Portfolio Problem III – Remote Sensing Portfolio

November 13, 2020

Mapping Tree Cover for Landscape Restoration

In this problem, we considered the challenges of and the role of image resolution in mapping tree cover restoration efforts. We worked with Landsat-8 imagery (resolution: 30 m), and Sentinel-2 imagery (resolution: 10 m) to create landcover classifications in our study region in Malawi, with the goal of isolating tree cover data for the year 2019.

Create Gap-Filled Imagery for Sentinel-2 and Landsat-8

The first task was to determine what time of year is best for tree cover mapping. For the Landsat imagery, enough data was available to choose to map the dry season (August-November, for years 2018-2020). This time of year was chosen for its lack of clouds, which made for higher-quality imagery, and for relatively easy distinction of tree cover (Fig. 1a).

^{Fig. 1a. Gap-filled image created from Landsat imagery for the dry season (05/01 – 11/30) in Malawi from the years 2018, 2019, and 2020. Link to Code}

However, Sentinel data for this area is only available for 2019 and 2020, and most date combinations left large gaps in the composite image. This meant that all available images form 2019 and 2020 were used to create a gap-filled image that was not riddled with holes (Fig. 1b).

^{Fig. 1b. Sentinel data: gap-filled imagery from the dry season (05/01/19 – 11/30/19 and 05/01/20 – 11/30/20) shown at left; gap-filled image using all images from 2019 and 2020 shown at right. Link to Code}

Classify Land Cover Types in Gap-Filled Images

The next task was to classify the gap-filled images based on landcover categories. We had previously created cover classes for Forest, Cropland, Grassland, Scrub, and Water. For the current question, we updated the training categories to distinguish between forest plantations and natural forests, as natural forests are a better indicator of restoration, and added a class for the built environment, where there was some confusion with tree cover. In addition, we added and shifted some points to include more examples of what to classify as “tree cover.”

Collect Validation Points and Create Confusion Matrices

Students working in Malawi then collaborated to create validation points, which indicated whether or not a training point was forested. These points were then compared to the forested pixels assigned by the classifier to the Landsat and Sentinel images (Fig. 2). This allowed “truthing” of the classifiers with relative confidence in the validation points and provides context for comparing classification of images from the two datasets. However, only 61 validation points were created for a very large study area, and the task of identifying landcover types was a somewhat subjective one, so our reliance on the validation in this case must be cautious at best.

^{Fig. 2. Validation points (black) compared to Landsat tree cover pixels (purple). Landsat image shown in background for context. Link to Code}

The results of the confusion matrices indicate that the tree cover classification for the Landsat data was more accurate than that of the Sentinel data (Table 1, Fig. 3).

	Overall Accuracy (%)	Kappa Coefficient
Landsat-8	78.3	0.41
Sentinel-2	53.3	0.15

^{Table 1. Accuracies for Landsat-8 and Sentinel-2 before and after adjusting for expected accuracy. The “No Tree” class seems to be overrepresented in our analysis, which likely explains our low Kappa coefficient.}

^{Fig. 3. Complete confusion matrices for Landsat-8 and Sentinel-2.}

The results of the confusion matrices indicate that the tree cover classification for the Landsat data was more accurate than that of the Sentinel data. As mentioned before, we must interpret the results of the confusion matrix with some caution. However, looking at the map visualizations, it does indeed seem that the Landsat classification was more successful in terms of picking out large forested areas. Upon closer inspection, it seems that some of the inaccuracy of the Sentinel classification may have to do with interference from clouds (Fig. 4a.).

^{Fig. 4a. Left: Landsat-8 imagery underneath tree cover layer (bright green). Right: Sentinel-2 image from the same area, underneath tree cover layer (bright green). The “fog” in the upper center-left region of the Sentinel image indicates that there was strong interference from clouds in the classification of this image. Link to Code}

For visual purposes, a Sentinel gap-filled image of the same extent as the Landsat image was used. In areas with low cloud interference, it is possible Sentinel data did capture several more dispersed areas of tree cover, likely due to its lower resolution (Fig. 4b). However, it may have been preferable to prioritize low-cloud imagery over extent of coverage in this analysis.