Introduction

Detection and monitoring of changes in remote sensing data remain a central topic within the Earth observation community over recent years . A wide range of applications emerge from remote sensing change detection, including natural disaster damage assessment , urban development monitoring , or deforestation detection . In particular, the task of bi-temporal change detection consists in finding the semantic differences between a pair of co-registered satellite or aerial images acquired at different dates, given a set of land cover classes of interest.

State-of-the-art methods for this task use deep neural networks  trained on datasets of image pairs annotated with pixel-level change maps  or even semantic maps at both dates . Nevertheless, gathering such data is both expensive and time consuming. As a result, change detection data is often spatially or temporally clustered, and methods trained on them hardly generalize to new, unseen locations. To address the scarcity of labeled data, several works have attempted to train change detection models in an unsupervised  or weakly supervised fashion , learning from much cheaper image-level annotations. However, these methods still lag behind fully supervised approaches, particularly in the spatial accuracy of predicted changes.

To benefit from quality semantic annotations at a larger scale, several works attempt to leverage single-temporal annotated datasets in order to train change detection model. This is the case of synthetic change datasets generated from such existing data , but also of learning frameworks considering non-overlapping images as training pairs , pairing images from different locations and thus generating fake change pairs. However, such attempts do not take advantage of the growing availability of remote sensing data. Satellite and aerial image acquisition campaigns are indeed often programmed to occur regularly by national GIS institutes  or to last a long period of time by spatial agencies .

In this paper, we propose to expand single-date semantic segmentation datasets into bi-temporal collections, leveraging easily accessible imagery. We introduce a new weakly supervised paradigm, where only single-temporal semantic maps are available to train a change detection network on bi-temporal pairs. Introducing this temporal variability exposes the model to naturally occurring variations that do not constitute semantically meaning changes, making it more robust. Because this weak temporal supervision is by nature noisy, we propose three key training ideas. First, we balance real bi-temporal pairs and fake non-overlapping pairs during training. Secondly, inspired by , we supervised fake pairs with an sIoU-based  change map. Third, we iteratively clean the extended dataset, filtering out real pairs exhibiting land-cover changes, so that such pairs can be considered as unchanged during training.

We validate our methodology by expanding two existing datasets, FLAIR  and IAILD . Through several experiments, we demonstrate that one can use additional temporal acquisitions in addition to semantic segmentation datasets to improve change detection with zero labeling cost. Our contributions are summarized as follows:

• We extend FLAIR and IAILD datasets from single-date to bi-temporal and release them for anyone to use. We additionally curate and release a test set for in-domain evaluation of building change detection methods trained on our extension of FLAIR.

• We develop an approach to leverage the additional, non-annotated images for training a change detection network.

• Through extensive experiments on several datasets, we show that our method improves the performance of change detection models. Thus, we demonstrate strong zero-shot performance, impressive low-data regime results, and compelling large-scale qualitative results, shown in Fig. [fig:teaser].

Related Works

In this paper, we tackle the problem of detecting land cover changes in bi-temporal remote sensing image pairs. We refer to this task as semantic change detection (SCD).

Over the past few years, SCD in remote sensing images has gained significant interest, leading to a large number of publications on the topic and several field surveys . Most state-of-the-art methods rely on deep learning in order to train 3-branch neural networks. First introduced by Daudt et al. , such architectures output, for a given bi-temporal image pair, a triplet consisting of two semantic maps and a binary change map. Multiple variants  improve different aspects of the procedure, including the multitask objective , the fusion mechanisms , the consistency between the three outputs , the quality of the predicted change boundaries , or the computational cost . All these methods require pixel-level annotated image pairs in order to train a model in a fully supervised manner. Because gathering such labeled data is costly and time-consuming, we instead propose a weakly supervised framework based on single-temporal annotations, leveraging new temporal images without additional annotation cost.

Weakly supervised learning encompasses any training algorithm that enables performing different or more complex tasks than typically possible with the available data. However, in the context of SCD, it specifically describes methods that infer pixel-level change predictions from weak image-level labels . For instance, Andermatt et al.  train a SCD model with image-level labels such as “forests to agricultural surfaces” for a given image pair. Between image- and pixel-level supervision, other weak signals such as bounding boxes or coarse masks , noisy labels , and low-resolution annotations  have been used to train SCD networks. Toker et al. presented DynamicEarthNet , a dataset of daily satellite image time series for which only the first image of each month is annotated, with a focus on semi- rather than weak supervision. Beyond this work, and to the best of our knowledge, weak settings in which semantic information is completely missing for one image of the pair are often overlooked in the literature, despite corresponding to cases that commonly occur in practice due to the recurrent acquisition of imagery by space and GIS agencies.

