Source-linked AI summary
Urban Change Detection for Multispectral Earth Observation Using Convolutional Neural Networks
Rodrigo Caye Daudt, Bertrand Le Saux, Alexandre Boulch, Yann Gousseau
TL;DR
The paper tackles limited labeled data for supervised urban change detection in multispectral Sentinel-2 imagery. It introduces the OSCD benchmark and trains Siamese and Early Fusion CNNs from scratch, finding that Early Fusion generally performs better and added channels generally improve classification, while the dataset’s resolution limits smaller-change detection.
Problem
Open labeled Sentinel-2 datasets are lacking for quantitative evaluation and for training supervised change detection systems.
Method
The paper creates the openly available OSCD dataset and trains Siamese and Early Fusion CNN architectures end-to-end on multispectral image pairs.
Results
Early Fusion networks generally outperform corresponding Siamese networks, while adding spectral channels generally improves classification performance nonlinearly and architecture-dependently.
Takeaways & Limitations
OSCD provides a benchmark for comparing urban change detection algorithms, and the trained CNNs detect urban changes using the dataset alone.
Takeaways & Limitations
Sentinel-2’s relatively low resolution supports detection of large buildings but may obscure smaller changes and produce differing manual change maps.
Abstract
from arXiv · showhide
The Copernicus Sentinel-2 program now provides multispectral images at a global scale with a high revisit rate. In this paper we explore the usage of convolutional neural networks for urban change detection using such multispectral images. We first present the new change detection dataset that was used for training the proposed networks, which will be openly available to serve as a benchmark. The Onera Satellite Change Detection (OSCD) dataset is composed of pairs of multispectral aerial images, and the changes were manually annotated at pixel level. We then propose two architectures to detect changes, Siamese and Early Fusion, and compare the impact of using different numbers of spectral channels as inputs. These architectures are trained from scratch using the provided dataset.
1. INTRODUCTION
The paper addresses the lack of open labeled Sentinel-2 datasets for supervised urban change detection by introducing OSCD and proposing two end-to-end CNN architectures.
- Open Sentinel-2 imagery is abundant, but open labeled datasets are lacking for quantitative comparison and evaluation of change detection algorithms.
- This data gap limits the application of supervised learning and CNNs to change detection despite their success in computer vision and image processing.
- The OSCD dataset provides openly available Sentinel-2 image pairs with pixel-wise labels for training, testing, and comparing urban change detection algorithms.
- The proposed CNN architectures learn end-to-end urban change detection from multispectral image pairs in a fully supervised manner.
- The paper presents the dataset, describes the supervised learning methods, and evaluates their reach and limitations.
2. DATASET
The OSCD dataset provides standardized, pixel-wise urban change labels from Sentinel-2 image pairs across diverse regions, while its resolution and temporal coverage constrain detectable changes and class balance.
- Dataset objective: OSCD standardizes comparison of urban change detection algorithms using pixel-wise labels that include urban growth while excluding natural changes.The dataset is intended for both algorithm benchmarking and supervised learning.
- Dataset construction: Twenty-four worldwide regions were selected from Sentinel-2 imagery, covering approximately 600x600 pixels at 10m resolution and 13 spectral bands.The regions represent various levels of urbanization where urban changes were visible.
- Dataset construction: Image selection and cloud control required manual verification because Sentinel-2 data often contain black regions and varying cloud conditions.The raw-data variability prevented completely scripted patch generation.
- Dataset construction: Ground-truth maps were manually produced from true-colour image pairs after additional image registration, using the older image as reference.The newer image was transformed to align with the older one.
- Challenges and limitations: Sentinel-2’s relatively low resolution may reveal large buildings but obscure smaller structures, extensions, or road-lane additions, causing analyst disagreement.Different analysts may therefore generate different change maps.
- Challenges and limitations: OpenStreetMap-based automated labeling was unsuccessful because map additions did not reliably correspond to construction dates and older map dates were imprecise.Before 2017, many years had only one available map.
- Challenges and limitations: Sentinel-2 coverage limits temporal separation to roughly two and a half years or less, producing many more no-change than change pixels.The satellite data cannot extend earlier than June 2015.
3. CHANGE DETECTION METHODS
The paper trains end-to-end CNNs to classify urban change from paired image patches, comparing Early Fusion and Siamese architectures with adjustable spectral inputs.
- Method overview: The networks classify the central pixel of a 15x15 patch pair as change or no change using supervised end-to-end learning.The intended labels distinguish artificialization changes from natural changes.
- Method overview: Each network receives two 15x15xC patches and outputs probabilities for the change and no-change classes.Thresholding the change probability beyond 0.5 can adjust the balance between false positives and false negatives.
- Early Fusion: Early Fusion concatenates the two patches into a single 15x15x2C input before seven convolutional and two fully connected layers.Its final softmax layer has two outputs corresponding to change and no change.
- Siamese: The Siamese network processes the two patches through parallel four-layer convolutional branches with shared weights before concatenating their outputs.Two fully connected layers then produce the two class outputs.
- Inference: Full-image change maps are generated by classifying patches with a larger extraction stride and combining overlapping votes using a Gaussian distance weighting.Pixels closer to a patch center receive greater voting weight.
4. RESULTS
The evaluation trains and tests the CNNs on split OSCD images while varying preprocessing and spectral inputs; reported results favor Early Fusion and generally benefit from additional channels.
- Experimental setup: Fourteen images were used for training and ten for testing, with flips and 90-degree rotations augmenting the training patches.Augmentation addresses the dataset’s limited size.
- Experimental setup: Channels with resolutions below 10m were upsampled so all spectral inputs could be concatenated with aligned pixels.This preprocessing enables comparison across channel configurations.
- Experimental setup: Training assigns higher penalty weight to the change class, with weights inversely proportional to class-example counts.This addresses the larger number of no-change pixels.
- Experimental setup: The channel comparison uses RGB with 3 channels, RGB plus infrared with 4, 10 channels through 20m resolution, and all 13 channels through 60m resolution.These cases test how spectral input breadth affects classification.
- Results: CNN variants outperform the cited difference-image methods, while Early Fusion generally exceeds its corresponding Siamese network.Adding colour channels generally improves classification, but the effect is nonlinear and architecture-dependent.
- Results: Generated change maps indicate that the trained networks detect urban changes in the test images.The figures also support visual comparison of architectures and input-channel counts.
5. CONCLUSION
The paper introduces the openly available OSCD dataset and two supervised CNN approaches for urban change detection, achieving excellent test performances without external image data. It identifies dataset expansion and broader change-labeling and temporal-sequence capabilities as future directions.
- 5. CONCLUSION: The authors present the openly available Onera Satellite Change Detection dataset for urban change detection.The dataset is based on Sentinel-2 imagery and supports supervised development and evaluation.
- 5. CONCLUSION: Two CNN approaches are trained fully supervised on image pairs from the dataset and achieve excellent test performances without other image data.The approaches are evaluated using the presented dataset alone.
- 5. CONCLUSION: Future work includes enlarging the dataset across cities and imaging modalities, including possible Sentinel-1 imagery.The stated expansion targets both geographic coverage and sensor diversity.
- 5. CONCLUSION: The authors propose fully convolutional networks to reduce patch effects when generating labels for all image pixels.This is presented as a next step for the current methods.
- 5. CONCLUSION: Further directions include semantic labeling of changes and change detection on image sequences rather than image pairs.These extensions are intended to improve interpretation and broaden the temporal input setting.