L5 specialists carried out rapid monitoring of the flood situation in the Republic of Dagestan from March 30 to April 12, 2026. The analysis was based on radar imagery from the Sentinel-1 satellites in Interferometric Wide Swath (IW) mode, with 10-meter spatial resolution and dual VV+VH polarization. Radar imaging can detect water surfaces regardless of cloud cover or time of day, making it the primary tool for rapid response during spring flooding.
Within the monitored area, covering roughly 55×55 km in central Dagestan, two separate flood zones were identified. The main zone was a large-scale flood in the lowlands, first detected on March 30, 2026. The flooded area peaked at 87.9 km². Water spread across large stretches of agricultural land near the villages of Adilotar, Kadyrotar, and Tukita, as well as near the settlements of Novoye Kheletyuri, Sulevkent, Narysh, Shagada, Arkhida, and Kutan Butush. As of April 12, the flooded area had shrunk to 26.0 km² — more than a threefold decrease from the peak, indicating the flood was actively receding.
The second flood zone was located in the Sulak River valley, where the river overflowed its banks, causing localized flooding of the floodplain. The flooded area peaked at 1.6 km² on April 5 and had shrunk to 0.7 km² by April 12.
These results were produced using a flood-segmentation algorithm developed in-house at L5, based on a consensus decision from an ensemble of five deep learning models, combined with multi-stage post-processing. This approach significantly reduces false positives and improves the accuracy of mapping actual flood boundaries.
The algorithm's accuracy is on par with the best comparable systems worldwide, and in some cases surpasses them. That said, this work represents just the first stage in developing the system further.
