Data-fusion Sand-oriented Land Cover Classification: Modified Normalized Difference Sand Index (MNDSI)

A guest blog post by Niccolò Tolio and Andrea Semenzato.

Foreword by Sentinel Hub

This post is part of a series of guest blog posts written by authors talking about their entries to the Sentinel Hub Custom Script Contest. The author is one of the winners of our special edition of the Contest — Urban Growth in Africa — which ran from May to September 2021. The award for the best custom script went to the Regione del Veneto team — Niccolò Tolio, Andrea Semenzato, Umberto Trivelloni, Silvano De Zorzi, Daniele Piccolo, Mirko Frigerio and Alessandro La Rocca. Their winning Data-fusion Sand-oriented Land Cover Classification script is available on our Custom Script Repository.

Intense phenomena such as desertification and sandstorms constantly affect the African territories, slowing down the social and economic development of cities and communities [1]. Meanwhile, vacant sandy soil constitutes a large portion of land that could accommodate for new urban areas [2]. Therefore, sand detection can be crucial for a wide variety of applications, not only confined to the African continent, but also for other urban and natural environments. Such applications range from monitoring natural hazard, coastline erosion, and other issues related to climate change, to urban development and agricultural applications [3].

The study related to this post started in Dakar, whose region suffers from various development issues associated with environment deterioration, such as the decrease of green areas, farmlands and wetlands. Therefore, economic activities of primary sectors, such as agriculture and fishing, are damaged due to the lack of fertile land and water bodies. Consequently, sand detection and monitoring from satellite imagery data, within the Dakar’s arid environment, can have a significant reflection on the issues mentioned above. The research was then extended to other regions with different environments.

Figure 1. The image displays the evolution of Dakar’s urban development since the early XX century (see [2] for more details).

The study consists of a custom script based on a land-cover classification, from which soils and surfaces connected to sand materials are extracted. The classification relies on data fusion, consisting of joining different datasets within a single analysis; in this case, data are collected from Sentinel-1 (radar SAR) and Sentinel-2 (optical imagery). SAR data have shown to be particularly useful to extract buildings (e.g., see [3]), while optical products are exploited to classify most of natural features on the surface. In particular, spectral indices are used for land cover classification.

The script is structured as follows. Firstly, we compute VV and VH polarization values from Sentinel-1 SAR data to extract buildings and built-up areas, displayed in white. NDVI (Normalized Difference Vegetation Index) [4] is computed to identify two classes of vegetation, such as shrub and flourishing flora [5], which appear in yellow and green, respectively. A combination between MNDWI (Modified Normalized Difference Water Index) [6] from Sentinel-2 and VH polarization from Sentinel-1 is used to extract water bodies, classified in blue. BSI (Bare Soil Index) [7] applied to Sentinel-2 data classifies all bare soil and unvegetated land in red. Lastly, we propose a new index, named Modified Normalized Difference Sand Index (MNDSI), which operates in the visible wavelength on the Sentinel-2 sensor, considering the red and blue channel (see formula below), as opposed to NDSI (proposed by [8]), which takes into account the SWIR and the red bands from Landsat TM.

If analysed together with all the previous variables (following the specified order of declaration), we believe MNDSI is able to better identify reddish surfaces, such as sand (displayed in orange), which otherwise can be more likely misclassified as other unvegetated or bare soil, if considering the SWIR wavelength (NDSI). Unclassified pixels are in black.

The resulting ‘sand-oriented’ land cover classification offers a promising result for the MNDSI’s application: the surface of Dakar covered by sand seems to be thoroughly classified, especially in the coastal areas (Figure 2).

Figure 2. Dakar’s territory classified by the script. This classification highlights arid soils and sand in particular. Built-up areas appear in white, such as the highly urbanized area of Plateau in the southern part of the peninsula, and the dense neighborhoods in Pikine and Parcelles Assainies (white frame). Inland water bodies are well identified, as well, including the Lac Rose (blue frame), with a reddish color related to its peculiar composition. Frames within the main image are extracted from Maxar’s high-resolution imagery (World-View data), specifically available for the contest. All frames are in ‘true color’, except for the green one, which displays the Maxar NDVI related to the marshland in the central region of Dakar. (🌐 EO Browser)

The script has been tested in other areas, confirming its reliability in different environments. For example, the region of the Saloum delta, south of Dakar, is a marshy zone made of swamps, wetlands and small islands that are mostly covered by sand and shrubs. The script highlights sandy soils in orange (Figure 3).

Figure 3. Saloum Delta, south of Dakar. Comprised mostly of swamps, wetlands and small islands, this natural environment is widely covered by sand surfaces (especially in the northern sector), which are thoroughly detected by the script. (🌐 EO Browser)

Furthermore, another test in the Sahara Desert shows encouraging results, as well: sand is correctly classified in orange, while red stripes correspond to elongated hills that are not covered by sand, exposing bare soils and rock surfaces, due to the wind’s effect (Figure 4).

