S2Looking: A Satellite Side-Looking Dataset for Building Change Detection

Building-change detection underpins many important applications, especially in the military and crisis-management domains. Recent methods used for change detection have shifted towards deep learning, which depends on the quality of its training data. The assembly of large-scale annotated satellite imagery datasets is therefore essential for global building-change surveillance. Existing datasets almost exclusively offer near-nadir viewing angles. This limits the range of changes that can be detected. By offering larger observation ranges, the scroll imaging mode of optical satellites presents an opportunity to overcome this restriction. This paper therefore introduces S2Looking, a building-change-detection dataset that contains large-scale side-looking satellite images captured at various off-nadir angles. The dataset consists of 5000 bitemporal image pairs of rural areas and more than 65,920 annotated instances of changes throughout the world. The dataset can be used to train deep-learning-based change-detection algorithms. It expands upon existing datasets by providing (1) larger viewing angles; (2) large illumination variances; and (3) the added complexity of rural images. To facilitate the use of the dataset, a benchmark task has been established, and preliminary tests suggest that deep-learning algorithms find the dataset significantly more challenging than the closest-competing near-nadir dataset, LEVIR-CD+. S2Looking may therefore promote important advances in existing building-change-detection algorithms.

An Image Change Detection Algorithm Based on Multi-Feature Self-Attention Fusion Mechanism UNet Network

Various UNet architecture-based image change detection algorithms promote the development of image change detection, but there are still some defects. First, under the encoder–decoder framework, the low-level features are extracted many times in multiple dimensions, which generates redundant information; second, the relationship between each feature layer is not modeled so sufficiently that it cannot produce the optimal feature differentiation representation. This paper proposes a remote image change detection algorithm based on the multi-feature self-attention fusion mechanism UNet network, abbreviated as MFSAF UNet (multi-feature self-attention fusion UNet). We attempt to add multi-feature self-attention mechanism between the encoder and decoder of UNet to obtain richer context dependence and overcome the two above-mentioned restrictions. Since the capacity of convolution-based UNet network is directly proportional to network depth, and a deeper convolutional network means more training parameters, so the convolution of each layer of UNet is replaced as a separated convolution, which makes the entire network to be lighter and the model’s execution efficiency is slightly better than the traditional convolution operation. In addition to these, another innovation point of this paper is using preference to control loss function and meet the demands for different accuracies and recall rates. The simulation test results verify the validity and robustness of this approach.

COMPARISON OF PIXEL-BASED AND FEATURE-BASED APPROACH FOR SMALL OBJECT CHANGE DETECTION

Satellite image resolution has evolved to daily revisit and sub-meter GSD. Main targets of previous remote sensing were forest, vegetation, damage area by disasters, land use and land cover. Developments in satellite images have brought expectations on more sophisticated and various change detection of objects. Accordingly, we focused on unsupervised change detection of small objects, such as vehicles and ships. In this paper, existing change detection methods were applied to analyze their performances for pixel-based and feature-based change of small objects. We used KOMPSAT-3A images for tests. Firstly, we applied two change detection algorithms, MAD and IR-MAD, which are most well-known pixel-based change detection algorithms, to the images. We created a change magnitude map using the change detection methods. Thresholding was applied to determine change and non-change pixels. Next, the satellite images were transformed as 8-bit images for extracting feature points. We extracted feature points using SIFT and SURF methods to analyze feature-based change detection. We assumed to remove false alarms by eliminating feature points of non-changed objects. Therefore, we applied a feature-based matcher and matched feature points on identical image locations were eliminated. We used non-matched feature points for change/non-change analysis. We observed changes by creating a 5x5 size ROI around extracted feature points in the change/non-change map. We determined that change has occurred on feature points if the rate of change pixels with ROI was more than 50%. We analyzed the performance of pixel-based and feature-based change detection using ground truths. The F1-score, AUC value, and ROC were used to compare the performance of change detection. Performance showed that feature-based approaches performed better than pixel-based approaches.

