of large-scale infrastructure. By leveraging advanced image processing techniques and machine learning, the algorithm should provide improved visual clarity and finer detail (x3 as minimum upscaling factor), minimizing the reliance on costly high-resolution datasets.
The project aims to enhance operational efficiency and the accuracy of assessments in scenarios such as construction site progress monitoring, with the aim of identifying structural components (i.e. wind turbines, PV modules, PV structures, earth movements, …);, vegetation growth monitoring (i.e. proximity to power lines, …); identification of small/medium photovoltaic installations (i.e. covering 500-1.400m2); identification of abandoned / damaged / new buildings in of cities, etc., tracking project progress, and analyzing the evolution of extensive infrastructure. The super-resolution images will be used as an input of computer vision and change detection algorithms, these latter being outside of the scope of this challenge, that will eventually allow to support the abovementioned use cases, tracking evolution of conditions. This will enable more informed and timely decision-making through improved access to visual data.
The developed augmentation algorithm will need to be tested by the participants to this challenge by using reference geographical areas and infrastructure sites for which Enel has existing data available for comparison. These include, but are not limited to, Solar/Wind/BESS Power Plant construction sites, for progress monitoring purpose, urban areas for detection of new photovoltaic installations and building shape & status identification.
The testing phase, that the challenge participants will perform, will involve a detailed analysis of the algorithm's structure and performance by comparing the augmented outputs with the existing baseline data and based on the specific references and key performance indicators (KPIs) outlined below.
The aim is to ensure that the algorithm meets predefined standards of accuracy and reliability, providing valuable insights into its practical applicability for monitoring and analyzing large-scale infrastructure projects the abovementioned use cases.
The validation results will help the challenge winner to refine the algorithm, ensuring its robustness and effectiveness in diverse real-world Enel scenarios during a second phase project which is out of the challenge process.
SOLUTION’S MUST HAVE:
a. TRL>6
b. Input data sources used: the algorithm must exclusively utilize free satellite data sources, such as Copernicus (i.e. Sentinel-2), Cosmo-skyMed, Prisma, Landsat or other similar publicly available datasets.
c. Test Areas: each participant can apply to more than one of the following Use Cases (find attached the KML files of the areas of test):
1. Solar Power Plant Centurion, Italy, for Construction progress on 24/08/23, 08/11/23 e 23/01/24
2. Province of Bologna, Italy, for detection of new photovoltaic installations equal or bigger than 500m2 in the time frame September 2024-December 2024.
3. Municipality of Bitonto, Italy, for abandoned / damaged / new Building recognition and classification on 11 and 12/04/2022
4. A Wind Farm under construction, for Construction progress in Impofu (south Africa) any time (Ground Truth)
5. BESS Pian di Giorgio Farm under construction, for Construction progress on 07/11/24
d. The solution must be validated by the Participant by applying the following “KPI for validation”:
i. Comparison with Reference Data (Ground Truth) respect to dataset coming from survey
ii. PSNR (Peak Signal-to-Noise Ratio): Measures the ratio between the signal and noise, indicating the quality of the reconstructed image compared to the original. A higher PSNR value indicates better image quality.
iii. SSIM (Structural Similarity Index Measure): Evaluates the structural similarity between the original and enhanced images. It considers brightness, contrast, and structure, providing an indication of perceived visual quality.
iv. RMSE (Root Mean Square Error): Measures the average squared error between the enhanced image and the reference image. A lower RMSE value indicates higher accuracy.
v. Processing Time: Measures the time taken by the algorithm to process the image. Efficient software should be able to handle large volumes of data in a reasonable timeframe.
vi. Resource Utilization (CPU/GPU): Evaluates the efficiency of the algorithm in terms of memory usage and computational power.
vii. Spectral Similarity Index: Measures how well the enhanced image retains the spectral data compared to the reference image.
viii. Testing Under Different Environmental Conditions: Evaluates the algorithm's performance on images captured under various conditions (e.g., cloud cover, different seasons) to ensure it can generalize well across different scenarios.
ix. Testing with Different Resolutions and Satellite Datasets: The software should be tested on data from various satellites (e.g., Sentinel-2, Landsat) and different resolutions to assess its versatility.
x. LPIPS (Learned Perceptual Image Patch Similarity):
A deep-learning metric that evaluates perceptual similarity using feature maps from neural networks.
xi. Number of Use Cases treated successfully according to the abovementioned KPIs.
The participant guarantee that any proposed solution featuring an AI technology is compliant with EU and Italian law on AI and with the Regulation of the Challenge.
To accurately evaluate the performance of the algorithm, all the necessary data for calculating the previously mentioned KPIs must be provided at the end of the testing phase. This includes, for example, information such as the number of pixels, bit depth, image dimensions, and other relevant metadata.
DELIVERABLES
Proposals must be submitted to the platform openinnovability.com in a single stage, and must include the following information in English:
- Solution name/title and overview
- A detailed description of the solution and model used
- Demo of at least 5 images in the selected Use Cases- max file upload is 35MB- please provide link to which the evaluation team can access to verify the demo (please take care of the validity period of the link)
- For each processed image, the corresponding source image from the earth observation service used must be provided, including its metadata
- The calculated KPIs for the images and the method of calculation (except for KPI i, x, xi Comparison with Reference Data (Ground Truth)
- A development roadmap to market, indicating the estimated time for design, certification and prototypes
- The budget estimation for industrialization
- A detailed and clear estimation of the final cost of the solution
- Supporting Documentation: Any additional supporting materials, diagrams, simulations, or research that could help in understanding and evaluating the proposed solution
