Portfolio 3: Landsat-8 vs. Sentinel-2 | Mapping of tree cover for landscape restoration

11/12/20

.

Time-Frame Rationale

I decided to use imagery from May through August because it was during the dry season, meaning there were fewer clouds, and when inspecting individual images from different months, I found that trees were less distinct between August and October. I hypothesize that this is because as dry season progressed (the dry season is May through October [source]), the trees became more water strained and therefore less green (Figure 1). 

Figure 1. Sentinel-2 imagery from 5/22/20 (left) and 9/11/19 (right). Forest is more distinct in May than September. While neither are great quality imagery, these were the best I could get when sifting through tiles.
Click image for code

Land Cover Classes

I actually did not adapt my classes because they were already very similar to the ones provided. I did not see any tree plantations, could not distinguish between herbaceous grassland and shrub grassland, did not see any wetlands, and the few built up sites would not be mistaken for forest. I maintained forestland (previously termed “trees”), cropland (previously termed “agriculture”, bare soil (previously termed “bare ground”), and grassland (previously termed “grass/shrub”). Looking back I could have included open water as a class, but because none of the validation points were in open water, this did not affect my confusion matrices.

Mapathon Dataset: advantages & disadvantages

Some advantages of the Mapathon dataset are that it allows for centralized, long-term monitoring of land cover in a relatively accessible, inclusive way. It involves locals who can validate land cover, increasing confidence in the data and a better sense of the accuracy of satellite imagery and classification algorithms.

A disadvantage is that the process of validating points is tedious, requiring manually looking at Google Earth Pro or having to physically be at each point. These limitations in scale are reflected in the fact that the validation points from Mapathon used in my analysis only covered a small part of my study area.

Accuracy Comparison between Landsat-8 and Sentinel-2

The overall accuracy of the Landsat-8 classification was 70%, versus 75% for Sentinel-2. The kappa coefficient for Landsat-8 was 0.35, versus 0.49 for Sentinel-2. For both Landsat-8 and Sentinel-2, approximately a third of the validation points where there were no trees were omitted. Where there were trees, a fifth were omitted in the Landsat-8 classification, versus no error of omission for validated trees in the Sentinel-2 classification. For Landsat-8, there was a very high error of commission (58%) for areas without trees; for Sentinel-2, it was 52%. Interestingly, the Sentinel-2 classification had no error of commission where classified as no trees, versus in Landsat-8 there was a 9% error (Tables 1-2).

Table 1: Sentinel-2 confusion matrix.
Table 2: Landsat-8 confusion matrix.
Figure 2. Sentinel gap-filled image vs. tree classification (trees = white, no trees = black)
Click image for code
Figure 3. Landsat gap-filled image vs. tree classification (trees = white, no trees = black)
Click image for code

Despite the seemingly low accuracy and relatively high commission and omission rates in the confusion matrix, when looking at the study area as a whole, the tree cover classifications look very accurate for both Landsat-8 and Sentinel-2 (Figures 2-3).

Although both may be true (low calculated accuracy and high visual accuracy), it is also possible that because the Mapathon imagery was from 2017 and the gap-filled image is from 2018-2020, the gap-filled images may have shown slightly different land cover, decreasing congruence between the validation points and classification. There is likely little difference between these two time periods, but it is still something to consider when assessing how accurate the validation points are.