Estimating VAIA Windstorm Damaged Forest Area in Italy Using Time Series Sentinel-2 Imagery and Continuous Change Detection Algorithms

Mapping forest disturbances is an essential component of forest monitoring systems both to support local decisions and for international reporting. Between the 28 and 29 October 2018, the VAIA storm hit the Northeast regions of Italy with wind gusts exceeding 200 km h−1. The forests in these regions have been seriously damaged. Over 490 Municipalities in six administrative Regions in Northern Italy registered forest damages caused by VAIA, that destroyed or intensely damaged forest stands spread over an area of 67,000 km2. The present work tested the use of two continuous change detection algorithms, i.e., the Bayesian estimator of abrupt change, seasonal change, and trend (BEAST) and the continuous change detection and classification (CCDC) to map and estimate forest windstorm damage area using a normalized burned ration (NBR) time series calculated on three years Sentinel-2 (S2) images collection (i.e., January 2017–October 2019). We analyzed the accuracy of the maps and the damaged forest area using a probability-based stratified estimation within 12 months after the storm with an independent validation dataset. The results showed that close to the storm (i.e., 1 to 6 months November 2018–March 2019) it is not possible to obtain accurate results independently of the algorithm used, while accurate results were observed between 7 and 12 months from the storm (i.e., May 2019 – October 2019) in terms of Standard Error (SE), percentage SE (SE%), overall accuracy (OA), producer accuracy (PA), user accuracy (UA), and gmean for both BEAST and CCDC (SE< 3725.3 ha , SE% < 9.69 , OA > 89.7, PA and UA > 0.87, gmean > 0.83).

DETECÇÃO DE ÁREAS DE FLORESTAS INVARIANTES EM SÉRIES TEMPORAIS UTILIZANDO RANDOM FOREST

O surgimento de algoritmos de detecção de mudanças na vegetação na última década é impressionante. Mas os resultados gerados ainda possuem ruído que precisa ser tratado com a utilização de resultados de outros mapeamentos de cobertura vegetal. Além disso, a necessidade de gerar classes de uso do solo invariantes é importante para o melhor entendimento de processos que ocorrem em áreas florestais. Pensando nisso, este trabalho busca criar uma nova forma de mapear essas áreas invariáveis que possam ser utilizadas para mascarar ruídos e também como subsídio para outros estudos de conservação e restauração. A metodologia proposta aqui usa a plataforma Google Earth Engine e um algoritmo de aprendizado de máquina: o Random Forest, para classificar áreas de floresta invariáveis usando todo o acervo de imagens da série temporal Landsat, de uma só vez. Os resultados mostraram que a nova abordagem teve melhor desempenho do que o uso de técnicas mais tradicionais como a agregação de mapeamentos de uso e cobertura anuais, com uma acurácia global de 91,7%. O trabalho busca ainda contribuir com a comunidade de sensoriamento remoto ao apresentar, após exaustivos testes, as melhores opções de variáveis a serem utilizadas neste tipo de classificação. Palavras-chave: Séries Temporais, Detecção de Mudanças, Florestas, Google Earth Engine, Random Forest.DETECTION OF INVARIANT VEGETATION AREAS IN TIME SERIES USING RANDOM FOREST ALGORITHMAbstract: The emergence of vegetation change detection algorithms in the last decade is impressive. But the results still have a lot of noise that needs to be cleaned. And the data cleaning process still uses other landcover mapping results. Besides that, the necessity to generate invariant land use classes is important to know particularly to forest areas. Thinking about that, this paper seeks to create a new form of mapping these invariant areas that can be used to mask noise and as an input on other conservation and restoration studies. The methodology proposed here uses the Google Earth Engine platform and a Random Forest algorithm to classify invariant forest areas using all the image’s collection in the time series at once. The results showed that the new approach performed better than the use of more traditional techniques such as the aggregation of annual land-use and land-cover mappings, with an overall accuracy of 91.7%. Also, this paper seeks to contribute to the remote sensing community showing after exhaustive testing, good options of variables to use on this type of work. Keywords: Time Series, Change Detection, Forests, Google Earth Engine, Random Forest.DETECCIÓN DE ÁREAS DE VEGETACIÓN INVARIANTES EN SÉRIES TEMPORALES UTILIZANDO ALGORITMO RANDOM FORESTResumen: La aparición de algoritmos de detección de cambios en la vegetación en la última década es impresionante. Pero los resultados todavía tienen muchos ruidos que deben ser eliminados. Además, el proceso de limpieza de datos se basa en otros mapas de cobertura de la tierra. Además de eso, es importante conocer la necesidad de generar clases de uso de la tierra invariables, particularmente en las áreas forestales. Pensando en eso, este artículo busca crear una nueva forma de mapear estas áreas invariantes que se pueden utilizar para enmascarar el ruido y como un aporte para otros estudios de conservación y restauración. La metodología propuesta aquí utiliza la plataforma Google Earth Engine y un algoritmo de aprendizaje de máquina: o Random Forest para clasificar áreas invariantes de bosque, utilizando a la vez todas las imágenes de la serie temporal Landsat. Los resultados encontraron que el nuevo enfoque tuvo mejor desempeño que el uso de técnicas tradicionales, con una precisión global del 91,7%. Este trabajo busca además contribuir con la comunidad de la teledetección, mostrando mediante de exhaustivas pruebas, mejores opciones de variables para utilizar en este tipo de clasificación. Palabras clave: Series de Tiempo, Detección de Cambios, Bosques, Google Earth Engine, Random Forest.

