Remote Sensing Intro (part 2)

Here's part 1

1. Spectral analysis

Here you can see the mapping results for 3 different analysis algorithms for the iron samples from my last post, which - remember - is a drop in the ocean of existing algorithms. The resources in the comments will lead you to different reviews of a large number of the different analyses types out there.

A comparison of mappings by the Spectral Angle Mapper (SAM), the ReSens+ algorithm (ReSens+) and a minimum wavelength mapping (MWL) is shown for the iron ore samples and for different stages of downsampling. It compares the accuracy of mapping to a validation image, for SAM and ReSens+ this accuracy of mapping is shown in [%], for MWL this is shown in relative comparison. 

SAM checks the similarity between the known (library) spectrum and the unknown pixel spectrum, by calculating the angle between them, treating them as vectors in a bandnumber-dimensional space. In case you don't have access to the commercial ENVI software license, SAM is also part of the open-source SPy python toolbox.

The ReSens+ algorithm is an enhanced version of the Bi-triangular feature fitting, that aims at multispectral data, triangulating within the spectrum, reconstructing and defining the spectral features by their triangle parameters. By comparing these parameters between the library and the unknown spectrum, the lib spectra best fitting the pixel spectrum can be identified and mapped. The method is not published in English yet but was used in numerous satellite jobs this year and is going to be published in a German Journal in the context of archaeological mapping for a church altar in Germany and for Sentinel-2 data exemplary for a few World Cultural Heritage Sites (Koellner et al., 2020, Mielke et al. 2020). Both SAM and ReSens+ can be applied to all kinds of data, hyper-, super-, and multispectral data alike.

A minimum wavelength mapping (e.g. doable with the HypPy python toolbox (Bakker, 2002; Bakker, 2014, vRuitenbeek, 2014; van der Meer, 2018; Hecker, 2019)) makes sense for distinct absorption features in order to show, where e.g. contaminants like clay dominate or where a distinct feature position change indicates a change in geochemistry that is relevant for the deposit (e.g. white mica AlOH feature changes for copper deposits Dalm et al., 2014 & 2017). Roughly, MWL checks for the position of the minimum (lowest point) of an absorption feature in a given range and calculates the feature depth of the feature. I have done this for the AlOH feature clearly visible in the 24% Fe sample (around 2200nm) which is only pronounced in a few of the other samples.

To distinguish the different ore grades it did not prove helpful, at least not in my tests during my PhD and for distinguishing these samples. MWL can only be feasibly applied to hyperspectral data which is able to distinguish these distinct features and changes within these features (change of only a few nm in wavelength position). In contrast to SAM and ReSens+, it does however not require a spectral library. Minerals can be found and mapped by their published absorption feature positions (e.g. Krupnik & Khan, 2019). As MWL is not my expertise I'm happy to have the people chime in that use it fairly regularly!

2: Application of a spectral library to satellite data

This is a region close to Tom Price in the Pilbara region, Western Australia, famous for being one of the hottest regions on earth and for being the centre of Australia's iron ore mining. Remember my post from Port Hedland a few weeks ago? This is where the ore is being mined from and it's being transported via rail towards the ports.

Ideally, a spectral geologist would have been on-site at least once, having a look at the conditions, the ore that is mined and desirable, talked to the geologist on-site and sampled together with them to ensure a complete spectral picture of the area.

Copernicus Sentinel-2 data from Nov. 2016

Copernicus Sentinel-2 data from Nov. 2020

I've checked the Copernicus Sentinel-2 archives for data from the area and found two lovely cloud-free images from Nov 2016 and Nov 2020. We should be able to see the development of the mines nicely in the 4-year time span. Have a look at the spectral library in S2 resolution, it's supposed to highlight how much information is lost from the spectra when using multispectral imagery. Ideally, I'd have used a ca. 100 band superspectral sensor or a hyperspectral one.

A fully functional site-specific spectral library would, of course, include fresh and weathered samples of ore-carriers and surrounding lithologies. 

