Dual-Cue Feature Fusion with Conformal Reliability for Efficient Remote Sensing Change Detection
Official implementation. Manuscript submitted to (Springer Journal).
Abdulkadhem A. Abdulkadhem — Department of Artificial Intelligence, College of Sciences, Al-Mustaqbal University, Hillah, Babil, Iraq — a.abdulkadhem@uomus.edu.iq
The manuscript is not in this repository. It is under review; this repository carries the code, the results and the figures. Everything needed to reproduce every number is here.
Two things, and they are independent. You can take either without the other.
1. An accurate convolutional change detector. 13.6 M parameters, no attention, no transformer, no state-space scan. It beats a reproduced FC-Siam-Diff baseline on five public benchmarks and reaches parity with far heavier transformer and state-space models on the protocol-matched ones. It is not the smallest detector in the literature — TinyCD reaches comparable accuracy at a fiftieth of the size — so the efficiency claim here is against the transformer family only.
2. A reliability layer that works on any change detector, including yours. Given a model that outputs a probability map, it calibrates a single threshold so that the expected per-image false-discovery proportion on unseen data is provably at most α, with no assumption about the data distribution. If no threshold can honour the budget, it says so and abstains rather than issuing a certificate it cannot support.
The second point is the one most likely to be useful to other people's work. It is post-hoc, training-free and model-agnostic: about forty lines of numpy applied to whatever probability maps your model already produces.
The task, and the two things an accurate detector still fails to tell its user.
One LEVIR-CD tile pushed through the trained model. Every panel is the model's own response: the difference cue reacts to change magnitude and lights up the replaced buildings, the concatenation cue holds the surrounding context, and the fused feature sharpens the changed footprints the decoder then isolates.
In-domain test F1. Both models trained by us under an identical protocol, one shared recipe with no per-dataset tuning.
| Dataset | FC-Siam-Diff (reproduced) | DCFF-Net | Δ |
|---|---|---|---|
| LEVIR-CD (+MS-TTA) | 85.31 | 90.69 | +5.4 |
| SYSU-CD | 79.24 | 83.57 | +4.3 |
| EGY-BCD | 68.10 | 81.39 | +13.3 |
| CLCD | 50.46 | 77.25 | +26.8 |
| DSIFN-CD | 57.31 | 67.61 | +10.3 |
Conformal-FDR on LEVIR-CD, calibrated on validation and evaluated once on test:
| Target α | Threshold λ̂ | Realised test FDP | Recall | Guarantee |
|---|---|---|---|---|
| 0.05 | 0.780 | 0.045 | 0.756 | held |
| 0.10 | 0.080 | 0.097 | 0.835 | held |
| 0.20 | 0.020 | 0.125 | 0.852 | held |
| fixed 0.50 | 0.500 | 0.059 | 0.791 | none |
At α ≥ 0.10 the certified threshold reaches higher recall than the default 0.5 threshold while also bounding false alarms. The conventional threshold is not merely unguaranteed here; it is dominated.
We are not the accuracy state of the art and we do not present ourselves as such. On LEVIR-CD we trail ChangeTitans by 0.83 F1. On SYSU-CD our two-seed mean is 82.89 ± 0.96, which sits among ChangeTitans (83.24), ChangeMamba (83.11), CFNet (82.87) and TransY-Net (82.84) rather than above them — we report the mean rather than our better run (83.57) because the seed spread is as large as the margin separating those four methods. Our CLCD copy is a 256-crop re-split and is not comparable with published CLCD numbers. On EGY-BCD and DSIFN-CD we remain below published results and make no competitiveness claim.
The claim is parity with the transformer and state-space family at a parameter count we report and most of them do not, plus a guarantee none of them offers.
git clone https://github.com/Abdulkadhem/DCFF-Net.git
cd DCFF-Net
pip install -r requirements.txtCPU is enough for evaluation and for the conformal layer. Training wants a GPU. Tested on Python 3.11–3.13, PyTorch 2.0–2.13, Linux and Windows.
Download a checkpoint from the Releases page into weights/, then:
python code/eval_local.py --ckpt weights/dcff_final_best.ptExpected, exactly:
[RESULT] P=0.8948 R=0.9057 F1=0.9002 IoU=0.8185
That is the 90.02 / 81.85 of Table 1. The LEVIR-CD test split downloads on first use and is cached. On a laptop CPU this takes roughly ten minutes.
