Project Summary

• Developed and tested a U-Net deep learning model for the automatic detection of Small Agricultural Reservoirs (SmARs) from Sentinel-1 SAR imagery in Tuscany.
• Built the workflow on a pre-existing SmAR mapping dataset, using reference masks derived from the 2016 LaMMA dataset and refined with Sentinel-2 data from 2021 to focus on reservoirs effectively containing surface water.
• Prepared a dataset of 5193 image patches (256 × 256 pixels), split into training (60%), validation (25%), and test (15%) subsets to support model development and evaluation.
• Assessed the potential of SAR data for reservoir detection in conditions where optical imagery is limited by cloud cover and variable illumination.
• Obtained promising preliminary results, showing that the model can correctly identify small reservoirs in complex agricultural landscapes and providing a methodological basis for future developments on temporal monitoring and multi-source EO integration.

Development Tools

The project combined Sentinel-1 SAR imagery, Sentinel-2 optical data, and deep learning for the automatic detection of Small Agricultural Reservoirs. Sentinel-1 data were used as the main input for the segmentation model, while Sentinel-2 data supported the refinement of the reference masks. The core of the work was the implementation of a U-Net convolutional neural network in PyTorch for pixel-wise image segmentation. The workflow also included geospatial data preparation, image patch extraction, dataset splitting into training, validation, and test subsets, and model evaluation. More broadly, the project relied on remote sensing analysis and GIS-based preprocessing for small water body detection, using QGIS and Google Earth Engine (GEE) as supporting tools.

Development Outputs

• A methodological workflow for the automatic detection of Small Agricultural Reservoirs from Sentinel-1 SAR imagery.
• A structured dataset of 5,193 image patches (256 × 256 pixels) covering Tuscany, organised into training, validation, and test subsets.
• Refined reference masks for the segmentation task, based on the 2016 LaMMA dataset and updated using Sentinel-2 observations from 2021.
• Preliminary U-Net segmentation results showing the feasibility of detecting small agricultural reservoirs from SAR data.

• A methodological basis for the continuation of the research during the second visiting period, focused on temporal dynamics and multi-source EO time series.
• Mannucci, N., Bertoli, G., Lompi, M., Pacetti, T., Goodarzi, M. S., Ebel, P., Chiarelli, D. D., Azzari, M., Caporali, E. (2025), “Small agricultural reservoirs detection with satellite data and OpenStreetMap integration for sustainable water management: a contribution to the CASTLE project”, Conference abstract – PICO presentation: EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-12476; Conference abstract – poster presentation: PhD Day 2025 – School of Engineering, Florence, Italy, 22 May 2025; Conference abstract – poster presentation: Living Planet Symposium 2025, Vienna, Austria, 23–27 June 2025.

Project Description

During my first visiting period at ESA Φ-lab, from 1 November to 31 December 2024, I worked on the automatic detection of Small Agricultural Reservoirs (SmARs) using Sentinel-1 Synthetic Aperture Radar (SAR) imagery and a U-Net deep learning model for semantic segmentation. This activity was part of my broader PhD research, which focuses on the use of Earth Observation data for water-related applications, with particular attention to the role of small and distributed water storage systems in agricultural landscapes.

Small Agricultural Reservoirs are of growing interest in the context of climate change adaptation, especially in Mediterranean regions affected by increasing water stress, drought frequency, and seasonal variability in water availability. Although individually small, these reservoirs can collectively play an important role in local irrigation practices and in supporting agricultural resilience. However, their mapping and monitoring remain technically challenging. Their dimensions are often limited, their shapes irregular, and their visibility can vary significantly depending on seasonal filling conditions, surrounding land cover, and the spatial resolution of available datasets. For this reason, improving methodologies for their identification is an important step toward a better understanding of their distribution and hydrological relevance.

