Italian
Flood Atlas.

Before training, I checked acquisition times, coverage, preprocessing and labels. The Italian emergency maps were based on observations taken at different times from the radar images. They also lacked an explicit dry class, so they could not serve as labels for measuring the model’s accuracy on those images.

I used Kuro Siwo for training and evaluation because it provides labels interpreted from the radar imagery itself. The Italian event is a separate, qualitative demonstration. Processing code, parameters, evaluation splits and source records are versioned; raw imagery and model checkpoints stay outside Git.

The baseline identifies dark radar pixels whose reflected signal, or backscatter, drops from the pre-event mean. I fitted its thresholds on training data and froze them before scoring. The U-Net uses both radar channels, VV and VH, across three dates. I excluded elevation after finding artifacts in the benchmark’s terrain data.

Some training and validation image tiles covered the same locations. To prevent that overlap from influencing model selection, I selected the model using only validation events with no training overlap. Once the model and evaluation policy were fixed, I scored both methods once on the same 15,896 test tiles across 10 event activations.

Kuro Siwo test set · flood class only

Method	Precision	Recall	IoU
U-Net v1	0.821	0.739	0.636
VV change baseline	0.651	0.764	0.542

Precision measures how much of the predicted flood matches the labels; recall measures how much of the labeled flood is detected. Intersection over union (IoU) divides the overlap between predicted and labeled flood by their combined extent. These scores apply to the benchmark alone.

The U-Net has higher precision and IoU than the baseline, with lower recall: it makes fewer false flood predictions but misses more labeled flood. It still confuses flood with permanent water in one test activation, and moderate backscatter decreases can cause false flood predictions. Because the labels were interpreted from the same radar imagery, these scores measure agreement with expert radar interpretation.

The Italian radar image was acquired on 22 May, about five days after the main flood. The figure pairs it with Copernicus Emergency Management Service (CEMS) outlines derived from an optical image taken roughly five hours later. It shows the radar and reference maps; model predictions can be compared in the live Atlas.

• Flooded area
• Flood trace
• Reference coverage

Both methods predict flood in only a small part of the area mapped as flooded by CEMS. The radar image shows little dark open water. Water may have receded, while vegetation and buildings can also affect its radar appearance; this comparison does not establish which explanation accounts for the difference.

One of the pre-event images, from 10 May, also follows an earlier flood. Any water still present on both dates could be classified as permanent water. The interface therefore keeps acquisition dates and reference sources visible. Overlap with the CEMS outlines describes the maps; it does not measure the model’s accuracy in Italy.

To check the ONNX export, I compared its predictions with the original model. Across 135,859,080 validation pixels, ONNX Runtime and PyTorch float32 disagreed on 1 pixel. CPU measurements showed that memory use limits the area processed in one pass. Inference runs offline, and its outputs are published as static files.

The Atlas serves static layers and self-hosted basemap tiles through Cloudflare Pages. SvelteKit generates the pages, and MapLibre provides the interactive comparison map. Keyboard controls and shareable views support map navigation; a static figure remains available without JavaScript. The personal site builds independently of the Atlas, using the same shared design package.

Prepared layers let the map work without an inference server or database. Automated checks cover published values, layer provenance, asset sizes and accessibility. The Atlas remains a portfolio study of historical floods and has not been validated for operational use in Italy.