Due to the high cost of pixel-level change annotations, several works have explored leveraging single-temporal semantic segmentation datasets to enable change detection without requiring ground truth change labels. A common approach is post-classification change detection, often referred to as “the most obvious method of change detection” . It compares the outputs of a single-temporal segmentation model on each image of a bi-temporal pair. In this case, multi-temporal acquisitions may serve as data augmentation at training time . However, this method suffers from prediction error accumulation and lacks temporal modeling . Other methods construct artificial change pairs by randomly pairing images from different locations and computing change maps based on label differences, STAR  or . Another line of work consists in creating synthetic SCD datasets based on a single-temporal land cover real-world dataset, using techniques such as GANs , or diffusion . Building on the observation that annotation is significantly more expensive than data acquisition itself, we instead propose to extend existing single-temporal datasets with new, aligned temporal acquisitions, using their existing annotations as weak supervision for bi-temporal change detection.

Methodology

We focus on learning to detect changes from a single-temporal satellite or aerial image dataset annotated for the semantic segmentation task. Based on the observation that accessing new acquisitions is less costly than obtaining change labels, our pipeline starts by extending such datasets temporally (Sec. 3.1). Then we leverage these new bi-temporal pairs, along with the single-temporal annotations in order to train a semantic change detection model (Sec. 3.2). We detail our implementation in Sec. 3.3.

Let $`\mathcal{D}`$ be a single-temporal dataset of $`N`$ aerial or satellite annotated images $`I_{t_i}^i`$ in $`\mathbb{R}^{C\times H\times W}`$ acquired at time $`t_i`$, for $`i \in \{1, \ldots, N\}`$. $`C`$, $`H`$, and $`W`$ respectively refer to the number of channels, height, and width of the images. Each $`I_{t_i}^i`$ is annotated with a semantic mask $`S_{t_i}^i`$ in $`\{1, ...,K\}^{H\times W}`$, with $`K`$ the number of semantic classes. Now, let $`f_\theta`$ be a deep neural network with learnable parameters $`\theta`$, returning, for a given input pair $`(I_{t}, I_{t'})`$ of bi-temporal images, a change map of pixels whose semantic class changes between the acquisition timestamps $`t`$ and $`t'`$. Our goal is to train such a model $`f_\theta`$, using the dataset $`\mathcal{D}`$ extended with new, easily-available, non-annotated temporal acquisitions $`I_{t'_i}`$.

Data

To verify our methodology, we apply it to two existing aerial datasets, FLAIR  and the Inria aerial image labeling dataset (IAILD) , spanning areas in three countries (France, USA and Austria). Since we extend these datasets so they contain bi-temporal pairs, we prepend their name with “b-” to distinguish them from the original datasets.

FLAIR is based on aerial images from the French National Institute of Geographical and Forest Information (IGN)’s BD ORTHO : a database of aerial orthophotographies covering the French territory at 0.2 meters per pixel (m/px). All $`512\times512`$ FLAIR image patches are annotated with semantic maps that assign each pixel to one of 19 land cover classes. BD ORTHO is renewed on average every three years, making it possible to collect aerial images of the same locations at different dates. We thus extract, for every FLAIR image, an additional observation from the BD ORTHO database at around three years before or after the original FLAIR acquisition. b-FLAIR is challenging for weakly supervised change detection with single-date annotations for two reasons. First, FLAIR areas were selected for land cover diversity with no particular temporal consideration, probably resulting in a low proportion of pairs with relevant semantic change. Second, approximately 30% of pairs have acquisition dates differing by two or more months, introducing seasonal variations that, combined with varying acquisition times within the day, may introduce significant radiometric and shadow changes.

IAILD covers urban areas in the USA and Austria. We only extended IAILD’s training set, as its test set is not publicly available. It consists of 180 patches of size $`5000\times5000`$ at a resolution of 0.3 m/px, annotated with binary building footprint masks. For each location, we gathered a patch from the most recent available orthoimage acquisition campaign released by the corresponding agency (USGS for the USA, and the respective GIS agencies for Austrian provinces). The added images have acquisition dates spanning 2022 to 2024, while IAILD’s original images were acquired before 2017. Because the new acquisitions have spatial resolutions varying from 0.15 to 0.6 m/px, we standardized b-IAILD to 0.6 m/px. Due to IAILD’s urban focus, we expect b-IAILD to contain more examples of changes related to artificialization and building construction or demolition.

