Fully automated land cover maps have been increasingly developed, as satellite imagery reflectance data in various seasons have become increasingly known through processes of classification. With learning-based algorithms continually improving the understanding of signal data, classification and land cover generation has become far easier.
Land Cover Mapping at Country and Global Levels
Examples of work include Inglada et al.’s (2017) recent work that applies automated supervised classification, where country-wide level maps can be produced using 17 different identified land cover classes. Such data can be built up over time where long-term evaluation allows relatively accurate maps to be more easily produced as learning algorithms can better adapt to variations and ranges within classification-based data.[1]
While such work has developed country-level data, global-based information is now possible as well. With deforestation and land use change issues being of crucial importance, particularly as land use change occurs rapidly in forest and rainforest regions, the need for developing automated land cover maps has become more crucial.[2]
Most methods of classification are semi-supervised or supervised techniques, while variations have now also included spectral-slope-based methods that look at spectral range.[3]
Mapping Vegetation Seasons in Deserts
For desert regions, one challenge has been to map the short vegetation seasons they have. Traditionally, Normalized Difference Vegetation Index (NDVI) is a method used to quantify vegetation signatures in a region to monitor rainfall and vegetation growth.
Nonlinear fitting methods (e.g., polynomial fitting-based scheme) have been shown to be the best way to auto-estimate land cover change, using NDVI results, in desert regions, where factors of urban growth as well as other unexpected changes could affect readings.[4]
Mapping Water Surfaces
Identifying water surfaces is also important to distinguish these regions from other forms of land cover. Often, in Landsat imagery for instance, water appears as darker in the visible spectrum. Signatures could also be confused for other darker regions, making it possible that land cover may be confused or unclear for areas where water is present. Using known ranges of water within an index and continually learning and improving results based on false positive results, the Automated Water Extraction Index (AWEI) has been created to facilitate likely signatures of water. The index, relative to other methods such as Maximum Likelihood (ML), has been shown to perform better in capturing the variety and range in which water may appear on imagery data.[5]
The Need for Semi-automated Approaches for Land Cover Mapping
While methods for obtaining reflectance data, applying atmospheric corrections, and running classification have now become more automated, applying semi-automated approaches are still often necessary, particularly in higher resolution imagery (e.g., 30 meter resolution) such as Landsat 7 data, where error and variation from expected results becomes more likely. Techniques include random sampling and application of sampled data to not only test the accuracy of the classification but also improve results by informing on what the location is known to be. Even with this method, however, it can be fully automated once enough training data are applied, as classification can be improved over time through the development of better indices for specific land cover types.[6]
As knowledge about the Earth’s land cover has improved with the acquisition of higher resolution imagery and with various seasonal changes, better automated approaches are now being developed that employ classification techniques, often using machine learning methods such as random forest. The next challenge will be to create better regional and global scale maps using even higher resolution imagery (e.g., 1-meter resolution satellite data) where there is greater likelihood for error in automated methods. This will also require more scientists having access to high performance computing resources so various data could be more easily processed.
References
[1] For more information on Inglada’s et al.’s work, see: Inglada, J., Vincent, A., Arias, M., Tardy, B., et al. (2017) Operational High Resolution Land Cover Map Production at the Country Scale Using Satellite Image Time Series. Remote Sensing. [Online] 9 (1), 95.
[2] For a global-level automated map creation in relation to forested landscapes, see: Hansen, M.C., Potapov, P.V., Moore, R., Hancher, M., et al. (2013) High-Resolution Global Maps of 21st-Century Forest Cover Change. Science. [Online] 342 (6160), 850–853.
[3] For more on spectral-slope-based, see: Aswatha, S.M., Mukherjee, J., Biswas, P.K. & Aikat, S. (2017) Toward Automated Land Cover Classification in Landsat Images Using Spectral Slopes at Different Bands. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing. [Online] 10 (3), 1096–1104.
[4] For more on nonlinear methods for automated land cover classification in deserts using NDVI data, see: Jamali, S., Seaquist, J., Eklundh, L. & Ardö, J. (2014) Automated mapping of vegetation trends with polynomials using NDVI imagery over the Sahel. Remote Sensing of Environment. [Online] 141, 79–89.
[5] For more on automated water mapping techniques, see: Feyisa, G.L., Meilby, H., Fensholt, R. & Proud, S.R. (2014) Automated Water Extraction Index: A new technique for surface water mapping using Landsat imagery. Remote Sensing of Environment. [Online] 140, 23–35.
[6] For more on semi-automated methods for assisting classification and automated land cover maps, see: Mack, B., Leinenkugel, P., Kuenzer, C. & Dech, S. (2017) A semi-automated approach for the generation of a new land use and land cover product for Germany based on Landsat time-series and Lucas in-situ data. Remote Sensing Letters. [Online] 8 (3), 244–253.
