Advancing Satellite-Based Crop Monitoring with Machine Learning and HPC – Collaboration with industrial (end) users

NCC BULGARIA

NCC presenting the success story:

NCC Bulgaria is the National Competence Centre for HPC and HPDA, operating within the EuroCC3 network and EuroHPC ecosystem. Coordinated by the Institute of Information and Communication Technologies at the Bulgarian Academy of Sciences (IICT-BAS) in partnership with the Sofia University “St. Kliment Ohridski” (SU) and the University of National and World Economy (UNWE), it supports industry, academia, and public organizations through specialized training, consulting, technical expertise, and access to national and European HPC resources.

Companies involved

Utilink Ltd. (https://utilink.xyz/ ) is a machine learning platform that integrates model training, fine-tuning, and deployment. It provides scalable computing resources, no-code tools, and solutions for businesses, developers, and researchers. The platform supports applications in healthcare, agriculture, education, transportation, and other sectors.
Utilink provides a platform for HPC resource management, job scheduling and monitoring, as well as AI model hosting and inference through its Model Hub. In this collaboration, Utilink connected the researcher’s workflows with the HEMUS infrastructure and enabled rapid model testing.

Technical/scientific Challenge

Sentinel-2 and Landsat 8/9 observe the same farmland at different spatial resolutions (10 m versus 30 m) and spectral bands, making consistent crop monitoring difficult. Conventional harmonisation adds processing steps and can compromise detail. The challenge was to estimate Leaf Area Index (LAI) with one sensor-aware neural network despite scarce field measurements and imbalanced satellite datasets. Preparing thousands of image tiles, repeating GPU-intensive training, and validating predictions across regions and seasons required HPC resources.

Solution, thanks to this collaboration: The researchers developed a multi-sensor deep learning model based on a U-Net architecture with sensor-specific adaptation and multi-scale feature extraction. Satellite imagery from Sentinel-2 and Landsat 8/9 was processed at its native resolution, while vegetation indices provided training labels where field measurements were limited. The HEMUS supercomputer at IICT-BAS supplied NVIDIA A100 GPUs and storage for model training, inference, and repeated experiments. Utilink enabled efficient job submission, scheduling, and monitoring on HEMUS, while its Model Hub supported rapid testing of trained models. A five-level evaluation framework assessed prediction accuracy, physical consistency, and generalisation across unseen regions and seasons. What did you manage to do together that you could not have done separately? The collaboration combined scientific expertise from IMI-BAS, high-performance computing (HPC) resources from IICT-BAS, and Utilink’s HPC management platform. HEMUS provided the GPU capacity and storage required for repeated training, large-scale satellite processing and validation, while Utilink simplified experiment scheduling, monitoring and model deployment. Together, the partners developed and evaluated a multi-sensor neural network that produced consistent Leaf Area Index maps from Sentinel-2 and Landsat 8/9 without separate data harmonization. The model achieved lower prediction errors than the tested single-sensor approaches and vegetation indices and demonstrated generalization to unseen regions and seasons, supporting a practical path from doctoral research to reusable crop-monitoring tools. Validation against direct field measurements remains ongoing.

Benefits: 

  • More frequent crop monitoring using two free satellite missions.
  • Faster GPU-accelerated research through HEMUS and Utilink.
  • Reuse the model in new regions without retraining.
  • Rapid testing through the Utilink Model Hub.

Success story # Highlights:

  • Keywords: High-Performance Computing; Machine Learning; Deep Learning; Remote Sensing; Sentinel-2; Landsat 8/9; Leaf Area Index (LAI); Precision Agriculture; Satellite Data Processing.
  • Technology: HPC (HEMUS Supercomputer); NVIDIA A100 GPUs; Deep Learning; U-Net Neural Networks; Multi-Sensor Data Processing; Utilink HPC Management Platform; Model Hub.
  • Industry sectors: Agriculture; IT/HPC Systems, Services & Software Providers.

Figure 1.Research workflow: satellite data preparation, model development, GPU training, inference, and evaluation using HEMUS and Utilink Model Hub

Contact: