Unsupervised hashing for remote sensing (RS) image retrieval first extracts image features and then use these features to construct supervised information (e.g., pseudo-labels) to train hashing networks. Existing methods usually regard RS images as natural images to extract unisource features. However, these features only contain partial information about ground objects and cannot produce reliable pseudo-labels. In addition, existing methods only generate a pseudo single-label to annotate each RS image, which cannot accurately represent multiple scenes in a RS image. To address these drawbacks, this paper proposes a new Multisource data reconstruction-based deep unsupervised Hashing method, called MrHash, which explores the characteristics of RS images to construct reliable pseudo-labels. In particular, we first use geographic coordinates to obtain different satellite images and develop a novel autoencoder network to extract multisource features from these images. Then pseudo multi-labels are designed to deal with the coexistence of multiple scenes in a single image. These labels are generated by a custom probability function with extracted multisource features. Finally, we propose a novel multi-semantic hash loss by using the Kull-back–Leibler (KL) divergence to preserve the semantic similarity of these pseudo multi-labels in Hamming space. Our newly developed MrHash only uses multisource images to construct supervised information, and hash code generation still relies on a unisource input image. Experiments on benchmark datasets clearly show the superiority of the proposed method over state-of-the-art baselines. https://github.com/sunyuxi/MrHash.

In recent years, the world witnessed many devastating wildfires that resulted in destructive human and environmental impacts across the globe. Emergency response and rapid response for mitigation calls for effective approaches for near real-time wildfire monitoring. Capable of penetrating clouds and smoke, and imaging day and night, Synthetic Aperture Radar (SAR) can play a critical role in wildfire monitoring. In this communication, we investigated and demonstrated the potential of Sentinel-1 SAR time series with a deep learning framework for near real-time wildfire progression monitoring. The deep learning framework, based on a Convolutional Neural Network (CNN), is developed to detect burnt areas automatically using every new SAR image acquired during the wildfires and by exploiting all available pre-fire SAR time series to characterize the temporal backscatter variations. The results show that Sentinel-1 SAR backscatter can detect wildfires and capture their temporal progression as demonstrated for three large and impactful wildfires: the 2017 Elephant Hill Fire in British Columbia, Canada, the 2018 Camp Fire in California, USA, and the 2019 Chuckegg Creek Fire in northern Alberta, Canada. Compared to the traditional log-ratio operator, CNN-based deep learning framework can better distinguish burnt areas with higher accuracy. These findings demonstrate that spaceborne SAR time series with deep learning can play a significant role for near real-time wildfire monitoring when the data becomes available at daily and hourly intervals with the launches of RADARSAT Constellation Missions in 2019, and SAR CubeSat constellations.

Mapping Earth’s surface and its rapid changes with remotely sensed data is a crucial task to understand the impact of an increasingly urban world population on the environment. However, the impressive amount of available Earth observation data is only marginally exploited in common classifications. In this study, we use the computational power of Google Earth Engine and Google Cloud Platform to generate an oversized feature set in which we explore feature importance and analyze the influence of dimensionality reduction methods to object-based land cover classification with Support Vector Machines. We propose a methodology to extract the most relevant features and optimize an SVM classifier hyperparameters to achieve higher classification accuracy. The proposed approach is evaluated in two different urban study areas of Stockholm and Beijing. Despite different training set sizes in the two study sites, the averaged feature importance ranking showed similar results for the top-ranking features. In particular, Sentinel-2 NDVI, NDWI, and Sentinel-1 VV temporal means are the highest ranked features and the experiment results strongly indicated that the fusion of these features improved the separability between urban land cover and land use classes. Overall classification accuracies of 94% and 93% were achieved in Stockholm and Beijing study sites, respectively. The test demonstrated the viability of the methodology in a cloud-computing environment to incorporate dimensionality reduction as a key step in the land cover classification process, which we consider essential for the exploitation of the growing Earth observation big data. To encourage further research and development of reliable workflows, we share our datasets and publish the developed Google Earth Engine and Python scripts as free and open-source software.

