Data concerning land use, land cover, and changes in land over time.  Includes agriculture, forests, and restoration.

land

Land

The Dynamic World (Land Cover product) dataset shows globally consistent, high resolution, near real-time (NRT) land use land cover (LULC) classification leveraging deep learning on 10 meters Sentinel-2 imagery. The dataset was created utilizing a highly scalable cloud-based system to apply this approach and provide an open, continuous feed of  LULC predictions in parallel with Sentinel-2 acquisitions. The dataset includes the following Land Cover types: water, trees, grass, flooded vegetation, crops, shrub/scrub, built area, bare ground, and snow/ice.  You can learn more about the dataset [here](http://dynamicworld.app).  

Regularly updated global land use land cover (LULC) datasets provide the basis for understanding the status, trends, and pressures of human activity on carbon cycles, biodiversity, and other natural and anthropogenic processes. Annual maps of global LULC have been developed by many groups and while widely used, many important LULC change processes are difficult or impossible to observe at a spatial resolution greater than 100 meters and annual temporal resolution, such as emerging settlements and small-scale agriculture and early stages of deforestation and wetland/grassland conversion. Inability to resolve these processes introduces significant errors in our understanding of ecological dynamics and carbon budgets. Thus, there is a critical need for spatially explicit, moderate resolution (10-30 meters/pixel) LULC products that are updated with greater temporal frequency.  

The Dynamic World dataset is a first-of-its-kind NRT product, which accommodates a variety of user needs ranging from extremely up-to-date LULC data to custom global composites representing user-specified date ranges. Furthermore, the continuous nature of the product's outputs enables refinement, extension, and even redefinition of the LULC classification. In combination, these attributes enable unprecedented flexibility for a diverse community of users across a variety of disciplines. 
  
  
### Methodology  
  
The classification schema for Dynamic World was determined after a review of global LULC maps. The Dynamic World taxonomy maintains a close semblance to the land use classes presented in the [IPCC Good Practice Guidance](https://www.ipcc-nggip.iges.or.jp/public/gpglulucf/gpglulucf.html) to ensure easier application of the resulting data for estimating carbon stocks and greenhouse gas emissions.   

The modeling approach relies on semi-supervised deep learning and requires spatially dense annotations. To collect a diverse set of training and evaluation data, the authors divided the world into three regions, then further divided each region by the [14 RESOLVE Ecoregions biomes](https://academic.oup.com/bioscience/article/67/6/534/3102935). At each sample location, an initial selection of Sentinel-2 images from 2019 scenes based on image cloudiness metadata reported in the Sentinel-2 tile’s QA60 band was selected. The authors further filtered scenes to remove images with many masked pixels. Individual tiles of 510x510 pixels centered on the sample sites from random dates in 2019 were finally extracted.   

Further steps were then taken to obtain an “as balanced as possible” training dataset with respect to the LULC classifications from the respective LULC products. To achieve a large dataset of labeled Sentinel-2 scenes, the authors worked with two groups of annotators: photo-interpretation experts (validating 4,000 tiles) and non-experts (validating 20,000 tiles). A final validation set of 409 image tiles were held back from the modeling effort and used for evaluation as described in the Technical Validation section of the source methodology. To account for an annotation skill differential between the non-expert and expert groups, they one-hot encode the labeled pixels and smooth them according to the confidence in a binary label of the individual annotator (expert/non-expert).  

For training data collection, the authors used the Sentinel-2 Level-2A (L2A) product, which provides
radiometrically calibrated surface reflectance (SR). However, systematically produced Sentinel-2 SR products are currently only available from 2017 onwards. Therefore, for the modeling approach, the authors used the Level-1C (L1C) product, which was generated from 2015 and represents Top-of-Atmosphere (TOA) reflectance measurements. For any L2A image, there is a corresponding L1C image, allowing them to directly map annotations performed using L2A imagery to the L1C imagery used in model training.  

In addition to their preliminary cloud filtering in training image selection, the authors adopted and applied a novel masking solution that combines several existing products and techniques, including masking cloud shadows.   

The distribution of Sentinel-2 reflectance values is highly compressed towards the low end of the sensor range, with the remainder mostly occupied by high return phenomena like snow and ice, bare ground, and specular reflection. To combat this imbalance, the authors introduce a normalization scheme that better utilizes the useful range of Sentinel-2 reflectance values for each band.  
 
Their broad approach to transferring the supervised label data to a system that could be applied
globally was to train a Fully Convolutional Neural Network (FCNN). Conceptually, this approach transforms pre-processed Sentinel-2 optical bands into a discrete probability distribution of the classes in our taxonomy on the basis of spatial context. This is done per image on the hypothesis that sufficient spatial and spectral context is available to recover one of their taxonomic labels at a pixel. The authors
introduce a number of novel innovations that achieve state-of-the-art performance on LULC globally with a neural network architecture almost 100x smaller than architectures used for semantic segmentation or regression of ground-level camera imagery.  

To generate the NRT products, the authors apply the normalization described earlier to the raw Sentinel-2 L1C imagery and pass all normalized bands except B1, B8A, B9, and B10 after bilinear upscaling to ee.Model.predictImage. This output is then masked using the cloud mask derived from the unnormalized L1C image. The creation of these images is triggered automatically when new Sentinel-2 L1C and S2C images are available.The NRT collection is continuously updated with new results.  

For the full documentation, please click on the “Learn more” button.  
  
    
### Disclaimer   
   
Excerpts of this description page were taken from the source metadata.

Land & Carbon Lab

Dynamic World

Dynamic World: April 2022

Dynamic World: March 2022

Dynamic World: May 2022

Dynamic World: July 2022

Dynamic World: June 2022