An Approach Based on Low Resolution Land-Cover-Maps and Domain Adaptation to Define Representative Training Sets at Large Scale

2021 ◽  
Author(s):  
Iwona Podsiadlo ◽  
Claudia Paris ◽  
Lorenzo Bruzzone
Keyword(s):  
Land Cover ◽  
Large Scale ◽  
Domain Adaptation ◽  
Low Resolution ◽  
Land Cover Maps ◽  
Training Sets

scholarly journals Domain-Adversarial Training of Self-Attention-Based Networks for Land Cover Classification Using Multi-Temporal Sentinel-2 Satellite Imagery

2021 ◽  
Vol 13 (13) ◽  
pp. 2564
Author(s):  
Mauro Martini ◽  
Vittorio Mazzia ◽  
Aleem Khaliq ◽  
Marcello Chiaberge
Keyword(s):  
Land Cover ◽  
Large Scale ◽  
Domain Adaptation ◽  
Learning Approaches ◽  
Atmospheric Conditions ◽  
Practical Applications ◽  
Area Of Interest ◽  
Multi Temporal ◽  
Significant Performance ◽  
Adversarial Training

The increasing availability of large-scale remote sensing labeled data has prompted researchers to develop increasingly precise and accurate data-driven models for land cover and crop classification (LC&CC). Moreover, with the introduction of self-attention and introspection mechanisms, deep learning approaches have shown promising results in processing long temporal sequences in the multi-spectral domain with a contained computational request. Nevertheless, most practical applications cannot rely on labeled data, and in the field, surveys are a time-consuming solution that pose strict limitations to the number of collected samples. Moreover, atmospheric conditions and specific geographical region characteristics constitute a relevant domain gap that does not allow direct applicability of a trained model on the available dataset to the area of interest. In this paper, we investigate adversarial training of deep neural networks to bridge the domain discrepancy between distinct geographical zones. In particular, we perform a thorough analysis of domain adaptation applied to challenging multi-spectral, multi-temporal data, accurately highlighting the advantages of adapting state-of-the-art self-attention-based models for LC&CC to different target zones where labeled data are not available. Extensive experimentation demonstrated significant performance and generalization gain in applying domain-adversarial training to source and target regions with marked dissimilarities between the distribution of extracted features.

scholarly journals Domain Adversarial Neural Networks for Large-Scale Land Cover Classification

2019 ◽  
Vol 11 (10) ◽  
pp. 1153 ◽  
Author(s):  
Mesay Belete Bejiga ◽  
Farid Melgani ◽  
Pietro Beraldini
Keyword(s):  
Neural Networks ◽  
Land Cover ◽  
Transfer Learning ◽  
Large Scale ◽  
Domain Adaptation ◽  
Land Cover Classification ◽  
Target Domain ◽  
Training Samples ◽  
Latent Space ◽  
Large Scale Land

Learning classification models require sufficiently labeled training samples, however, collecting labeled samples for every new problem is time-consuming and costly. An alternative approach is to transfer knowledge from one problem to another, which is called transfer learning. Domain adaptation (DA) is a type of transfer learning that aims to find a new latent space where the domain discrepancy between the source and the target domain is negligible. In this work, we propose an unsupervised DA technique called domain adversarial neural networks (DANNs), composed of a feature extractor, a class predictor, and domain classifier blocks, for large-scale land cover classification. Contrary to the traditional methods that perform representation and classifier learning in separate stages, DANNs combine them into a single stage, thereby learning a new representation of the input data that is both domain-invariant and discriminative. Once trained, the classifier of a DANN can be used to predict both source and target domain labels. Additionally, we also modify the domain classifier of a DANN to evaluate its suitability for multi-target domain adaptation problems. Experimental results obtained for both single and multiple target DA problems show that the proposed method provides a performance gain of up to 40%.

scholarly journals Automated Production of a Land Cover/Use Map of Europe Based on Sentinel-2 Imagery