To verify our methodology beyond very high resolution aerial imagery, we also created a satellite variant of b-FLAIR. For each bi-temporal pair, we downloaded the corresponding SPOT-6/7 images from IGN’s ORTHO-SAT database  for the same acquisition year and location. The images have a spatial resolution of 1.5 m/px, which we resampled to 1.6 m/px to obtain $`64\times64`$ patches. The corresponding single-date semantic maps were also resampled to the same resolution via nearest neighbor interpolation. We refer to this satellite dataset as b-FLAIR-spot.

b-FLAIR, b-IAILD and b-FLAIR-spot thus contain triplets ($`S_{t}`$, $`I_{t}`$, $`I_{t'}`$). We show example of such triplets in Fig. 1. Further details on these datasets, and additional example images can be found in the Supplementary Material.

We evaluate competing methods on 5 datasets. For the in-domain evaluation of methods trained on b-FLAIR and b-FLAIR-spot, we release b-FLAIR-test and b-FLAIR-test-spot, two evaluation sets of 1730 image pairs annotated with a binary building change mask. The images are in the same format as FLAIR images and were acquired, processed, and formatted following the procedure described in . The images were extracted from 9 different French administrative departments and do not intersect the FLAIR dataset, which allows for a sound evaluation of methods trained on FLAIR or on our bi-temporal extensions of FLAIR. The pairs were annotated by photointerpretation experts and verified by a non-expert assessor. The pairs either show new building constructions, or no building change at all ($`\thicksim`$30% of the pairs), and no pair exhibits building destruction. For the out-of-domain evaluation of all methods, we selected LEVIR-CD , WHU-CD , and S2Looking , three datasets commonly used in building change detection benchmarks. Hyperparameter studies are performed using a subset of b-FLAIR for training, and b-FLAIR-test for evaluation. More details on the evaluation can be found in the Supplementary Material.

Training with single date labels

Given a set of triplets $`(S_{t_i}^i, I_{t_i}^i, I_{t_i'}^i)_{i=1,...,N}`$, our goal is to train a bi-temporal change detection neural network $`f_\theta`$. Our methodology can be applied to any model predicting, for an input image pair ($`I_{t}`$, $`I_{t'}`$), a triplet ($`\hat{S}_{t}`$, $`\hat{S}_{t'}`$, $`\hat{M}`$) respectively consisting of a semantic mask for each frame and a binary change mask. A widely accepted architecture for semantic change detection indeed consists of a triple-branch network  with two semantic segmentation branches and a binary change branch. We now detail the three main components of our training strategy: balanced batch sampling, change map generation, and iterative refinement.

Our extended datasets typically contain a small fraction of pairs exhibiting actual change events. Zheng et al.  compensate the lack of annotated change datasets in the literature by artificially pairing images from different locations, using as change labels a difference of their respective semantic masks. Building on this idea, we choose to mix in a single batch such fake pairs with real bi-temporal pairs. We thus introduce a parameter $`p_{\text{real}}`$ in $`[0,1]`$, corresponding to the proportion of real pairs in each batch. Specifically, for a batch of size $`B`$ of pairs ($`I_{t_i}^i, I_{t_i'}^i)_{i=1,...,B}`$ sampled from $`\mathcal{D}`$, the method splits it into $`N_{\text{real}}`$ bi-temporal examples and $`N_{\text{fake}}`$ unaligned examples, following:

\begin{equation}
N_{\text{real}} = \lfloor B \times p_{\text{real}} \rfloor~, \quad N_{\text{fake}} = B - N_{\text{real}}~.
\end{equation}

We apply a random permutation to the $`N_{\text{fake}}`$ samples, such that the image $`I_{t_i}^i`$ is paired with $`I_{t_j'}^j`$ with $`j \neq i`$.

The models trained with our method usually supervise their predicted maps ($`\hat{S}_t`$, $`\hat{S}_{t'}`$, $`\hat{M}`$) with ground-truth maps ($`S_t`$, $`S_{t'}`$, $`M`$). In our weakly supervised setting, we only have access to $`S_t`$. We create missing semantic and change maps distinguishing two cases, depending on if the pair is real or fake. For real pairs, we make the rough assumption that there is no semantically relevant change between the two acquisitions, and define:

