irfanhossainsust/mewtp-convnext-tiny

MEWTP ConvNeXt-Tiny — water treatment plants in Sentinel-2 imagery

Binary classifiers that decide whether a 1 km × 1 km Sentinel-2 chip contains a water treatment plant (wastewater, drinking water or desalination) in the Middle East. Trained on Middle East Water Treatment Plants in Satellite Imagery (Kaggle mirror).

File	Input	Threshold
convnext_tiny_rgb_v2.pt ⭐ recommended	RGB PNG chip (B04/B03/B02), 256 px → 224	0.05
convnext_tiny_6band_v2.pt	6-band array B02 B03 B04 B08 B11 B12, 100×100 at 10 m → 224	0.20
ensemble (mean of both)	both	0.30

Probabilities are not calibrated (heavy class weighting); use the validation-tuned thresholds in config.json.

Results (test split: 1,772 images from areas never seen in training)

Model	ROC-AUC	Accuracy	Precision	Recall	F1	FPR ordinary neg.	FPR hard neg.	FPR urban
ConvNeXt-Tiny RGB v2	0.957	0.904	0.847	0.817	0.832	4.6%	9.4%	8.0%
ConvNeXt-Tiny 6-band v2	0.954	0.888	0.796	0.827	0.811	7.2%	12.1%	14.7%
Ensemble	0.958	0.899	0.832	0.817	0.824	5.0%	11.1%	9.7%
ConvNeXt-Tiny RGB v1 (no extra cleaning)	0.957	0.897	0.816	0.831	0.824	6.9%	9.4%	13.9%

Per-country test ROC-AUC (RGB v2): AE 0.974 · EG 0.981 · IL 0.986 · IQ 0.944 · JO 0.989 · KW 0.987 · OM 0.909 · SA 0.918.

Usage

from huggingface_hub import snapshot_download
import sys
d = snapshot_download("irfanhossainsust/mewtp-convnext-tiny")
sys.path.insert(0, d)
import inference
probs, is_plant = inference.predict_rgb(["chip.png"])          # RGB PNG chips
# probs, is_plant = inference.predict_six_band(X)              # X: uint16 [N, 6, 100, 100]

Or from the command line: python inference.py chip1.png chip2.png.

Inputs should match the training chips: Sentinel-2 L2A, ~1000 m extent centred on the site, RGB stretched linearly over 0–0.40 BOA reflectance (see the dataset card).

Training

• ImageNet-initialised torchvision ConvNeXt-Tiny, single-logit head, AdamW (lr 3e-4, wd 0.05), cosine schedule with 1-epoch warm-up, 12 epochs, batch 64, mixed precision, Kaggle Tesla T4 (~11 min each).
• Augmentation: random resized crop (0.75–1), flips, 90° rotations, colour jitter. TTA: identity + 2 flips.
• Label cleaning (train only): 5-fold out-of-fold ConvNeXt (OOF ROC-AUC 0.927) dropped 447 positive images of 149 facilities with mean P(plant) < 0.25 and up-weighted ×3 192 confusable negatives (mostly commercial/urban, water storage, ponds). Validation and test labels are untouched.
• 6-band model: first conv expanded to 6 channels (RGB weights mapped to B04/B03/B02, NIR/SWIR initialised from their mean), per-band standardisation.
• Code: Kaggle notebook.

Limitations — read before use

• Labels are weak. Dataset positives are model-filtered, not human-verified; test metrics include label noise.
• Not for unattended mapping. About 1 in 7 positive predictions is wrong; at real-world prevalence (plants are rare) false alarms dominate. Use as a screening tool with human confirmation.
• Confusers: tank farms, industrial ponds, warehouses and dense urban blocks.
• Trained on 11 Middle-East countries at 10 m resolution; small package plants may be invisible; other regions and sensors are untested.
• The 6-band model did not outperform RGB on test.

License and attribution

Weights: CC BY 4.0. Training imagery: Contains modified Copernicus Sentinel data 2019–2025. Facility labels derive from OpenStreetMap (© OpenStreetMap contributors, ODbL), HydroWASTE (CC BY 4.0, Ehalt Macedo et al. 2022) and Wikidata (CC0). ImageNet initialisation from torchvision (BSD-3-Clause).

Citation

@misc{hossain2026mewtp,
  title  = {Middle East Water Treatment Plants in Satellite Imagery},
  author = {Hossain, Irfan},
  year   = {2026},
  url    = {https://huggingface.co/datasets/irfanhossainsust/middle-east-water-treatment-satellite}
}