rohitm9/ufo-segformer-b2

UFO SegFormer B2

Archived SegFormer B2 checkpoint supplied for the Urban Flood Observations (UFO) revision, together with verified binary prediction mosaics and their evaluation CSV. The checkpoint is provided in its original PyTorch format and as a standard Transformers model using safetensors.

Model and input contract

• Architecture: SegFormer B2; encoder depths [3, 4, 6, 3].
• Decoder hidden dimension: 768.
• Input: three channels, confirmed by the first convolution's [64, 3, 7, 7] weight shape.
• Output: two classes, 0 non-inundated and 1 surface water.

The original three-band selection, band order, input scaling and normalization were not included with the checkpoint. No image processor is supplied, and no RGB, ImageNet or four-band normalization is assumed. The four-band PlanetScope training pipeline in the linked GitHub repository is a separate input contract.

The checkpoint contains a state dictionary without the original training configuration. Its configuration was reconstructed using the tensor shapes and the NVIDIA B2 reference configuration. Non-tensor architecture settings use that reference. Only the leading model. wrapper prefix was removed when converting the weights.

Loading

import torch
import torch.nn.functional as F
from transformers import AutoModelForSemanticSegmentation

model = AutoModelForSemanticSegmentation.from_pretrained(
    "rohitm9/ufo-segformer-b2"
).eval()

@torch.inference_mode()
def segment_prepared_tensor(pixel_values):
    # Float tensor [batch, 3, height, width], prepared with the original
    # training band selection/order and normalization.
    if pixel_values.ndim != 4 or pixel_values.shape[1] != 3:
        raise ValueError("Expected a prepared tensor with three input channels")
    logits = model(pixel_values=pixel_values).logits
    logits = F.interpolate(
        logits, size=pixel_values.shape[-2:],
        mode="bilinear", align_corners=False,
    )
    return logits.argmax(dim=1)

Confirm the original preprocessing before using the checkpoint on imagery. Loading and conversion were tested with Transformers 4.47.1 and PyTorch 2.6.0. All checkpoint tensors survived the safetensors round trip bit for bit. A small synthetic CPU forward pass produced finite, identical logits before and after conversion. These checks validate serialization and execution, not prediction accuracy or recovery of the original preprocessing.

Archived prediction evaluation

The supplied B2-named prediction ZIP contains 215 full-chip masks across 14 UFO events. Each mask is a single-band 1024 x 1024 GeoTIFF with values 0 and 1. The supplied CSV has 215 image rows, one overall mean row and one totals row.

Every filename and event matches the corresponding local UFO image and label. Dimensions, coordinate reference systems and affine transforms match for all 215 pairs. Recomputed TP, FP, TN and FN match the CSV exactly for every image; Precision, Recall, IoU, F1 and Accuracy agree within 1e-6.

Water IoU aggregation	Score
Mean of the 215 per-image IoUs (CSV OVERALL_MEAN)	77.340578%
Pooled-pixel IoU	85.081899%
Mean of the 14 event mean IoUs	73.937440%

These scores describe the archived prediction artifacts. Their linkage to this individual checkpoint has not been independently verified. Fold-specific checkpoint identities and training/test separation were not supplied with these artifacts, so this release does not certify that the single checkpoint reproduces the scores or the full leave-one-event-out experiment.

Total confusion counts: TP 74,159,184; FP 7,193,473; TN 138,281,723; FN 5,809,460.

Event counts: BEI 2, BNA 5, CMO 14, CTO 7, DKA 12, GIL 20, HTX 39, KTM 23, MID 4, NSW 27, PNE 13, QUE 12, SLC 6, SPS 31.

Files

File	Contents
config.json, model.safetensors	Transformers checkpoint, three input channels
segformer_b2_best.pt	Original supplied state dictionary
checkpoint_verification.json	Architecture, loading and conversion checks
release_manifest.json	Source links, sizes and SHA-256 checksums
evaluation/PS_256_localNorm_segformer_b2_Preds_ProcessedMosaics.zip	215 binary prediction mosaics
evaluation/PS_256_localNorm_segformer_b2_Evaluation_mosaics.csv	Original evaluation CSV
evaluation/matching_report.json	Dataset matching and recomputed aggregate scores
evaluation/per_image_verification.json	Checks for each image

Sources and license

• UFO code repository.
• UFO dataset, released under CC BY 4.0.
• Original checkpoint.
• Prediction ZIP.
• Evaluation CSV.

The supplied checkpoint did not specify a model-weights license. The dataset license should not be interpreted as a license for the model weights.