For this final task, I will be sharing our group’s experience creating training points for land cover classification in our area of study in western Rwanda.
Our group made 8 different sets of training points to feed into the classifier. Each member made the decision on what classes to include in their set based on their observations and own reasoning.
Based on this information, I will be providing an evaluation of similarities and differences in our groups training points class schema decisions to conclude with a recommendation on what training points to include when making the land cover classification on this area.
The standard classification schema used to compare all individual classes was the following provided by Prof. Niwaeli
| Land Cover Class | Land cover sub-class | Example |
|---|---|---|
| Forestland | Natural |  |
| Plantation |  | |
| Cropland |  | |
| Grassland | Herbaceous |  |
| Shrubs |  | |
| Wetland | Vegetated |  |
| Water |  | |
| Built–Up |  | |
| Other | Bare-Soil |  |
Using this schema, I renamed all classes established in my peer’s classification so that I would be able to compare them. Then, I compiled them in a table of frequency which is visualized in Fig. 1. The main schemas used were forestsland_natural, bare soil, built-up, croplands and water.
Now it is important to evaluate whether all the classes shown above are relevant for the classification in the study region.
Looking closely at the study region there are land uses listed in the class schema that are not present within the study region and therefore are not a relevant source of differentiation for to the classifier.
Area of Interest in context of Rwanda 
landscape variability within the study region
There are no major bodies water or wetlands so these could be left out of the classification. Moreover, the most prominent features are natural forestland, herbaceous and shrub grasslands, built-up areas and bare-soil.
When it comes to the specifics of the land use in each of this categories, we had differences in our training. There was overlap between the bare soil and cropland classes. For example, as seen in the images below, for the Cropland class, some members of the group includes bare agricultural fields within the class, while the rest only included green agricultural fields.
green 
bare
This problem also occurred in the bare-soil class, where some members used exclusive barren areas, whereas others also included bare agricultural land.
Bare agricultural field 
Barren land
To resolve this issue, I would recommend following the these suggestions:
- Generate training points for these categories and clarification
- Cropland:
- Only include green agricultural fields.
- Forestland_natural:
- Include large forests but also be sure to include other discreet smaller forest patches as well. This will help the classifier include forests that are near villages, or around agricultural areas.
- Bare soil:
- Include barren and tilled or brown agricultural fields.
- Grassland shrubland:
- Even though there is not much of it, it could get mixed up with the natural forest classification.
- Grassland herbaceous:
- Include smooth looking green fields
- Built-up:
- Be sure to include smaller villages that are near forested areas, not only big urban areas.
- Forestland_plantation:
- This one is at your discretion, and you may include it if you would prefer the final tree cover map to not include these areas.
- Cropland:
- Ignore the following classes:
- Wetland_water: there is no significant body of water in the study region that could be confused as forest land
- Wetland_vegetated: likewise this land use is not present in the study region
Finally, because the classification takes as input an image larger than the study region, a final observation would be to be sure to include training points mainly from within the study region, but also allocating a portion to areas outside of it so the classifier has more diversity within each class.
If you would like to inspect all of our training points feel free to do so with the links below: