The PASCAL Visual Object Classes Challenge

A freely available dataset produced from three million Landsat satellite images reveals substantial changes in the distribution of global surface water over the past 32 years and their causes, from climate change to human actions. The distribution of surface water has been mapped globally, and local-to-regional studies have tracked changes over time. But to date, there has been no global and methodologically consistent quantification of changes in surface water over time. Jean-Francois Pekel and colleagues have analysed more than three million Landsat images to quantify month-to-month changes in surface water at a resolution of 30 metres and over a 32-year period. They find that surface waters have declined by almost 90,000 square kilometres—largely in the Middle East and Central Asia—but that surface waters equivalent to about twice that area have been created elsewhere. Drought, reservoir creation and water extraction appear to have driven most of the changes in surface water over the past decades. The location and persistence of surface water (inland and coastal) is both affected by climate and human activity1 and affects climate2,3, biological diversity4 and human wellbeing5,6. Global data sets documenting surface water location and seasonality have been produced from inventories and national descriptions7, statistical extrapolation of regional data8 and satellite imagery9,10,11,12, but measuring long-term changes at high resolution remains a challenge. Here, using three million Landsat satellite images13, we quantify changes in global surface water over the past 32 years at 30-metre resolution. We record the months and years when water was present, where occurrence changed and what form changes took in terms of seasonality and persistence. Between 1984 and 2015 permanent surface water has disappeared from an area of almost 90,000 square kilometres, roughly equivalent to that of Lake Superior, though new permanent bodies of surface water covering 184,000 square kilometres have formed elsewhere. All continental regions show a net increase in permanent water, except Oceania, which has a fractional (one per cent) net loss. Much of the increase is from reservoir filling, although climate change14 is also implicated. Loss is more geographically concentrated than gain. Over 70 per cent of global net permanent water loss occurred in the Middle East and Central Asia, linked to drought and human actions including river diversion or damming and unregulated withdrawal15,16. Losses in Australia17 and the USA18 linked to long-term droughts are also evident. This globally consistent, validated data set shows that impacts of climate change and climate oscillations on surface water occurrence can be measured and that evidence can be gathered to show how surface water is altered by human activities. We anticipate that this freely available data will improve the modelling of surface forcing, provide evidence of state and change in wetland ecotones (the transition areas between biomes), and inform water-management decision-making.

High-resolution mapping of global surface water and its long-term changes

Remote sensing image scene classification plays an important role in a wide range of applications and hence has been receiving remarkable attention. During the past years, significant efforts have been made to develop various datasets or present a variety of approaches for scene classification from remote sensing images. However, a systematic review of the literature concerning datasets and methods for scene classification is still lacking. In addition, almost all existing datasets have a number of limitations, including the small scale of scene classes and the image numbers, the lack of image variations and diversity, and the saturation of accuracy. These limitations severely limit the development of new approaches especially deep learning-based methods. This paper first provides a comprehensive review of the recent progress. Then, we propose a large-scale dataset, termed "NWPU-RESISC45", which is a publicly available benchmark for REmote Sensing Image Scene Classification (RESISC), created by Northwestern Polytechnical University (NWPU). This dataset contains 31,500 images, covering 45 scene classes with 700 images in each class. The proposed NWPU-RESISC45 (i) is large-scale on the scene classes and the total image number, (ii) holds big variations in translation, spatial resolution, viewpoint, object pose, illumination, background, and occlusion, and (iii) has high within-class diversity and between-class similarity. The creation of this dataset will enable the community to develop and evaluate various data-driven algorithms. Finally, several representative methods are evaluated using the proposed dataset and the results are reported as a useful baseline for future research.

Remote Sensing Image Scene Classification: Benchmark and State of the Art

Remote sensing image scene classification plays an important role in a wide range of applications and hence has been receiving remarkable attention. During the past years, significant efforts have been made to develop various data sets or present a variety of approaches for scene classification from remote sensing images. However, a systematic review of the literature concerning data sets and methods for scene classification is still lacking. In addition, almost all existing data sets have a number of limitations, including the small scale of scene classes and the image numbers, the lack of image variations and diversity, and the saturation of accuracy. These limitations severely limit the development of new approaches especially deep learning-based methods. This paper first provides a comprehensive review of the recent progress. Then, we propose a large-scale data set, termed “NWPU-RESISC45,” which is a publicly available benchmark for REmote Sensing Image Scene Classification (RESISC), created by Northwestern Polytechnical University (NWPU). This data set contains 31 500 images, covering 45 scene classes with 700 images in each class. The proposed NWPU-RESISC45 1) is large-scale on the scene classes and the total image number; 2) holds big variations in translation, spatial resolution, viewpoint, object pose, illumination, background, and occlusion; and 3) has high within-class diversity and between-class similarity. The creation of this data set will enable the community to develop and evaluate various data-driven algorithms. Finally, several representative methods are evaluated using the proposed data set, and the results are reported as a useful baseline for future research.

Deep learning methods such as convolutional neural networks (CNNs) can deliver highly accurate classification results when provided with large enough data sets and respective labels. However, using CNNs along with limited labeled data can be problematic, as this leads to extensive overfitting. In this letter, we propose a novel method by considering a pretrained CNN designed for tackling an entirely different classification problem, namely, the ImageNet challenge, and exploit it to extract an initial set of representations. The derived representations are then transferred into a supervised CNN classifier, along with their class labels, effectively training the system. Through this two-stage framework, we successfully deal with the limited-data problem in an end-to-end processing scheme. Comparative results over the UC Merced Land Use benchmark prove that our method significantly outperforms the previously best stated results, improving the overall accuracy from 83.1% up to 92.4%. Apart from statistical improvements, our method introduces a novel feature fusion algorithm that effectively tackles the large data dimensionality by using a simple and computationally efficient approach.