Accurate burned area information is needed to assess the impacts of wildfires on people, communities, and natural ecosystems. Various burned area detection methods have been developed using satellite remote sensing measurements with wide coverage and frequent revisits. Our study aims to expound on the capability of deep learning (DL) models for automatically mapping burned areas from uni-temporal multispectral imagery. Specifically, several semantic segmentation network architectures, i.e., U-Net, HRNet, Fast-SCNN, and DeepLabv3+, and machine learning (ML) algorithms were applied to Sentinel-2 imagery and Landsat-8 imagery in three wildfire sites in two different local climate zones. The validation results show that the DL algorithms outperform the ML methods in two of the three cases with the compact burned scars, while ML methods seem to be more suitable for mapping dispersed burn in boreal forests. Using Sentinel-2 images, U-Net and HRNet exhibit comparatively identical performance with higher kappa (around 0.9) in one heterogeneous Mediterranean fire site in Greece; Fast-SCNN performs better than others with kappa over 0.79 in one compact boreal forest fire with various burn severity in Sweden. Furthermore, directly transferring the trained models to corresponding Landsat-8 data, HRNet dominates in the three test sites among DL models and can preserve the high accuracy. The results demonstrated that DL models can make full use of contextual information and capture spatial details in multiple scales from fire-sensitive spectral bands to map burned areas. Using only a post-fire image, the DL methods not only provide automatic, accurate, and bias-free large-scale mapping option with cross-sensor applicability, but also have potential to be used for onboard processing in the next Earth observation satellites.

Producing accurate land cover maps is time-consuming and estimating land cover changes between two generated maps is affected by error propagation. The increased availability of analysis-ready Earth Observation (EO) data and the access to big data analytics capabilities on Google Earth Engine (GEE) have opened the opportunities for continuous monitoring of environment changing patterns. This research proposed a framework for analyzing urban land cover change trajectories based on Landsat time series and LandTrendr, a well-known spectral-temporal segmentation algorithm for land-based disturbance and recovery detection. The framework involved the use of baseline land cover maps generated at the beginning and at the end of the considered time interval and proposed a new approach to merge the LandTrendr results using multiple indices for reconstructing dense annual land cover maps within the considered period. A supervised support vector machine (SVM) classification was first performed on the two Landsat scenes, respectively, acquired in 1987 and 2019 over Kigali, Rwanda. The resulting land cover maps were then imported in the GEE platform and used to label the interannual LandTrendr-derived changes. The changes in duration, year, and magnitude of land cover disturbance were derived from six different indices/bands using the LandTrendr algorithm. The interannual change LandTrendr results were then combined using a robust estimation procedure based on principal component analysis (PCA) for reconstructing the annual land cover change maps. The produced yearly land cover maps were assessed using validation data and the GEE-based Area Estimation and Accuracy Assessment (Area²) application. The results were used to study the Kigali’s urbanization in the last three decades since 1987. The results illustrated that from 1987 to 1998, the urbanization was characterized by slow development, with less than a 2% annual growth rate. The post-conflict period was characterized by accelerated urbanization, with a 4.5% annual growth rate, particularly from 2004 onwards due to migration flows and investment promotion in the construction industry. The five-year interval analysis from 1990 to 2019 revealed that impervious surfaces increased from 4233.5 to 12116 hectares, with a 3.7% average annual growth rate. The proposed scheme was found to be cost-effective and useful for continuously monitoring the complex urban land cover dynamics, especially in environments with EO data affordability issues, and in data-sparse regions.

Urbanization is progressing rapidly around the world. With sub-weekly revisits at global scale, Sentinel-1 synthetic aperture radar (SAR) and Sentinel-2 multispectral imager (MSI) data can play an important role for monitoring urban sprawl to support sustainable development. In this letter, we proposed an urban change detection (CD) approach featuring a new network architecture for the fusion of SAR and optical data. Specifically, a dual stream concept was introduced to process different data modalities separately, before combining extracted features at a later decision stage. The individual streams are based on U-Net architecture that is one of the most popular fully convolutional networks used for semantic segmentation. The effectiveness of the proposed approach was demonstrated using the Onera Satellite CD (OSCD) dataset. The proposed strategy outperformed other U-Net-based approaches in combination with unimodal data and multimodal data with feature level fusion. Furthermore, our approach achieved state-of-the-art performance on the urban CD problem posed by the OSCD dataset. Our Sentinel-1 SAR data and code are available on https://github.com/SebastianHafner/DS_UNet .