2020 ◽  
Vol 12 (21) ◽  
pp. 3523
Author(s):  
Radek Malinowski ◽  
Stanisław Lewiński ◽  
Marcin Rybicki ◽  
Ewa Gromny ◽  
Małgorzata Jenerowicz ◽  
...  
Keyword(s):  
High Resolution ◽  
Land Cover ◽  
Data Processing ◽  
Spatial Resolution ◽  
Large Scale ◽  
Observation Data ◽  
European Continent ◽  
Continental Scale ◽  
Land Cover Maps ◽  
Sentinel 2

Up-to-date information about the Earth’s surface provided by land cover maps is essential for numerous environmental and land management applications. There is, therefore, a clear need for the continuous and reliable monitoring of land cover and land cover changes. The growing availability of high resolution, regularly collected remote sensing data can support the increasing number of applications that require high spatial resolution products that are frequently updated (e.g., annually). However, large-scale operational mapping requires a highly-automated data processing workflow, which is currently lacking. To address this issue, we developed a methodology for the automated classification of multi-temporal Sentinel-2 imagery. The method uses a random forest classifier and existing land cover/use databases as the source of training samples. In order to demonstrate its operability, the method was implemented on a large part of the European continent, with CORINE Land Cover and High-Resolution Layers as training datasets. A land cover/use map for the year 2017 was produced, composed of 13 classes. An accuracy assessment, based on nearly 52,000 samples, revealed high thematic overall accuracy (86.1%) on a continental scale, and average overall accuracy of 86.5% at country level. Only low-frequency classes obtained lower accuracies and we recommend that their mapping should be improved in the future. Additional modifications to the classification legend, notably the fusion of thematically and spectrally similar vegetation classes, increased overall accuracy to 89.0%, and resulted in ten, general classes. A crucial aspect of the presented approach is that it embraces all of the most important elements of Earth observation data processing, enabling accurate and detailed (10 m spatial resolution) mapping with no manual user involvement. The presented methodology demonstrates possibility for frequent and repetitive operational production of large-scale land cover maps.

scholarly journals Combining Multiple Geospatial Data for Estimating Aboveground Biomass in North Carolina Forests

2021 ◽  
Vol 13 (14) ◽  
pp. 2731
Author(s):  
Leila Hashemi-Beni ◽  
Lyubov A. Kurkalova ◽  
Timothy J. Mulrooney ◽  
Chinazor S. Azubike
Keyword(s):  
North Carolina ◽  
Land Cover ◽  
Large Scale ◽  
Forest Biomass ◽  
Estimation Procedure ◽  
Geospatial Data ◽  
Inventory Analysis ◽  
Natural Resource Conservation ◽  
Management And Development ◽  
Land Cover Maps

Mapping and quantifying forest inventories are critical for the management and development of forests for natural resource conservation and for the evaluation of the aboveground forest biomass (AGFB) technically available for bioenergy production. The AGFB estimation procedures that rely on traditional, spatially sparse field inventory samples constitute a problem for geographically diverse regions such as the state of North Carolina in the southeastern U.S. We propose an alternative AGFB estimation procedure that combines multiple geospatial data. The procedure uses land cover maps to allocate forested land areas to alternative forest types; uses the light detection and ranging (LiDAR) data to evaluate tree heights; calculates the area-total AGFB using region- and tree-type-specific functions that relate the tree heights to the AGFB. We demonstrate the procedure for a selected North Carolina region, a 2.3 km2 area randomly chosen in Duplin County. The tree diameter functions are statistically estimated based on the Forest Inventory Analysis (FIA) data, and two publicly available, open source land cover maps, Crop Data Layer (CDL) and National Land Cover Database (NLCD), are compared and contrasted as a source of information on the location and typology of forests in the study area. The assessment of the consistency of forestland mapping derived from the CDL and the NLCD data lets us estimate how the disagreement between the two alternative, widely used maps affects the AGFB estimation. The methodology and the results we present are expected to complement and inform large-scale assessments of woody biomass in the region.

