Measuring forest loss with remote-sense data
We quantified changes in forest cover for each river catchment between the before and after by comparing year-matched (i.e., from 1 January to 31 December in the year fish specimens were collected) Landsat images on Google Earth Engine (GEE) (codes in Supplementary Material 1).
First, we downloaded all raw images available of our sites taken in the sampling year by USGS Landsat 5 (before time point) and USGS Landsat 8 (after time point). These were used to create cloud-free mosaics with the simple composite function on GEE. We delineated target river catchments using HydroSHEDS polygons (Lehner & Grill, 2013). In order to classify land-cover in the images with GEE’s Random Forest machine learning algorithm, we plotted ‘ground-truth’ polygons comprising pixels associated with forest and non-forest land-cover types. Our ‘ground-truth’ polygons were plotted manually over areas which were clearly associated with any of the above land-uses (e.g., residence complexes with urban land-use). We allocated 60% of our ‘ground-truth’ points to the training of the GEE Random Forest algorithm (or classifier) while the remaining data points were allocated for classifier testing and validation. We included the following spectral bands for land-cover classification: 1) RGB (red, green, and blue) bands; 2) near infrared; 3) shortwave infrared 1 & 2; and 4) thermal infrared 1 & 2. Generally, trained classifiers comprise models which have been parameterised with ‘ground-truth’ training data which can then be used to categorise all remaining pixels in relevant satellite images into one of the three land-cover types of interest. We tested the trained classifiers against our testing data (i.e., 40% of ‘ground-truth’ points) in a validation error matrix. Classification accuracy for all eight classifiers (two time points across four river catchments) were greater than the 95% threshold we set a priori . Finally, we used the trained Random Forest classifiers to produce land-cover raster layers of the catchments at both time points.
We quantified the following by comparing land-cover maps in thebefore and after time points:
  1. Proportion of total catchment area associated with forest loss (ΔFtotal ).
  2. Net proportion of catchment area associated with forest loss (ΔFnet ).
  3. Ratio of catchment area associated with forest loss versus forest gain (ΔFratio ).
In addition to catchment-wide changes, we were also interested in quantifying change in land-use cover in a sub-section of the overall catchment area, specifically, in the area upstream of our sampling points. To this end, we delineated upstream areas by processing void-filled NASA Shuttle Radar Topographic (SRTM) Digital Elevation Models (DEMs) (Jarvis et al., 2008) using watershed hydrological tools on Whitebox GAT (Lindsay, 2016). We then quantified the following:
  1. Proportion of catchment area associated with forest loss in the immediate upstream area of sampling points (ΔFsub.total ).
  2. Net proportion of catchment area associated with forest loss in the immediate upstream are of sampling points (ΔFsub.net ).
  3. Ratio of catchment area associated with forest loss versus forest gain in the immediate upstream area of sampling points (ΔFsub.ratio ).