\begin{equation}
S_{t'} = S_t~, \quad M = 0~.
\end{equation}

For fake pairs ($`I_{t_i}^i`$, $`I_{t_j'}^j`$), we use the semantic mask of $`I_{t_j}^j`$ as proxy of the unavailable mask $`I_{t_j'}^j`$, again assuming no change between the two acquisitions. In order to generate a change mask $`M`$ from this pair of non-spatially aligned maps $`S_{t_i}`$ and $`S_{t_j}`$, Zheng et al.  apply an XOR operation, which only consider the semantic temporal differences at the pixel level. Instead, we choose to use an object-aware metric that considers changes at the object level. To this end, we generate a change mask by thresholding the sIoU  between connected components of $`S_{t_i}`$ and $`S_{t_j}`$. The sIoU is a well-adopted object-level variation of the traditional intersection over union (IoU). Unlike the conventional IoU, which penalizes fragmented ground truth regions by assigning each prediction a moderate IoU score, the sIoU does not penalize predictions for a segmentation if other predicted segmentations sufficiently cover the remaining ground truth. This enables a robust evaluation of changes at the object level, especially when accounting for shadow occlusions, deformations, or label inconsistencies. Let $`\mathcal{C}_{t_i}`$ and $`\mathcal{C}_{t_j}`$ denote the sets of connected components in $`S_{t_i}`$ and $`S_{t_j}`$ respectively, where each component corresponds to a semantic class. We compute the sIoU between each component $`c^k`$ in $`\mathcal{C}_{t_i}`$, characterized by its semantic category $`k`$, with respect to $`\mathcal{C}_{t_j}`$ as

\begin{equation}
        \text{sIoU}_{\mathcal{C}_{t_j}}(c^k) := \frac{\left| c^k \cap \mathcal{C}_{t_j}(c^k) \right|}{\left| \left(c^k \cup \mathcal{C}_{t_j}(c^k) \right) \setminus \mathcal{A}(c^k) \right|}~,
\end{equation}

with

\begin{equation}
        \mathcal{C}_{t_j}(c^k) = \bigcup_{\substack{\bar{c}^l \in \mathcal{C}_{ t_j}\\ \bar{c}^l \cap c^k \neq \varnothing,\ l=k}}\bar{c}^l \text{ and } 
        \mathcal{A}(c^k) = \bigcup_{\substack{\bar{c}^l \in \mathcal{C}_{t_i} \setminus\{c^k\}\\l=k}} \bar{c}^l.
\end{equation}

The notation $`\left|\cdot\right|`$ denotes the size of a set in terms of number of pixels. The term $`\mathcal{A}(c^k)`$ excludes from the denominator any pixels belonging to other $`\mathcal{C}_{t_i}`$ components that overlap with $`c^k`$ and share the same semantic class $`k`$, preventing interference from nearby objects. During training, we apply this object-level comparison to each fake pair in the batch. For each pair of semantic maps $`(S_{t_i}, S_{t_j})`$, we compute the sIoU for all components in both directions: $`\mathcal{C}_{t_i}`$ with respect to $`\mathcal{C}_{t_j}`$, and vice versa. A pixel location ($`x,y`$) of the binary change mask $`M`$ is labeled as a change if it belongs to any connected component in either $`\mathcal{C}_{t_i}`$ or $`\mathcal{C}_{t_j}`$ for which the sIoU is below a predefined threshold $`\tau`$:

\begin{equation}
    M(x,y) = \begin{cases}
    1 & \text{if } (x,y)\in \displaystyle\bigcup_{\substack{c\in\mathcal{C}_{t_i}\\ \text{sIoU}_{\mathcal{C}_{t_j}(c)<\tau}}} c~, \\
    1 & \text{if } (x,y)\in \displaystyle\bigcup_{\substack{c\in\mathcal{C}_{t_j}\\ \text{sIoU}_{\mathcal{C}_{t_i}(c)<\tau}}} c~, \\
    0 & \text{otherwise}.
\end{cases}
\end{equation}

Fig. 2 provides a visual explanation of the motivations and effects of the object-level change map generation based on the sIoU. In particular, our method is consistent with labeling co-registered real pairs with zero change maps, whereas XOR change maps are sensitive to viewpoint changes.

The rough assumption that real bi-temporal pairs can be labeled with a zero change mask introduces noise in the supervision signal. Indeed, genuine semantic changes may have occurred during the several years separating the original and new acquisitions added in our extended datasets. This leads to mislabeled training samples where changes are incorrectly treated as unchanged, which can degrade model performance. To mitigate this, we adopt an iterative refinement strategy. First, we train a model on the entire dataset. We then use its predictions to identify and filter out image pairs likely to contain changes. In practice, we remove all image pairs for which the model predicts more than 2% of changed pixels. In the next iteration, we train a model from scratch using this cleaned dataset, therefore improving training stability and overall prediction quality. We repeat this for $`N_\text{iter}`$ iterations, where $`N_\text{iter}`$ is a hyperparameter of the method. Ignoring a subset of the dataset by removing noisy or incorrect samples, also known as data cleaning, has been proven successful in other works .

Implementation

Recent analyses show that simple baselines like UNet remain highly competitive for change detection despite the proliferation of complex task-specific architectures . As architectural design is beyond the scope of this paper, we adopt a 3-branch Dual UNet, following Benidir et al . Note that the proposed methodology is architecture agnostic and as such can also be applied to any 3-branch models.