This is the part that does not need our architecture. You need one thing: a
function that turns an image pair into a probability map in [0, 1].
import numpy as np
from conformal_fdr import crc_threshold
# 1. On held-out images, count what each candidate threshold would cost.
# fdp[i, j] = false-discovery proportion of image i at threshold lambdas[j].
lambdas = np.linspace(0.0, 1.0, 101)
fdp = np.zeros((len(calib_images), len(lambdas)))
for i, (t1, t2, y) in enumerate(calib_images):
p = your_model(t1, t2) # any model, any framework
for j, lam in enumerate(lambdas):
pred = p > lam
tp = (pred & y).sum()
fp = (pred & ~y).sum()
fdp[i, j] = fp / (tp + fp) if (tp + fp) else 0.0
# 2. Certify a threshold for the budget you choose.
# Returns (grid index, threshold, empirical risk curve); the first two are
# None when no threshold can honour the budget.
idx, lam_hat, risk = crc_threshold(fdp, lambdas, alpha=0.10)
if lam_hat is None:
print("no threshold can honour alpha=0.10 on this data — abstaining")
else:
print(f"use threshold {lam_hat:.3f}; E[FDP] on exchangeable test data <= 0.10")Two things are worth understanding before relying on it.
The guarantee holds under exchangeability, meaning the calibration images and the images you deploy on must be drawn alike. That is a real condition, not a formality. Calibrate on one city and deploy on another with a different sensor and the bound can fail.
The failure is informative. We measured how it fails: the gap between realised and target error grows with the size of the distribution shift. Calibrating on a matched split gives a valid certificate or an honest abstention; calibrating on a shifted split gives a certificate whose gap tells you how far the data has moved. Section 5.5 of the paper reports this across three datasets. In practice, calibrating on a held-out slice of the target data is both the correct procedure and the one that works.
Every number and every figure regenerates from this repository. The commands below are the ones that produced them.
python code/train.py --dataset levir --model dcff --fusion dual \
--no_cbam --no_boundary --batch 16 --lr 6e-4 --epochs 100 \
--tag levir_dcff --out results--dataset accepts levir, sysu, egy, clcd, dsifn. Each run writes a
*_result.json with the full argument set, parameter count and metrics, plus a
training log. Roughly four hours on one RTX 3090 for LEVIR-CD at 100 epochs.
python code/eval_local.py --ckpt weights/dcff_final_best.pt # 90.02, Table 1
python code/eval_mstta.py --ckpt weights/dcff_final_best.pt --dataset levir # 90.69, adopted
python code/eval_cross.py --ckpt weights/dcff_final_best.pt # zero-shot, Sec. 6.1python code/conformal_prep.py --ckpt weights/dcff_final_best.pt # per-image FDP counts
python code/conformal_fdr.py # Table 4
python code/conformal_multi.py --dataset egy --ckpt weights/egy_dcff_best.pt # Table 5python paper/extract_flow_assets.py # runs the model, saves the intermediate maps
python paper/make_problem.py # Fig 1
python paper/make_architecture_v2.py # Fig 2
python paper/make_dualcue_v2.py # Fig 3
python paper/make_dualcue_scales.py # Fig 4
python paper/make_training_curve.py # Fig 5
python paper/make_figures.py # Figs 6, 7, 9, 14
python paper/make_curves_v2.py # Figs 8, 15
python paper/make_ablation_curves.py # Figs 10, 11
python paper/make_conformal_multi.py # Fig 13
python code/make_qualitative.py --ckpt weights/dcff_final_best.pt # Figs 16–18
python code/qualitative_multi.py --dataset egy --ckpt weights/egy_dcff_best.pt # Figs 19–20Figure numbers above are the ones used in the paper. paper/figures/ holds the
same images under descriptive names.