/pdf/deep-learning-earth-observation-classification-using-1b6xulvhdp.pdf

Deep Learning Earth Observation Classification Using ImageNet Pretrained Networks

A general framework for processing high and very-high resolution imagery in support of a Global Human Settlement Layer (GHSL) is presented together with a discussion on the results of the first operational test of the production workflow. The test involved the mapping of 24.3 million km2 of the Earth surface spread in four continents, corresponding to an estimated population of 1.3 billion people in 2010. The resolution of the input image data ranges from 0.5 to 10 meters, collected by a heterogeneous set of platforms including satellite SPOT (2 and 5), CBERS 2B, RapidEye (2 and 4), WorldView (1 and 2), GeoEye 1, QuickBird 2, Ikonos 2, and airborne sensors. Several imaging modes were tested including panchromatic, multispectral and pan-sharpened images. A new fully automatic image information extraction, generalization and mosaic workflow is presented that is based on multiscale textural and morphological image features extraction. New image feature compression and optimization are introduced, together with new learning and classification techniques allowing for the processing of HR/VHR image data using low-resolution thematic layers as reference. A new systematic approach for quality control and validation allowing global spatial and thematic consistency checking is proposed and applied. The quality of the results are discussed by sensor, band, resolution, and eco-regions. Critical points, lessons learned and next steps are highlighted.

A Global Human Settlement Layer From Optical HR/VHR RS Data: Concept and First Results

The simple, elegant approach of training convolutional neural networks (CNNs) directly from RGB pixels has enjoyed overwhelming empirical success. But can more performance be squeezed out of networks by using different input representations? In this paper we propose and explore a simple idea: train CNNs directly on the blockwise discrete cosine transform (DCT) coefficients computed and available in the middle of the JPEG codec. Intuitively, when processing JPEG images using CNNs, it seems unnecessary to decompress a blockwise frequency representation to an expanded pixel representation, shuffle it from CPU to GPU, and then process it with a CNN that will learn something similar to a transform back to frequency representation in its first layers. Why not skip both steps and feed the frequency domain into the network directly? In this paper we modify \libjpeg to produce DCT coefficients directly, modify a ResNet-50 network to accommodate the differently sized and strided input, and evaluate performance on ImageNet. We find networks that are both faster and more accurate, as well as networks with about the same accuracy but 1.77x faster than ResNet-50.

/pdf/faster-neural-networks-straight-from-jpeg-13iyssh4br.pdf

Faster Neural Networks Straight from JPEG

Satellite imagery is a valuable source of information for assessing damages in distressed areas undergoing a calamity, such as an earthquake or an armed conflict. However, the sheer amount of data required to be inspected for this assessment makes it impractical to do it manually. To address this problem, we present a semi-supervised learning framework for large-scale damage detection in satellite imagery. We present a comparative evaluation of our framework using over 88 million images collected from 4, 665 KM2 from 12 different locations around the world. To enable accurate and efficient damage detection, we introduce a novel use of hierarchical shape features in the bags-of-visual words setting. We analyze how practical factors such as sun, sensor-resolution, satellite-angle, and registration differences impact the effectiveness our proposed representation, and compare it to five alternative features in multiple learning settings. Finally, we demonstrate through a user-study that our semi-supervised framework results in a ten-fold reduction in human annotation time at a minimal loss in detection accuracy compared to manual inspection

/pdf/large-scale-damage-detection-using-satellite-imagery-l40jsy78yk.pdf

Large-scale damage detection using satellite imagery

A new unsupervised change detection method for modeling nonlinear temporal dependences based on local information is proposed. A theoretical analysis is presented, demonstrating how to derive optimal parameters for automating the method. It is then validated on both simulated data and very high resolution remote sensing imagery. The results show a clear improvement in change detection using the proposed method compared to other state-of-the-art change detection techniques.

Change Detection Based on Information Measure

With the increased spatial resolution of current sensor constellations, more details are captured about our changing planet, enabling the recognition of a greater range of land use/land cover classes. While pixeland object-based classification approaches are widely used for extracting information from imagery, recent studies have shown the importance of spatial contexts for discriminating more specific and challenging classes. This paper proposes a new compact representation for the fast query/classification of compound structures from very high resolution optical remote sensing imagery. This bag-of-features representation relies on the multiscale segmentation of the input image and the quantization of image structures pooled into visual word distributions for the characterization of compound structures. A compressed form of the visual word distributions is described, allowing adaptive and fast queries/classification of image patterns. The proposed representation and the query methodology are evaluated for the classification of the UC Merced 21-class data set, for the detection of informal settlements and for the discrimination of challenging agricultural classes. The results show that the proposed representation competes with state-of-the-art techniques. In addition, the complexity analysis demonstrates that the representation requires about 5% of the image storage space while allowing us to perform queries at a speed down to 1 s/ 1000 km 2 /CPU for 2-m multispectral data.

Lionel Gueguen

Papers

A Global Human Settlement Layer From Optical HR/VHR RS Data: Concept and First Results

Faster Neural Networks Straight from JPEG

Large-scale damage detection using satellite imagery

Change Detection Based on Information Measure

Classifying Compound Structures in Satellite Images: A Compressed Representation for Fast Queries