We train our models in a multi-task setting, supervising the semantic segmentation at both dates as well as the binary change detection with corresponding focal losses. All models are trained using the AdamW optimizer  with a 1e-2 weight decay over 100 epochs and a learning rate of 1e-4. We keep the model with the best validation IoU on the segmentation task over the original sample $`t`$. We only use the RGB bands of the images in the datasets, and focus on building change detection when not explicitly stated otherwise. For example, with FLAIR-based datasets, we ignore infrared and elevation bands, merging all non-building classes into a “background” class. Our hyperparameters are set to $`p_\text{real} = 0.25`$, $`\tau = 0.25`$, and $`N_\text{iter} = 3`$.

Experiments

We compare our methodology with four frameworks that use single-temporal semantic segmentation annotations for change detection. These include post-classification with and without multi-temporal data augmentation, synthetic dataset learning, and STAR , which generates fake change pairs by pairing images from different locations in single-temporal datasets. For post-classification, we train a UNet  on the semantic segmentation task and detect building changes by comparing bi-temporal predictions. Bi-temporal data augmentation is achieved by using our extended data during training. Lastly, we consider two Dual UNet  frameworks trained on synthetic datasets: SyntheWorld  and FSC-180k .

We adopt F1-score (F1) and intersection over union (IoU) as evaluation metrics for building change detection. These scores are reported as percentages. We also report the false positive rates (FPR), and the number and size of connected components to evaluate false alarms.

Image It	Image It′	Ground truth	Post-classif.	Post-classif.	FSC-180k	STAR	Ours
				+ temp. aug.	pretraining

Tab. [tab:comparison_results] reports in-domain performance on the b-FLAIR-test and b-FLAIR-test-spot datasets (described in Sec. 3.1), with qualitative examples in Fig. 3. Our method achieves the second-best score on b-FLAIR-test, while performance slightly drops on b-FLAIR-test-spot. We attribute this to two main factors. (1) Our weak supervision relies on temporal augmentation, which can introduce minor misalignments or imperfect change annotations. This produces slightly dilated predictions, smoother boundaries and less defined corners, affecting IoU and F1 scores. (2) The model assumes most regions remain unchanged between acquisitions, biasing predictions toward the no-change class. Given change detection datasets are dominated by the background class, false negatives penalize IoU and F1 more than false positives.

Beyond quantitative results, Fig. 3 confirms that our model is notably more robust in no-change regions, producing fewer false positives. We argue that this behavior is desirable: in large-scale deployments, minimizing false alarms is often more critical than achieving perfectly sharp change boundaries. Therefore, we complement our evaluation with an analysis of the FPR and connected components. Tab. [tab:fpr_and_connected_components] reports FPR, number of change connected components per image pair, and average component size for our predictions, ground truth, and those of Benidir et al. . Component-level evaluation treats small isolated pixels equally to large false detections. While such small artifacts can be easily removed through simple post-processing, larger false detections are more challenging to handle. Following prior work , we report results both before and after applying a 5×5 median filter on the change maps. Across both metrics, our model produces substantially fewer false positives, consistent with the qualitative large-scale results in Fig. [fig:teaser], where our predictions closely align with ground truth while FSC-180k shows numerous false alarms.

We further evaluate zero-shot building change detection performance on LEVIR-CD, WHU-CD, and S2Looking, as reported in Tab. [tab:comparison_results]. Comparisons are made across models trained on the same data sources (FLAIR, SPOT, and IAILD), with the best results highlighted for each setting. Our method achieves the best overall performance on WHU-CD and S2Looking, and consistently outperforms direct competitors within each training domain. The only exception is the b-FLAIR variant on S2Looking, which ranks second by a marginal difference. On LEVIR-CD, the b-FLAIR-spot model surpasses its counterparts, although the b-FLAIR and b-IAILD variants report lower performances. We also point out that these datasets are “very” out-of-domain for models trained using FLAIR-spot because of the difference in resolution ($`\sim1.6`$m vs $`\sim20`$cm).

We finetune a Dual UNet, pretrained with our methodology, on small subsets of 1%, 10%, and 30% of training data for LEVIR-CD and S2Looking. This evaluates the quality of the learned representations when training with very few samples, a common practical scenario. Following the protocol of Benidir et al. , we randomly sample training sets from the target dataset and average results over 10 runs. Tab. 1 reports results for our b-FLAIR and b-IAILD models, compared with FSC-180k and SyntheWorld from . Our approach achieves substantial improvements over the state of the art when only minimal annotations are available (e.g., 1%), while performance converges as more data is available.

Hyperparameters analysis

For the hyperparameters $`\tau`$ and $`p_\text{real}`$, we report the scores of models trained on a subset of b-FLAIR.