Five public benchmarks, downloaded on demand through 🤗 datasets and cached:
| Short name | Dataset | HuggingFace id | Train / Val / Test |
|---|---|---|---|
levir |
LEVIR-CD | ericyu/LEVIRCD_Cropped_256 |
7120 / 1024 / 2048 |
sysu |
SYSU-CD | ericyu/SYSU_CD |
12000 / 4000 / 4000 |
egy |
EGY-BCD | ericyu/EGY_BCD |
3654 / 1219 / 1218 |
clcd |
CLCD | ericyu/CLCD_Cropped_256 |
1440 / 480 / 480 |
dsifn |
DSIFN-CD | EVER-Z/torchange_dsifn-cd |
3600 / 340 / 48 |
Our LEVIR-CD, SYSU-CD and DSIFN-CD splits reproduce the standard partitions, so those numbers are directly comparable with published work. The CLCD copy is a 256-crop re-split and is not.
No path is hardcoded anywhere in code/. Everything resolves through
code/paths.py, which you can override:
export DCFF_DATA=/mnt/big-disk/hf_cache # where datasets live
export DCFF_WEIGHTS=/mnt/big-disk/weights # where checkpoints liveTrained weights are published on the Releases page, not in git
(52 MB each). Download the one you need into weights/.
| File | Dataset | Test F1 | Use it for |
|---|---|---|---|
dcff_final_best.pt |
LEVIR-CD | 90.02 (90.69 +MS-TTA) | the headline result, the conformal tables, all LEVIR figures |
sysu_dcff_best.pt |
SYSU-CD | 83.57 | the protocol-matched comparison |
egy_dcff_best.pt |
EGY-BCD | 81.39 | cross-dataset soundness |
clcd_dcff_best.pt |
CLCD | 77.25 | the imbalanced-data case |
dsifn_dcff_best.pt |
DSIFN-CD | 67.61 | the hard case, where the layer abstains |
Every checkpoint is a plain state_dict for DCFFNet(backbone="resnet18", use_cbam=False, use_aspp=True). Nothing else is needed to load one:
import torch
from models import DCFFNet
net = DCFFNet(backbone="resnet18", pretrained=False, use_cbam=False, use_aspp=True).eval()
net.load_state_dict(torch.load("weights/dcff_final_best.pt", map_location="cpu"))code/
paths.py every data and checkpoint location, and nothing else
models.py DCFFNet, DualCueFusion, ASPP, the FC-Siam baseline
data_cd.py one loader for all five datasets
train.py training entry point
eval_local.py plain evaluation
eval_mstta.py multi-scale test-time augmentation (the adopted result)
eval_cross.py zero-shot cross-dataset transfer
conformal_prep.py per-image false-discovery counts
conformal_fdr.py the conformal threshold, and Table 4
conformal_multi.py cross-dataset soundness, Table 5
soup.py model soup — a reported negative result, kept for the record
make_qualitative.py qualitative panels
qualitative_multi.py the same for EGY-BCD and DSIFN-CD
paper/
figures/ every figure in the paper, as generated
figstyle.py shared drawing primitives for the diagrams
make_*.py one script per figure
extract_*.py run the model and dump the intermediate maps the figures use
tables.md every table, with the caveats attached
equations.md the 13 equations, each pointing at the code that implements it
results_log.md chronological record of every experiment, failures included
results/ *_result.json and training logs for every run in the paper
Every file in code/ produces something that appears in the paper. Scripts that
did not survive into the manuscript were removed rather than left lying around.
Seeds are fixed at 42; the SYSU-CD variance study additionally uses 7. Each run
writes its complete configuration next to its metrics, so a *_result.json in
results/ is enough to reconstruct the command that produced it.
We also report what did not work. Seven independent avenues for further accuracy were probed and six failed — longer schedules, remote-sensing self-supervised pretraining, Lovász-hinge training, Lovász polishing, model soup and FixRes fine-tuning. Only multi-scale test-time augmentation helped, and by 0.67 F1. Table 6 of the paper lists all seven. Two of the failures are worth knowing about: Sentinel-2 self-supervised pretraining transfers worse than ImageNet for 0.5 m aerial change detection, which suggests that matching resolution matters more than matching nominal domain; and model soup collapses across independently initialised runs, which do not share a loss basin.
If you use the conformal layer, please also cite the risk-control result it rests on:
MIT. The datasets keep their own licenses; see the sources listed above.
Questions, bug reports and reproduction failures are all welcome as issues. If a number in the paper does not reproduce for you, please open one — we would rather know.