Detection of Removed Objects in 3D Meshes Using Up-to-Date Images for Mixed-Reality Applications

Precise knowledge of the real environment is a prerequisite for the integration of the real and virtual worlds in mixed-reality applications. However, real-time updating of a real environment model is a costly and difficult process; therefore, hybrid approaches have been developed: An updated world model can be inferred from an offline acquisition of the 3D world, which is then updated online using live image sequences under the condition of developing fast and robust change detection algorithms. Current algorithms are biased toward object insertion and often fail in object removal detection; in an environment where there is uniformity in the background—in color and intensity—the disappearances of foreground objects between the 3D scan of a scene and the capture of several new pictures of said scene are difficult to detect. The novelty of our approach is that we circumvent this issue by focusing on areas of least change in parts of the scene that should be occluded by the foreground. Through experimentation on realistic datasets, we show that this approach results in better detection and localization of removed objects. This technique can be paired with an insertion detection algorithm to provide a complete change detection framework.

Diversity of Algorithm and Spectral Band Inputs Improves Landsat Monitoring of Forest Disturbance

Disturbance monitoring is an important application of the Landsat times series, both to monitor forest dynamics and to support wise forest management at a variety of spatial and temporal scales. In the last decade, there has been an acceleration in the development of approaches designed to put the Landsat archive to use towards these causes. Forest disturbance mapping has moved from using individual change-detection algorithms, which implement a single set of decision rules that may not apply well to a range of scenarios, to compiling ensembles of such algorithms. One approach that has greatly reduced disturbance detection error has been to combine individual algorithm outputs in Random Forest (RF) ensembles trained with disturbance reference data, a process called stacking (or secondary classification). Previous research has demonstrated more robust and sensitive detection of disturbance using stacking with both multialgorithm ensembles and multispectral ensembles (which make use of a single algorithm applied to multiple spectral bands). In this paper, we examined several additional dimensions of this problem, including: (1) type of algorithm (represented by processes using one image per year vs. all historical images); (2) spectral band choice (including both the basic Landsat reflectance bands and several popular indices based on those bands); (3) number of algorithm/spectral-band combinations needed; and (4) the value of including both algorithm and spectral band diversity in the ensembles. We found that ensemble performance substantially improved per number of model inputs if those inputs were drawn from a diversity of both algorithms and spectral bands. The best models included inputs from both algorithms, using different variants of shortwave-infrared (SWIR) and near-infrared (NIR) reflectance. Further disturbance detection improvement may depend upon the development of algorithms which either interrogate SWIR and NIR in new ways or better highlight disturbance signals in the visible wavelengths.