Anyhow, before we look at the mapping results (in the next article), here's what I'd keep in mind when interpreting them:

1. We're looking at open-pits, which are large scale, but if I had the choice I'd use imagery of higher spatial resolution than 20x20m pixels. Each pixel will represent a variety of surface material, bunching together infrastructure, different stockpiles and pit levels in the various pixels. So the mapping is more of an approximation than an accurate mapping in this case. It's supposed to give us an idea of what we can see.
2. The spectral library is based on fresh samples from Brazil varying between 68-24% iron. The rock composition differs between the two areas, even though the ore grade mined in Brazil is close to the ore grades in this area (up to 62% Fe based on my online "research"). 
3. We're mapping relatively "fresh" surface in the mines after extraction and the spectral library from the samples will be a good approximation for that. But even if this material is considerably less weathered than the untouched surfaces around the open pits it will still be influenced by surface processes and secondary mineralization - depending on how long it's been "untouched". A mapping of the different Fe % is therefor an approximation and more of a long shot, showing us how the iron content is distributed relative to each other (think "more or less iron" instead of "exactly 64% iron will be mined here")
4. As we've just established that the lab built spectral library does not contain samples with a weathered surface, I've added spectra from the S2 imagery itself that describe the iron-rich, weathered surface cover. They are shown in brown-colours. A fully functional site-specific spectral library would, of course, include fresh and weathered samples of ore-carriers and surrounding lithologies. 
5. Ideally, a spectral geologist would have been on-site at least once, having a look at the conditions, the ore that is mined and desirable, talked to the geologist on-site and sampled together with them to ensure a complete spectral picture of the area. This includes asking about the weather conditions when the soil and pits carry more or less surface water etc. I simply checked on Wikipedia for the time frame with the lowest precipitation and chose imagery based on their availability and cloud cover in that time frame, which ended up being Oct-Nov. To make the mapping comparable I chose both datasets from these months. Also, as mentioned in the comments below the original post, dust cover plays a crucial role here and an on-site visit would have shown what a dominant role it plays in this particular area and mining environment.

Let's discuss the mapping results in the next article.

As the question came up in the comments of the original post as to why I used a spectral library from Brazil instead of one from the CSIRO or the USGS from the Pilbara region.

The simple answer is: "I did not have access to one"

The longer answer is: "I'm sure they already have spectral libraries from the area, as will Rio Tinto and all companies sampling and exploring in Pilbara. But as I do not know how the spectral data was acquired etc. I prefer to use a speclib that I've compiled myself ;) And I'm not sure how CSIRO does it, but for most speclibs e.g. the USGS spectral library, the spectra stem from spectrally pure minerals (often powders). I, however, am looking for mixed, surface materials that are neither mono-minerals nor powdered and the samples and spectra from Brazil were the best fit for that with my resources."

The resources

Part 1:

• on SAM: Kruse et al. 1993;
• Again, giving you the links for Philip Lypaczewski papers here: 

1. https://www.researchgate.net/publication/341620859_Characterization_of_Mineralogy_in_the_Highland_Valley_Porphyry_Cu_District_Using_Hyperspectral_Imaging_and_Potential_Applications
2. https://www.researchgate.net/publication/336855837_Hyperspectral_characterization_of_white_mica_and_biotite_mineral_chemistry_across_the_Canadian_Malartic_gold_deposit_Quebec_Canada

Reviews and research on different mineral and material mapping techniques for you to read:

• https://www.tandfonline.com/doi/full/10.1080/19479832.2019.1589585 (Girija & Mayappan, 2018)
• https://ieeexplore.ieee.org/abstract/document/8314827 (Khan et al., 2018)
• https://www.researchgate.net/publication/259096595_Geological_mapping_using_remote_sensing_data_A_comparison_of_five_machine_learning_algorithms_their_response_to_variations_in_the_spatial_distribution_of_training_data_and_the_use_of_explicit_spatial_ (Cracknell & Reading, 2014)
• https://doi.org/10.1016/j.jag.2015.12.004 (Asadzadeh & Filho, 2014)
• https://www.researchgate.net/publication/258134138_Evaluating_Classification_Techniques_for_Mapping_Vertical_Geology_Using_Field-Based_Hyperspectral_Sensors (Murphy, Monteiro, Schneider, 2012)
• Felix M. Riese gives a nice intro into the use of Machine Learning for hyperspectral in his PhD thesis, maybe he can also publish the publication link 
• Links to the HypPy toolbox literature can be found here: https://blog.utwente.nl/bakker/hyppy/

Part 2:

• I got my Sentinel-2 data from https://earthexplorer.usgs.gov/.

and a bit of info about mines in the area here:

1. https://miningdataonline.com/property/738/Nammuldi-Mine.aspx
2. and https://miningdataonline.com/property/804/Brockman-4.aspx

• The spectral library used is shown in hyperspectral here: https://www.dhirubhai.net/posts/friederikekoerting_iron-ore-remotesensing101-activity-6741643188831846400-d6RD