Tab. 2 shows adding real pairs in a batch improves performance, with $`p_\text{real}=0`$ yielding the lowest scores. Performance peaks at $`p_\text{real}=0.25`$, while higher values degrade it, likely because overrepresented real pairs with no change bias the model toward predicting no change.

As shown in Tab. 2, a threshold of $`\tau=0.25`$ for sIoU-based change map generation outperforms the logical XOR operation. On the contrary, higher thresholds ($`\tau \in \{0.5, 0.75\}`$) tend to classify components as changed despite significant overlap between building footprints across the two masks, resulting in noisy supervision. Note that $`\tau=1`$ reduces our sIoU-based maps to the OR operation, which Zheng et al.  demonstrated to be inferior to XOR-based change maps for training change detection models. Visualizations comparing these different change map generation variants are provided in the Supplementary Material.

Our methodology relies on mixing bi-temporal image pairs with fake pairs of unaligned images, while assuming by default that all real pairs contain no change. This assumption is incorrect in practice in certain cases, particularly since our extension of existing single-temporal datasets is based on new acquisitions often captured several years apart. Tab. 3 therefore demonstrates the benefit of iterative cleaning of the extended datasets, through filtering of image pairs that would contain significant semantic changes. For b-FLAIR, the second iteration increases F1 by over 10pt and IoU by nearly 14pt on b-FLAIR-test, while gains at the third iteration are smaller. We therefore set $`N_\text{iter} = 3`$ by default. Fig. 4 illustrates filtered pairs, including building changes and blurred sensitive areas. On b-FLAIR-spot, successive iterations have little effect.

Conclusion

In this article, we introduced a novel approach for training remote sensing change detection models by extending single-date datasets and supervising changes through weak temporal supervision, without additional annotations. We assume no change occurs between two observations of the same location while providing change examples by pairing images from different locations. Extensive experiments across multiple datasets show that our methodology learns models with strong out-of-distribution generalization, can be fine-tuned with minimal annotations, and are robust to false positives in “in the wild” scenarios. Qualitative results demonstrate strong alignment with ground truth building construction and demolition maps, bridging the gap between weak supervision and reliable large-scale change monitoring. Overall, our framework offers a practical, annotation-efficient approach to scalable change detection for both aerial and remote sensing observations.

Acknowledgement

This work was funded by AID-DGA (l’Agence de l’Innovation de Défense à la Direction Générale de l’Armement—Ministère des Armées), and was performed using HPC resources from GENCI-IDRIS (grants 2023-AD011011801R3, 2023-AD011012453R2, 2023-AD011012458R2) and from the “Mésocentre” computing center of CentraleSupélec and ENS Paris-Saclay supported by CNRS and Région Île-de-France (http://mesocentre.universite-paris-saclay.fr ). Centre Borelli is also with Université Paris Cité, SSA and INSERM. This work is additionally supported by RGC-GRF project 11309925, Mathematical Formalization of GIS. We thank Etienne Bourgeat, Fabien Poilane and Floryne Roche for their valuable assistance in data curation and annotation.

We first provide additional details on the datasets used in this work (Sec. 6). Then we detail our implementation (Sec. 7). Finally, we report supplementary quantitative and qualitative results (Sec. 8). Our code and datasets will be released publicly for anyone to use.

Data

Fig. [fig:add_visual_examples] shows further examples of triplets of bi-temporal images and their corresponding single-temporal semantic masks, for our three extended datasets, b-FLAIR, b-FLAIR-spot and b-IAILD. We describe them below, and also give a brief description of the benchmark datasets we used (LEVIR-CD, WHU-CD and S2Looking).

b-FLAIR

The original FLAIR dataset is based on the French IGN’s BD ORTHO product  and its associated elevation data. For details on these data products, their pre-processing and their formatting into the FLAIR dataset, please refer to the FLAIR datapaper . The BD ORTHO product is updated every three years on average at each location with new aerial acquisitions. This allowed us to extend FLAIR dataset with a new temporal acquisition for each of its 77,762 patches. For each patch, we extracted a 5-band (RGB, infrared, elevation) patch from the closest temporal acquisition of the BD ORTHO product and its associated elevation data, following the same process as described in . We use the same training splits as in the original paper  (see Fig. 5), except for the hyperparameter studies for which a subset of the data is used (departments 13, 21, 44, 63, 80 for training, and 77 for validation). The average number of days between two acquisitions is 1097 days ($`\thicksim`$3 years). 82.9% of our added images correspond to images acquired after the corresponding original FLAIR patches, while the remaining were acquired before. The average number of days between two acquisitions within the same calendar year is 44 ($`\thicksim`$1.5 months), with 18% of the pairs having more than a 3-month difference within the same calendar year, indicating possible significant seasonal variations for a given image pair. In addition, the differences in the time of the acquisitions within the day—which is 153 min on average—and in the angle of acquisition, introduce significant radiometric, geometric, and shadow-related variations.

