Methodology:

Our methodology is based on: Xu, R., Lin, H., Lü, Y., Luo, Y., Ren, Y., Comber, A., 2018. A Modified Change Vector Approach for Quantifying Land Cover Change. Remote Sens. 10, 1578. https://doi.org/10.3390/rs10101578.

We acquired images from 2001-2003 from Landsat 7 ETM+ (legacy image) and from 2016-2020 from Landsat 8 OLI (contemporary image) to create a pair of image mosaics for each site.

Suitable images had low cloud cover and limited snow cover.

We selected six bands representing the visual and infrared wavelengths from the images for classification and topographically corrected the images.

We added three further bands to the image in the form of normalised difference snow index (NDSI), normalised difference vegetation index (NDVI), and normalised difference water index (NDWI).

We then merged each image pair to create single 18-band imaged, and added 3 further bands to describe the Euclidean distance, change vector angle and spectral angle mapper of each pair. This resulted in a 21-band image.

Using a modern classification of Byers' Peninsula (available from: Stringer, C. (2022). Contemporary (2016 - 2020) land cover classification across West Antarctica and the McMurdo Dry Valleys (Version 1.0) [Data set]. NERC EDS UK Polar Data Centre. https://doi.org/10.5285/5A5EE38C-E296-48A2-85D2-E29DB66E5E24 ) and a legacy classification of Byers' Peninsula produced using the same methodology as Stringer (2022).

We identified a series of class to class (ClTCl) changes using these classified images. We removed any ClTCl changes that represented less than 1 % of the points to reduce the risk of misclassification. The remaining points described 12 ClTCl change classes: wet ice to coarse/wet sediment (WITC), ice to fine & dry sediment (ITF), ice to coarse/wet sediment (ITC), ice to turbid water (ITT), coarse/wet sediment to turbid water (CTT), coarse/wet sediment to wet ice (CTWI), fine & dry sediment to bedrock (FTB), coarse/wet sediment to bedrock (CTB), coarse/wet sediment to fine & dry sediment (CTF), coarse/wet sediment to vegetation (CTV), bedrock to coarse sediment (BTC), fine and dry sediment to coarse/wet sediment (FTC).

Using 8,500 randomly selected points, we identified the spectral properties of each class to class change using the 21-band image. We then used a random forest classifier (500 trees) to classify the 21-band image at each site. We modified the training dataset for each of our five-proglacial sites to ensure that only changes between classes present in the modern land classification were possible.

Class to class changes should be considered in the context of the process they represent (e.g. land to water, ice to land, etc.) and not solely the ClCl value assigned to them.

The legacy land classification is described by band values of:

0 = No data

1 = Water

2 = Turbid water

3 = Wet ice

4 = Ice

5 = Land (non-differentiated)

6 = Bedrock

7 = Coarse/wet sediment

8 = Fine & dry sediment

9 = Vegetation

The land change classification is described by band values of:

0 = No change

12 = Wet ice to coarse/wet sediment (WITC)

16 = Ice to turbid water (ITT)

18 = Ice to coarse/wet sediment (ITC)

19 = Ice to fine & dry sediment (ITF)

25 = Bedrock to coarse/wet sediment (BTC)

29 = Coarse/wet sediment to turbid water (CTT)

30 = Coarse/wet sediment to wet ice (CTWI)

32 = Coarse/wet sediment to bedrock (CTB)

33 = Coarse/wet sediment to fine & dry sediment (CTF)

34 = Coarse/wet sediment to vegetation (CTV)

39 = Fine & dry sediment to bedrock (FTB)

40 = Fine & dry sediment to coarse/wet sediment (CTW)

Data quality:

First, we calculated the resubstitutuion error to assess the differences between the training data classes and predictions the classifier made. This shows the accuracy of the model based on the training data made available to it was 100.0 %. To validate the accuracy of our change maps, we reproduced the change detection analysis on Byers Peninsula with a 70/30 split of the training points between the classifier and validation. To ensure this split was unbiased, we randomly sorted the training points. This showed a total validation of accuracy of 80.1 %.