scholarly journals INTER-COMPARISON OF THE GLOBAL LAND COVER MAPS IN AFRICA SUPPLEMENTED BY SPATIAL ASSOCIATION OF ERRORS

2019 ◽  
Vol XLII-4/W14 ◽  
pp. 11-16
Author(s):  
G. Bratic ◽  
D. Oxoli ◽  
M. A. Brovelli
Keyword(s):  
Land Cover ◽  
Large Scale ◽  
Confusion Matrix ◽  
Spatial Association ◽  
Probability Of Error ◽  
Earth Observations ◽  
Area Of Interest ◽  
Global Land Cover ◽  
Land Cover Maps ◽  
Single Pixel

<p><strong>Abstract.</strong> Recent advances in Earth Observations supported development of high-resolution land cover (LC) maps on a large-scale. This is an important step forward, especially for developing countries, which experienced problems in the past due to absence of reliable LC information. Nevertheless, increasing number of LC products is imposing additional validation workload to confirm their quality. In this paper inter-comparison of two recent LC products (GlobeLand30 and S2 prototype LC 20m map of Africa) for country of Rwanda in Africa was done. It is a way to facilitate validation by identifying the areas with higher probability of error. Specific approach of comparison of single pixel of one map with multiple pixels of another map provided confusion matrix and sub-pixel agreement table. In this work, accuracy indexes based on the confusion matrix were computed as a measure of similarity between the two maps. Furthermore, Moran’s I index was computed for estimation of spatial association of the pixels in disagreement. Also, total disagreement, as well as disagreement of particularly confused classes was visualised to analyse their spatial distribution. The results are showing that similarity of the two maps is about 66%. Disagreements are spatially associated and the most evident in the eastern and north-western part of the area of interest. This coincides also with the distribution of the two most confused classes Wetland and Shrubland. The results delineate areas of inconsistency between the two maps, and therefore areas where careful accuracy analysis are needed.</p>

scholarly journals A new European land systems representation accounting for landscape characteristics

2021 ◽  
Author(s):  
Yue Dou ◽  
Francesca Cosentino ◽  
Ziga Malek ◽  
Luigi Maiorano ◽  
Wilfried Thuiller ◽  
...  
Keyword(s):  
Land Use ◽  
Land Use Change ◽  
Land Cover ◽  
Large Scale ◽  
Bird Species ◽  
Data Availability ◽  
Landscape Characteristics ◽  
Continental Scale ◽  
Land Systems ◽  
Land Cover Maps

Abstract Context While land use change is the main driver of biodiversity loss, most biodiversity assessments either ignore it or use a simple land cover representation. Land cover representations lack the representation of land use and landscape characteristics relevant to biodiversity modeling. Objectives We developed a comprehensive and high-resolution representation of European land systems on a 1-km2 grid integrating important land use and landscape characteristics. Methods Combining the recent data on land cover and land use intensities, we applied an expert-based hierarchical classification approach and identified land systems that are common in Europe and meaningful for studying biodiversity. We tested the benefits of using this map as compared to land cover information to predict the distribution of bird species having different vulnerability to landscape and land use change. Results Next to landscapes dominated by one land cover, mosaic landscapes cover 14.5% of European terrestrial surface. When using the land system map, species distribution models demonstrate substantially higher predictive ability (up to 19% higher) as compared to models based on land cover maps. Our map consistently contributes more to the spatial distribution of the tested species than the use of land cover data (3.9 to 39.1% higher). Conclusions A land systems classification including essential aspects of landscape and land management into a consistent classification can improve upon traditional land cover maps in large-scale biodiversity assessment. The classification balances data availability at continental scale with vital information needs for various ecological studies.

Sign in / Sign up

Export Citation Format

Share Document