b-FLAIR-spot

For all patch location and acquisition date of b-FLAIR, we download a corresponding SPOT-6/7 patch from the ORTHO-SAT database . These are RGB images in 8 bits, at a spatial resolution of 1.5 m/px. We resized the patches at the shape 64$`\times`$64, resulting in an effective spatial resolution of 1.6 m/px. The ORTHO-SAT is an annual product, and we align the acquisition year of the images of b-FLAIR-spot on the corresponding images in b-FLAIR. This allows us to reuse FLAIR annotations for b-FLAIR-spot, provided a nearest-neighbor resampling from 512$`\times`$512 to 64$`\times`$64.

b-IAILD

The original IAILD  public training set is composed of images from 5 cities from the USA (Austin, Chicago, Kitsap) and Austria (Tyrol, Vienna). For each city, the dataset contains 36 tiles of size 5000$`\times`$5000 px at the spatial resolution of 30 cm/px. We downloaded new acquisitions over the same locations from the USGS website² and the respective website of the Austrian provinces of Tyrol³ and Vienna⁴. Years of acquisition of the new added orthophotographies are the following : 2024 for Vienna, 2023 for Chicago, Kitsap and Tyrol, and 2022 for Austin. Because the spatial resolution varies across the areas of interest (15 cm/px for Vienna, 20 cm/px for Tyrol, and 60 cm/px for Austin, Chicago and Kitsap), we resample all the downloaded tiles—as well as the original IAILD tiles—at the common resolution of 60 cm/px. As in , we keep the first 5 tiles of each zone for validation, and use the other 31 for training. The resulting 2500$`\times`$2500 images are split in 100 patches of size 256$`\times`$256, over a grid with an overlap of 6 pixels between adjacent cells. This results in a dataset containing 15500 training pairs (3100 per area) and 2500 validation pairs (500 per area).

b-FLAIR-test & b-FLAIR-test-spot

b-FLAIR-test is composed of 1730 image pairs annotated with a binary building change mask. The images are in the same 5-band 512$`\times`$512 format as FLAIR images and where acquired, processed, and formatted following the procedure described in . The images were extracted from 9 different French administrative departments and do not intersect the FLAIR dataset, which allows for a sound evaluation of methods trained on FLAIR or on our bi-temporal extension of FLAIR (see Fig. 5). The pairs were annotated by photointerpretation experts and verified by a non-expert assessor. The pairs either show new building constructions, or no building change at all ($`\thicksim`$30% of the pairs), but no pair exhibits building destruction. b-FLAIR-test-spot is built from b-FLAIR-test similarly as how we obtained b-FLAIR-spot based on b-FLAIR, i.e. by downloading acquisition at the same date and location for each patch, and downsampling the annotation masks. Example triplets for these two datasets can be visualized in Fig. 6.

LEVIR-CD, WHU-CD and S2Looking

is composed of 637 pairs of aerial images of size 1024$`\times`$1024 at the spatial resolution of 50 cm/px. We keep the original data splits (445 images for training, 64 for validation, and 128 for testing). LEVIR-CD’s images are from 20 different regions in the state of Texas, US, and were acquired between 2002 and 2018. The temporal gap for two acquisitions of the same location in the dataset range from less than a year to 15 years.

is composed of 7620 pairs of aerial images of size 256$`\times`$256 at the spatial resolution of 7.5 cm/px. We keep the original data splits (6096 images for training, 762 for validation, and 762 for testing). WHU-CD’s images are of the area of Christchurch, New Zealand, and were acquired in 2012 (pre-change) and 2016 (post-change).

is composed of 5000 pairs of satellite images of size 1024$`\times`$1024 at a spatial resolution between 50 cm/px and 80 cm/px. We keep the original data splits (3500 images for training, 500 for validation, and 1000 for testing). S2Looking’s images are from 15 areas of interest spread across Europe, Asia, Africa, and North and South America. They were acquired between 2017 and 2020, with a temporal gap ranging from 1 to 3 years between bi-temporal acquisitions.

Tab. [tab:datasets] summarizes the main characteristics of these three datasets, as well as those of our two datasets, b-FLAIR-test and b-FLAIR-test-spot.

Implementation details

A general diagram of the proposed weak temporal supervision approach is shown in Fig. 9. We trained the b-IAILD and b-FLAIR-spot models on a single NVIDIA RTX 6000 GPU with batch sizes of 64 and 256, respectively. Training the b-IAILD model required approximately 26 hours, while b-FLAIR-spot completed in about 5.5 hours. The b-FLAIR dataset contains a significantly larger number of pixels than the other datasets, resulting in substantially higher computational complexity. Consequently, we trained the b-FLAIR models on a multi-GPU setup with 32 NVIDIA V100 GPUs and a batch size of 32, where a complete training took roughly 4 hours. For all models, we use image crops of $`256 \times 256`$ pixels (except b-FLAIR-test, for which images are of size $`64 \times 64`$) and normalize input images using the per-channel mean and standard deviation computed over the entire training dataset, following Benidir et al. . During inference and evaluation, target images are normalized using the same statistics as the training set.