Figure 4. Portion of the Sahara Desert, in Southern Morocco. The large orange area detected by the script relates to the extensive sand dune fields of the desert. Moreover, red pixels correctly represent high crests and hills that are not covered by sand, therefore exposing bare soils and rock surfaces. These portions can change periodically due to the wind. (🌐 EO Browser)

Finally, the script has been tested in a completely different environment - the Venice lagoon. The script highlights coastal beaches facing the Adriatic Sea, with great reliability (Figure 5).

Figure 5. Detail of the Venice lagoon. The script highlights extensive beaches (made of sand) in the touristic areas of the coastline facing the Adriatic Sea. (🌐 EO Browser)

Nonetheless, the script still holds wide margins of improvements. Primarily, working in the visible wavelength, MNDSI does not take into account the infrared spectral composition of sand surfaces, which could improve their identification. However, with the spectral resolution available in the infrared wavelength range, these surfaces are not easily distinguishable from other bare soil surfaces. Moreover, the optimal threshold values used to distinguish bare soil (BSI) from actual sand (MNDSI) were set through empirical tests and thus may vary depending on the target area. Furthermore, some harvested farmlands might present similar visible and infrared signature to sand surfaces.

Still, the script displays encouraging results in various contexts, especially those related to arid environments (such as that of Dakar). Therefore, the growing availability of imagery data with improved spectral diversity and resolution (from visible to thermal infrared wavelengths) would definitely increase the script accuracy and its applicability to different environments.

We believe the script can be useful to a wide range of issues that affect environments similar to Dakar, providing a preliminary tool that can help tackling the challenges mentioned above, as well as answering the needs of the 2030 Agenda for the Sustainable Development [9]. Firstly, facing desertification (Goal 15 “Life on land”, target 15.3) and therefore contributing to better agricultural productivity in order to face hunger (Goal 2 “Zero hunger”, target 2.4). Moreover, it can provide a useful support for protecting water-related ecosystems (Goal 6 “Clean water and sanitation”, target 6.6) and natural environments from climate change’s impact (Goal 13 “Climate action”, target 13.1). Lastly, sand detection and sand monitoring can be crucial to foster a sustainable urbanization, for example helping the slum dwellers in arid environments (such as those of Dakar), where sandstorms mostly affect those who live in inadequate structures, or monitoring the healthiness of urban green spaces (Goal 11 “Sustainable cities and communities”, targets 11.3 and 11.7, respectively).

Custom Script Contest Team: Niccolò Tolio, Andrea Semenzato, Umberto Trivelloni, Silvano De Zorzi, Delio Brentan, Daniele Piccolo, Mirko Frigerio, Alessandro La Rocca.

Bibliography

1. Darkoh M. B. K., (1998), “The nature, causes and consequences of desertification in the drylands of Africa.”, in Land Degradation & Development, 9(1), pp. 1–20.

2. Japan International Cooperation Agency 2016, Project for Urban Master Plan of Dakar and Neighboring Area for 2035. (link)

3. Sivakumar, M. V. (2005). Impacts of natural disasters in agriculture, rangeland and forestry: an overview. Natural disasters and extreme events in Agriculture, 1–22.

4. Semenzato, A., Pappalardo, S. E., Codato, D., Trivelloni, U., De Zorzi, S., Ferrari, S., De Marchi, M. & Massironi, M. (2020). Mapping and Monitoring Urban Environment through Sentinel-1 SAR Data: A Case Study in the Veneto Region (Italy). ISPRS International Journal of Geo-Information, 9(6), 375.

5. Xue, J., & Su, B. (2017). “Significant remote sensing vegetation indices: A review of developments and applications.”, in Journal of sensors, 2017.

6. Xu, H. (2006). Modification of normalised difference water index (NDWI) to enhance open water features in remotely sensed imagery. International journal of remote sensing, 27(14), 3025–3033.

7. Fadhil A. M., (2013), “Sand dunes monitoring using remote sensing and GIS techniques for some sites in Iraq.”, in PIAGENG 2013: Intelligent information, control, and communication technology for agricultural engineering, 8762, p. 876206.

8. Nguyen C. T., Chidthaisong A., Kieu Diem P., & Huo L. Z., (2021), “A Modified Bare Soil Index to Identify Bare Land Features during Agricultural Fallow-Period in Southeast Asia Using Landsat 8”, in Land, 10(3), p. 231.

9. European Global Navigation Satellite System and Copernicus: Supporting the Sustainable Development Goals. Building Blocks towards the 2030 Agenda; United Nations Office at Vienna: Vienna, Austria, 2018 (link).

The Sentinel Hub team would like to thank Niccolò, Andrea and the Regione del Veneto team for their participation in the Sentinel Hub Custom Script Contest.

To learn more about satellite imagery and custom scripts, we recommend the Sentinel Hub Educational page and the Custom Scripts webinar. You can also visit a dedicated topic in the Sentinel Hub Forum for more information. We’d also like to invite you to take a look at the other entries submitted to the Sentinel Hub Custom Script Contests, which can be found here and here.

Stay tuned for our next Contest which will highlight a burning issue in the world today. We look forward to receiving your submissions.

If you want to learn more about Sentinel Hub, make sure to listen the MapScaping Podcast: