Home arrow Projects arrow Projects arrow Atacama Desert Project - Chile
Atacama Desert Project - Chile PDF Print E-mail

Remote Sensing Specialist assisting a project in Chile

 

Atacama Desert Project (ADP) of Forensic Architecture and Goldsmiths University requested the assistance of a remote sensing specialist. ADP asked the advice of the volunteer in the selection of appropriate satellite data-sets/imagery to analyze vegetation health and vegetation decrease in several areas along the Loa River in Chile. Jim Norton, a remote sensing specialist from Canada was recruited for this project and recently finalized the analysis and submitted the following report.

 

By: Jim Norton, GISCorps Volunteer & Godofredo Pereira of Forensic Architecture

 

Introduction

 

1. The following is a technical report in so much that the process and rational behind performing a NDVI analysis on the two locations in question will be detailed but the results will not be discussed as this is out of the scope of the project.

 

Project Scope

2. Jim Norton (GISCorps) will perform NDVI (Normalized Difference Vegetation Index) using any available Landsat sensor imagery for as long a time period as is possible of two locations in the Atacama region of Chile (Quillagua:  21°39'11.44"S  69°32'22.01"W and Chiu Chiu:  22°20'47.80"S  68°38'48.99"W) within 20 hours in June 2013 (a time frame designated by GISCorps briefing document found at Annex A (GISCorps Partner Agency Form) in order to aid in the identification of vegetation health and decrease. 

Processing workflow and associated software:

3. The following is the workflow followed through the project:
  • Gather data sets – USGS Earth Explorer
  • Layer Stack all 30m layers – Erdas Imagine
  • Atmospheric correction – PCI Geomatica
  • NDVI – Erdas Imagine
  • Data extraction – ArcGIS
  • Data presentation – Microsoft Excel

Data

 

4. Several sensors were considered for this project but the limitations of the length of study period and the spectral requirements of the NDVI constrained the choice of sensors to Landsat.

 

5. Landsat; was the only real choice of sensor for this project as it met several of the criteria for project success: 
a. Temporal availability – 1973 – present
 

USGS Landsat availability
   
    i. The Landsat system provides a 16 day repeat cycle.
   
    ii.Other than priot to 1984 and after 2012; Landsat 5 was used for sensor consistency.
 
b. Spectral and radiometric characteristics that allow the NDVI process.
   
    i. An 11 or 16 bit red band covering the spectral range of 0.45 - 0.52 µm.
   
    ii. An 11 or 16 bit A Near Infrared band covering the spectral range of 0.76 - 0.90 µm
 
c. Spatial characteristics. 30 meter cell size

d. Geometric accuracy. Landsat scenes are processed to Standard Terrain Correction (Level 1T -precision and terrain correction) if possible.

 

6. The Landsat data was available via the United States Geological Survey’s (USGS) Earth Explorer web interface (http://earthexplorer.usgs.gov/).

 

7. The path/row (Worldwide Reference System-2) used was predominantly 001/075 as according to the path/row coverage was the one scene that consistently covered both sites. This assumption was discovered to be flawed later in the processing when it was found that Quillagua was near to the edge or off some of the 001/075 scenes and therefore 115/169 was used to fill in the gaps in coverage from year to year.

 

 

Area of interest with Landsat path/row coverage.

 

8. The dates of imagery used for this project needed to be consistent throughout the period of study as too great a fluctuation of annual temporal vegetation variation would skew to results. The season of the year most applicable to this study is mid spring and in particular April when what rains there are in combination with snow melt will allow greater vegetative growth and as a consequence be able to easier to quantify. Every attempt was made to stay consistent with image scenes of the area of interest of April.

 

9. This was sometimes not possible because the images were not available. Even on a 16 day revisit cycle; the area of interest was not necessarily imaged and / or processed. On more than a few years; images from March or May had to be used.

 

10. Imagery from before 1984 was sparse as until Landsat 5 was launched; the onboard components did not allow for much storage or continuous transmission of data as we have enjoyed since 1984. This coupled with the lack of commercial interest in the area has led to a lack of data in this date range.

 

11. Landsat 7 data was used for 2012 as Landsat 5 data was no longer available and Landsat 8 was not available yet. This proved problematic as in 2003 the scan line corrector failed causing gaps in the imagery. The only possible way to use this imagery is to perform some processing to fill the gaps. One approach that had been used with some success was to take imagery of previous years to fill in the gaps in the imagery. This process was not suitable for this project so another approach that fills the gaps using interpolation of data one 30 meter pixel at a time creates a reasonable result. The process is called a focal mean and it takes a 3x3 sample around the missing data and fills in the pixel with the result. This process is repeated until the gap is filled. The process can be found here (http://landsat.usgs.gov/ERDAS_Approach.php) and in Annex B to this report.

 

12. The process worked well for Quillagua as seen below:

 

 

Landsat 7 of Quillagua with SLC data gaps and no focal mean.