In Fig. 10, we show example of change maps generated with our implemented sIoU-based methods. We visually compare them to change maps computed with logical operations (OR and XOR). While XOR-based masks leave residual change pixels around almost-overlapping buildings, our masks ignore such pixels, though they may correspond to an actual difference in land cover. We believe these “cleaner”, object-based change masks constitute a better supervision in order to train models that are less prone to false alarms.

/>

Additional Results

This section provides additional quantitative and qualitative results that complement those of the main paper.

Different architectures

Figures 7 and 8 report the performance of our method using three different architectures: Dual UNet , SCanNet , and A2Net . A2Net is a lightweight model with 3.52 M parameters, SCanNet contains 27.9 M parameters, while Dual UNet is the largest architecture with 65.05 M parameters. Figure 7 presents the F1 scores on the b-FLAIR-test split for the FSC-180k  approach, the STAR  baseline and our method. Figure 8 reports the same evaluation in a zero-shot setting on the WHU-CD dataset. Our method achieves competitive performance when using Dual UNet and SCanNet backbones across both datasets. In contrast, A2Net performs poorly on WHU-CD. We attribute this behavior to its limited model capacity, which induces a strong bias toward the no-change class. This observation is consistent with the large discrepancy in parameter counts between A2Net (3.5 M) and the larger Dual UNet (65.1 M) and SCanNet (27.9 M) architectures.

Large-scale results

In Fig. 1 of the main paper, we present large-scale results over the metropolitan area of Lille, France (approximately 55.3 km², see Fig. 5 for location on France map). This section provides additional details on the inference setup and extends the analysis to SPOT-6/7 satellite imagery.

The large-scale inference on Lille uses BD ORTHO  bi-temporal aerial images from 2018 and 2021. We extract $`512 \times 512`$ crops with a 6-pixel overlap and feed them to our model pre-trained on b-FLAIR, as well as to the FSC-180K model of Benidir et al. . A $`5 \times 5`$ spatial median filter is applied to the model outputs.

Motivated by the strong performance on very high-resolution aerial data, we also evaluate our method at scale on more widely accessible satellite imagery. We extract $`64 \times 64`$ SPOT-6/7 crops from IGN’s ORTHO-SAT database  at 1.5 m/pixel resolution with a 5% overlap. We process these crops with the same inference pipeline and report results for both our b-FLAIR-spot model and its STAR counterpart in Fig. 11. As shown, the STAR model produces many false positives, similar to FSC-180K on the BD ORTHO experiment. In contrast, our model is substantially more robust to false alarms and successfully highlights most large change regions, although it remains more susceptible to false negatives.

Additional qualitative results

Tab. [tab:comparison_results_supp_mat] provides additional FPR results for all evaluated methods as an extension of the results presented in the main paper. Our method consistently leads to less false positive than baselines, though performing on par with them w.r.t other metrics on most datasets. This is especially the case on in-domain datasets and on WHU-CD. In Fig. 12, we can clearly see that, on b-FLAIR-test-spot, our method mostly predict no change, whereas its change predictions are accurate (see first row). On the contrary, the other methods show a significant number of false alarms, making them unusable in practical applications. This behavior can also be observed with the zero-shot qualitative results on LEVIR-CD (Fig. 13) and WHU-CD (Fig. 14). On S2Looking (Fig. 15), all methods globally perform poorly due to the important domain gap with the training data.

Detailed impact of the iterative refinement

Tab. [tab:iteration_comparison_bflairspot_full] reports the results for each iteration of our models trained on b-FLAIR, b-FLAIR-spot, and b-IAILD, respectively. Although we set the hyperparameter $`N_\text{iter}=3`$ based solely on in-domain dataset results, we also report the impact of iterative refinement on out-of-domain zero-shot performance. With the exception of the b-IAILD model on LEVIR-CD, where the first iteration significantly outperforms subsequent ones, removing detected change pairs from the dataset generally benefits out-of-domain performance. The number of filtered pairs at each iteration suggests that additional iterations could further improve model performance. However, given the marginal improvement between the second and third iterations on in-domain evaluation and considering computational cost constraints, we did not explore values of $`N_\text{iter}`$ greater than 3.

Image It	Image It′	Ground truth	Post-classif.	Post-classif.	STAR	Ours
				+ temporal aug.