An optical image of Kliuchevskoi volcano on the left, with a radar image on the left (Image credit: Michigan Tech Volcanology, Image Cropped)
Improving the accuracy of land cover classification in cloud persistent areas using optical and radar satellite image time series (2020), Lopes et al., Methods in Ecology and Evolution, https://doi.org/10.1111/2041-210X.13359
Most ecologist has at some point run across or used a land cover map in their career. Whether it’s used for figuring out the canopy diversity of a forest, or figuring out which habitat a species is using, land cover maps are incredibly useful tools for everyone from conservationists to architects. But have you ever wondered how they are produced?
Until recently, land cover maps were created using either images from optical satellites or images from radar satellites with a coarse to medium spatial resolution (check out the Did You Know Section for more details, or the image above for an example). Combined with classification algorithms, land cover maps can be created automatically. That makes it sound simple, but the final output depends greatly on the quality and amount of images you use for the classification. Since 2014, the Copernicus Programme has made satellite imaginary freely available at high spatial and temporal spatial resolution. Due to this, optical and radar images can be combined more efficiently to produce land cover classification maps with enhanced accuracy. This is especially useful in tropical and boreal areas, as optical images often don’t show the entire landscape due to persistent cloud over.
Another important factor to consider when producing land cover maps is seasonality. Does the area you intend to classify change over the months? In tropical areas, the answer would probably be no, not too much. To test this hypothesis, the authors of this study investigated whether the use of a time series of radar and optical images improves the classification accuracy of a tropical peatland in the Jambi Province, Indonesia.
What They Did
To classify an area, you need to decide on your land cover categories. In this study, the authors used prior knowledge of the region acquired during field visits and high spatial resolution Google Earth images to select the categories and to indicate where in the study area these categories are found. Marked as polygons on the map, they were later used by the classification algorithm to classify the remaining pixels of the image. They chose to map the following categories: water, peat swamp forest, urban, palm trees, Acacia trees, fern/scrublands, bare ground, and mangrove.
As satellite data is freely available online, they downloaded all optical images (Sentinel-2) and radar images (Sentinel-1) acquired by satellites in 2017. After performing the necessary editing procedures, they ran the classification algorithm with six different combinations of input images: the annual temporal average (the mean pixel value of all images) or the complete time series of either optical or radar images and a combination of both. This greatly eases the computational power of the process but at the cost of possibly important information, such as the importance of seasonality. Comparing the accuracy of the resulting maps, the authors assessed the influence of seasonality on the classification results and whether using a combination of optical and radar images produces better maps than using just one of the two.
Did You Know: Optical vs. Radar Images
The two main types of satellite data used in remote sensing are optical and radar images. Optical satellite imagery is a great way to view the world as the human eye does. However, optical sensors measure reflected solar light, thus only function in the daytime and can’t penetrate through clouds. Radar sensors on the other hand, can image at both day and night and in almost all weather conditions. Moreover, radar images can reveal “secrets” of the land cover that are not visible to the human eye, such as soil moisture or inundated vegetation, marine pollution, or forest biomass.
What They Found
Surprisingly, the most accurate land cover classification maps were produced by using the time series of all images instead of a temporal average, indicating that seasonality does play a role. Certain categories benefited more from this approach, with palm trees, fern/scrublands, and mangrove mapped more accurately when multiple images were used. Overall, the combination of optical and radar images (and the use of time series) created even better maps, minimising classification errors. For example, radar data often confused palm trees and forest, while in optical data undetected clouds were classified as urban areas.
The authors measured the classification accuracy with the so called Kappa-coefficient. Their results showed that classification accuracy based on radar time series was higher than classifications based on optical time series. However, more accurate results don’t necessarily mean better maps. Thus, when analyzing your maps, potential errors due to the similarity of categories in optical or radar data need to be considered.
In the past, land cover maps often were based on classifying single optical images. However, accurately monitoring land cover is critical to ensure effective conservation and restoration action, and to inform ongoing policies and strategies. Freely available satellite data, combined with recent computational advances, significantly improve our ability to implement advanced and more correct classifications. This study clearly demonstrated that efforts to go beyond classical approaches (using several images of optical and radar satellites) do pay off by significantly improving produced land cover maps.
Julia Ramsauer is a landscape ecologist currently working on the integration of ecosystem services in the Mediterranean region. You can see here recent work on Ecology for the Masses on her profile at this link, or to listen to the latest episode of her podcast, Environmental Science Careers, you can follow her on Twitter here.