 

 

2012 Landsat 7 (001/075) of Quillagua with SLC data gaps and 1 focal mean function applied.

 

 

2012 Landsat 7 (001/075) of Quillagua with SLC data gaps and 5 focal mean function applied.

 

13. But in Chui Chui; because of its proximity to the edge of the scene which is more affected by the data loss; this method produced less adequate results.

 

 

2012 Landsat 7 (001/075) of Chui Chui with SLC data gaps and no focal mean function applied.

 

 

2012 Landsat 7 (001/075) of Chui Chui with SLC data gaps and 6 focal mean functions applied.

 

14. The alternate and eventually the result that was used was derived from the adjacent path/row (115/169).

 

 

2012 Landsat 7 (233/076) of Chui Chui with SLC data gaps and no focal mean function applied.

 

 

2012 Landsat 7 (233/076) of Chui Chui with SLC data gaps and 5 focal mean function applied.

 

Atmospheric Correction

 

15. Every Image taken of an object not in direct contact with the sensor is prone to some atmospheric alteration of the digital numbers (DN). The further away the sensor is from the object; the more alteration of the DN from the actual radiance of the object. The purpose of atmospheric correction is to remove as much of the alteration from the actual radiance as possible. The atmosphere is complex and cannot be completely modeled other that by having in situ measurements available for the exact time the images were sensed. Theses in situ measurements are not available.

 

16. The following details the specific correction applied using PCI Geomaticas Atmospheric Correction function: http://www.pcigeomatics.com/pdf/Geomatica2013/TechSpecs/AtmosphericCorrection.pdf.

 

17. Further details of the atmospheric correction process taken can be found at Annex C to the report.

 

18. Because of its generalist nature and lack of ground truth; the results of the ATCOR are the most subjective of the processes.

 

 

Results of atmospheric correction (without on left)

 

NDVI

 

19. The Normalized Difference Vegetation Index is a well-documented method of deriving a vegetation index by using the red and infrared channels in a multispectral image to create the index. Once atmospherically corrected; the NDVI equation can be can be performed on the imagery. The results are between 0 and 1. Values of 1 indicate the maximum possible amount of vegetation (i.e. rainforest) and 0 indicates desert.

 

20. The following is a general overview of the NDVI:http://en.wikipedia.org/wiki/Normalized_Difference_Vegetation_Index.

 

21. The following is a practical guide to the NDVI process in Erdas Imagine: http://gis.stackexchange.com/questions/56944/how-to-create-an-ndvi-from-multispectral-imagery.

 

 

Chiu Chiu NDVI for 2001

 

Data Extraction

 

22. The NDVI raster gives a pixel by pixel account of the vegetative state of the imagery but to graph the results over the years; data needed to be extracted and put into graph format in Excel for easy and understandable analysis.

 

23. Using the NDVI results from 1984 (the first Landsat 5 scene); the vegetation indices with values of more than 0.1 were extracted and converted into an integer value of 1. Everything else was given a value of 0. This resulting raster was converted to a vector shapefile and everything but the two areas of interest were selected and deleted.

 

 

Chiu Chiu NDVI more than 0.1

 

 

Chiu Chiu 30m points

 

24. The values of each 30 meter pixel within the two boundaries were extracted into an Excel spreadsheet. This gave the basis for graphing to be made.

 

 

Process of extracting point values to spreadsheet.

 

Data presentation – Microsoft Excel

 

25. The data was recorded in such a way as to allow direct graphing of the results in Excel. Each recorded pixel value (some 7000 plus) for every year of the imagery available was assembled in a spread sheet and the following graphs were produced.

 

26. Average NDVI value for each site:

 

 

Chui Chui average annual NDVI

 

 

Quillagua average annual NDVI

 

27. The five sites with the highest NDVI value based on the 1984 scene were extracted and graphed

 

 

Chui Chui highest five NDVI sites from 1985

 

 

Chui Chui highest five NDVI sites from 1985 mapped

 

 

Quillagua highest five NDVI sites from 1985

 

 

Quillagua highest five NDVI sites from 1985 mapped

 

Conclusions

The process is a relatively straightforward one but takes some experimentation to ensure the appropriate corrections are applied to the relative sensors images.
 There is no guarantee that the results of the derived spreadsheets are representative of the vegetative states without ground truth and validation.

Recommendations

Acquire a new collect of Worldview 2 high resolution satellite imagery of each site to act a NDVI benchmark. Collect ground truth validation in situ using a spectrometer. Top of atmosphere (TOA) radiance has a linear relationship with ground reflectance.  By curve fitting TOA radiances against at least two known reflectance values, an empirical relationship between TOA radiance and ground reflectance can be found and applied to the image to provide highly accurate multispectral imagery and a more credible NDVI result.

This will allow some backwards calibration of the results of the Landsat analysis.

References: