Mapping Land Cover by Stacking Landsat Imagery

Ad:

Regions where there is a lot of dust or cloud cover or potentially change rapidly can be difficult to monitor using satellite data. This can make forecasts, such as harvests or even if a given region is becoming more or less urbanized, difficult make. New methods now allow multi-temporal integration and stacking of images such that the composite views allow better estimates of land cover land use (LCLU) change to be made. This has benefits for policy and decision-making for governments and agencies.

Using Knowledge-based Expert Systems to Map Land Cover

For Sub-Saharan Africa, a region often having a lot of dust or obstruction to satellite data, the use of knowledge-based expert systems (KBES) has enabled multiple images to be integrated into a dynamic mapping tool to assess LCLU change. The KBES approach allows a rule-based algorithm, based on normalized difference vegetation index (NDVI) calculations, to learn from training pixels that identify land types. Multiple images over time are used to make a composite image of the region that experiences high dust or obstruction. Classification was then made using the composite imagery, where obstructed areas are masked and replaced with and land type once an image is available that clearly shows an area. In Sub-Saharan Africa, it was shown that 52.5% of Nigeria is arable land, which is higher than what has been reported previously (38.4%). Thus, it is argued that this form of classification is more accurate and accounts for dynamic change to LCLU, particularly in environments where observation is limited and change can be rapid.[1]

Example of Google Earth high resolution imagery (left side) and the corresponding 30 m resolution Landsat 8-based LCLU map for northern Nigeria (right side). Dark green: stable natural vegetation; Light green: non-stable natural vegetation; Light yellow: rain-fed agriculture; Red: Irrigation agriculture. Figure: Sedano, Molini, & Azad, 2019

Example of Google Earth high resolution imagery (left side) and the corresponding 30 m resolution Landsat 8-based LCLU map for northern Nigeria (right side). Dark green: stable natural vegetation; Light green: non-stable natural vegetation; Light yellow: rain-fed agriculture; Red: Irrigation agriculture. Figure: Sedano, Molini, & Azad, 2019

Similar approaches have been developed using time series data that apply machine learning techniques, specifically random forest methods, that apply a post-processing procedure to then determine LCLU types. This was applied on irrigated agricultural lands to determine areas irrigated and how that has changed from season to season. Similar to the previous study, the monitoring using Landsat 8 allowed for both stacking of images while also classifying areas as they change from scene to scene up to near real time.[2]

Block diagram showing the key steps of the methodology used in this study. For each Major Class (MC), Steps 1–3 were repeated to develop single class maps, followed by temporal aggregation. Figure: Pareeth et al., 2019

Block diagram showing the key steps of the methodology used in this study. For each Major Class (MC), Steps 1–3 were repeated to develop single class maps, followed by temporal aggregation. Figure: Pareeth et al., 2019

Another similar method to the Sedano et al. (2019) approach utilised NDVI to help classify an area obscured by cloud cover in China. In this approach, a normalized difference vegetation index coefficient of variation (NDVI-CV), rather than standard NDVI, was utilized, which accounts for phenological change to the vegetation as well as overall LCLU changes. However, this method attempted to minimize the use of images with heavy cloud cover. This was seen as a more accurate method to account for variation in LCLU in potentially obstructed regions because it can be a better account for seasonality in images when landscapes change naturally in addition to human-induced changes.[3]

Combining Imagery Types for Better Land Use Mapping

Other methods have attempted to combine imagery types, and not just multiple images from the same satellite system. Using Landsat 8 and RADARSAT-2 VH, a Canadian satellite, has enabled more accurate LCLU designation than only Landsat 8 by combining multiple seasons of Landsat 8 with that of RADARSAT-2 VH imagery. This helped to classify more accurately rangelands, forage areas, and croplands.[4]

Challenges to Stacking Landsat Imagery

Challenges remain using Landsat 8, including the fact that its resolution at 15 meters is not ideal for classifying very heterogenous landscapes. Nevertheless, training algorithms that apply random forest and eXtreme Gradient Boosting (XGBoost) have been integrated together to allow time series and varied, changing landscapes particularly along coastal regions that are heavily built-up, such as in Bangladesh, to be better mapped. The training algorithms can learn patterns of variation and measure the extremes of change, which helps to better classify if the change is a gradual more natural shift or one induced through LCLU change. In this case, it was observed that both agricultural land and built-up areas increased over 5% and 4% respectively since 2000.[5]

Landsat 8 is offered as free, open data, which has led to new innovation in methods to better utilize such data for time series and complex scenes that are sometimes obstructed by clouds or dust. While challenges still remain, such data have enabled more accurate maps and forecasts to be made that account for dynamic LCLU classification, often improving on government statistics and other measures used to guide estimates affecting policy or decision-making.

References

[1]    For more on the KBES approach using multiple Landsat 8 images, see: Sedano, F., Molini, V., & Azad, M. (2019). A Mapping Framework to Characterize Land Use in the Sudan-Sahel Region from Dense Stacks of Landsat Data. Remote Sensing, 11(6), 648. https://doi.org/10.3390/rs11060648.

[2]    For more on the random forest technique used to classify LCLU areas for irrigation, see: Pareeth, S., Karimi, P., Shafiei, M., & De Fraiture, C. (2019). Mapping Agricultural Landuse Patterns from Time Series of Landsat 8 Using Random Forest Based Hierarchial Approach. Remote Sensing, 11(5), 601. https://doi.org/10.3390/rs11050601.

[3]    For more on the NDVI-CV method, see:  Yang, Y., Wu, T., Wang, S., Li, J., & Muhanmmad, F. (2019). The NDVI-CV Method for Mapping Evergreen Trees in Complex Urban Areas Using Reconstructed Landsat 8 Time-Series Data. Forests, 10(2), 139. https://doi.org/10.3390/f10020139.

[4]    For more on the method that integrates Landsat 8 with RADARSAT-2 VH, see: Lindsay, E. J., King, D. J., Davidson, A. M., & Daneshfar, B. (2019). Canadian Prairie Rangeland and Seeded Forage Classification Using Multiseason Landsat 8 and Summer RADARSAT-2. Rangeland Ecology & Management, 72(1), 92–102. https://doi.org/10.1016/j.rama.2018.07.005.

[5]    For more on heterogeneousregions and how they can be mapped dynamically using Landsat 8 time series, see:  Abdullah, A. Y. M., Masrur, A., Adnan, M. S. G., Baky, M. A. A., Hassan, Q. K., & Dewan, A. (2019). Spatio-temporal Patterns of Land Use/Land Cover Change in the Heterogeneous Coastal Region of Bangladesh between 1990 and 2017. Remote Sensing, 11(7), 790. https://doi.org/10.3390/rs11070790.

See Also


Advertising



Like this article and want more?

Enter your email to receive the weekly GIS Lounge newsletter:

Advertising