This work built on a previous mapping effort I had carried out on SmARs in Italy using optical satellite data and ancillary geospatial information. That earlier phase of research provided an essential starting point for the work at ESA Φ-lab, since it made available a first reference inventory and associated spatial information. For the segmentation task developed during the visiting period, I used masks derived from the 2016 LaMMA dataset, which were then refined using Sentinel-2 observations from 2021. The purpose of this refinement was to distinguish, among the known reservoirs, those that actually contained surface water during the selected period. This step was important because it allowed me to work with labels that were more consistent with the real spectral and spatial conditions represented in the imagery, rather than relying on a purely static inventory.

One of the main motivations for focusing on Sentinel-1 was the need to overcome some of the limitations of optical satellite imagery in water monitoring applications. Optical data, such as Sentinel-2, are highly valuable for detecting open water and are widely used in surface water mapping. However, their effectiveness can be strongly reduced by cloud cover, particularly in periods when frequent observations are needed or in regions where atmospheric conditions limit the availability of usable scenes. Sentinel-1, by contrast, carries a C-band SAR sensor, which can acquire data independently of cloud cover and solar illumination. This makes SAR especially useful for environmental monitoring workflows that require a more stable temporal continuity. In the case of SmARs, this is particularly relevant because these reservoirs can change rapidly over time and may need to be observed under a range of seasonal and weather conditions.

The methodological core of the collaboration was the implementation of a U-Net convolutional neural network for semantic segmentation of Sentinel-1 imagery. U-Net was selected because it is one of the most established architectures for pixel-level segmentation tasks and is well suited to the extraction of fine spatial patterns from remotely sensed images. Its encoder-decoder structure allows the model to combine local image details with larger contextual information, which is particularly important when working with small and irregular features embedded in heterogeneous rural landscapes. In the case of SmARs, this ability is crucial, because the reservoirs are often surrounded by complex land cover types and may not present a simple or uniform radar response.

To build the dataset, I extracted 5193 image patches, each of 256 × 256 pixels, from Sentinel-1 scenes covering the Tuscany region. Tuscany represented a suitable study area because it contains a significant number of agricultural reservoirs distributed across different environmental and land-use settings. The dataset was split into 60% for training, 25% for validation, and 15% for testing. This subdivision allowed the modelling process to be carried out in a structured and transparent way. The training set was used to optimise the parameters of the model, the validation set to monitor its behaviour during development and support model tuning, and the independent test set to assess the generalisation capacity of the final model on unseen data. This separation was essential to avoid overestimating performance and to ensure that the results reflected a more robust methodological evaluation.

A relevant methodological challenge was the issue of class imbalance. Small reservoirs occupy only a limited fraction of the image area compared with the surrounding non-water classes, which means that the model is trained on data in which the target class is relatively rare. This is a common problem in segmentation of small objects and requires careful interpretation of model behaviour and performance metrics. In addition, water bodies in SAR imagery may sometimes be confused with other land surface conditions that produce similar backscatter responses under specific circumstances. For this reason, the work also involved examining the outputs qualitatively and considering potential sources of ambiguity in the classification.

The first results of the model were encouraging. In several test cases, the U-Net was able to correctly identify small reservoirs in Sentinel-1 imagery and to distinguish them from surrounding landscape features with good accuracy. These early results suggested that SAR data, combined with a deep learning segmentation framework, can provide a promising basis for the automatic detection of SmARs. While this first stage of the work was mainly methodological and exploratory, it helped demonstrate the feasibility of the approach and provided a stronger scientific basis for further development.

Beyond the immediate technical outputs, the visiting period was also valuable because it allowed me to discuss methodological choices and research directions in close interaction with the ESA Φ-lab environment. The activity was carried out in regular exchange with Dr. Patrick Ebel and Dr. Nicolas Longépé, who provided scientific feedback on the technical aspects of the work. At the same time, Dr. Sabrina Ricci supported the coordination of the visiting period and helped ensure alignment with the broader objectives of the lab. This collaborative setting was important not only for refining the methodological approach, but also for placing the work within a wider perspective on innovative EO applications.

The next phase will investigate seasonal fluctuations, changes in storage conditions, and longer-term behaviour of SmARs, with the goal of developing approaches that are increasingly robust, scalable, and relevant for future high-frequency, high-resolution data streams such as those expected from the upcoming IRIDE constellation.