Creating my classification
| Class | Screenshot example | Class definition | Why include this class? |
|---|---|---|---|
| Forest | This class groups all dense forest in the area. There are clusters of forest throughout the area of interest and they show in darker green. | This is the primary class of interest for this assignment. This class is making sure that forests are not omitted from the classification. | |
| Green fields | There is a vast amount of these green fields that seem to be either grass fields or used for agriculture. They are mainly surrounding forested areas. | Having this class would allow for better distinction between the dark green reflectance of forests and the lighter green reflectance of fields. This would hypothetically allow the classifier to commit less errors of commission (classifying green fields as forest). | |
| Bare ground | This encompasses all bare ground fields (brown/ beige-looking). | Some patches of forest are next or very near bare fields, so having this class would allow the classifier to distinguish these two feature better reducing errors of commission as well. | |
| Water | This class represents all water bodies: mainly rivers and lakes. | I am including this class because there seem to be some eutrophic water bodies that the classifier may confuse as forest or fields. This would reduce those errors of commission. |
Interpretation
My classifier did a good job a representing forest areas. As you can see in Fig. 1. below, the red circle encompasses an intricate portion of forest that was well captured by the classifier in the bottom image. Nevertheless, clouds were classified as bare ground (beige), and the shadows of clouds as water(blue). However, that only occurred on the western section of the area of interest (AOI) where clouds were present (the rest of the AOI was virtually cloud free).
While inspecting my tree cover map and comparing it to Hansen’s tree cover map at different thresholds, I noticed that overall, my classifier tends to be more accurate on the location of the forest’s patches. Moreover, as seen in Table 1. below, levels of agreement varied between thresholds.
| 30-threshold:At this threshold, the Hansen tree cover map commits more errors of commission (yellow bar), but it also omits a significant area of forest (blue bar) | |
| 60-threshold: Here there is less errors of commission, although there are more omissions. This may be due to misclassification due to clouds. | |
| 80-threshold: The Hansen’s tree cover map omits a large part of the area classified by my forest cover map. |
Context: Suppose that your tree map is the truth against which Hansen’s thresholds are evaluated. At what threshold is the Hansen map truly reflecting tree cover on the ground? For how to report this section, use the style of “True Positive/True Negative/Balanced Rate” employed in Adjognon, et al’s paper for guidance.
| 30 -threshold | 50 -threshold | 80-threshold | 90-threshold | |
| TPR | 0.71 | 0.50 | 0.22 | 0.11 |
| TNR | 0.77 | 0.93 | 0.98 | 0.99 |
| Balanced | 0.74 | 0.71 | 0.60 | 0.55 |
As seen in the table above, the Hansen’s tree cover map was most in agreement with my tree cover map at thresholds around 50 percent. Even though the 30- percent threshold had a more reported a more balanced outcome, it was clear in my visual inspection that Hansen’s tree cover map was not nearly accurate at that threshold. I inspected both tree cover maps in cloud-free areas of my map, and notice that the there was most agreement between both maps on ranges between 55-65 percent threshold on Hansen’s tree cover map. Moreover, the table above may have skewed or not-too-reliable data because the heavy presence of clouds on the western side of the area probably account for a big area of forest that Hansen’s tree cover map is classifying as forest, but that my classifier is omitting. For that reason, I trusted more my observations by manually changing the thresholds and inspecting different areas of the map, than the True Positive/True Negative/Balanced Rate” table.