Introduction
Evaluating imagery for the use of specific analyses and purposes is a key skill in remote sensing. This portfolio entry evaluates NICFI imagery to provide insight and input that may be of use to the World Resource Institute’s (WRI) restoration efforts. Dr. Arakwiye from WRI has provided us with study locations (of which I have chosen one based in Rwanda) with which she is working on developing baseline tree cover maps with intention of being consistently updated. She also hopes to write about factors governing tree cover trends using this imagery. In this entry, I evaluate NICFI imagery at the study location in Rwanda and compare it to similar Landsat and Sentinel program imagery by assessing key remote sensing domains. I then provide recommendations on the use of NICFI imagery for Dr. Arakwiye’s purposes.
Background on NICFI Imagery
Norway’s International Climate and Forest Initiative (NICFI) is an international development fund that pledges to improve the health of tropical forests around the world as well as the livelihoods of those who live near and depend on these forests. Through this program, one can access high-resolution, analysis-ready imagery of the world’s tropical forests. As the goal is to reduce and reverse tropical forest loss, this dataset is largely geared towards identifying forest cover.
NICFI imagery for Rwanda is accessible through Google Earth Engine (GEE) under the dataset “Planet & NICFI Basemaps for Tropical Forest Monitoring – Tropical Africa”. The datasets are also available through other Planet GIS integrations in QGIS or ArcGIS Pro through which basemaps or subsets can be downloaded.
Comparisons are drawn between the above NICFI imagery, the USGS Landsat 8 Level 2, Collection 2, Tier 1 imagery, and the Sentinel-2 MSI: MultiSpectral Instrument, Level-2A. All of these datasets are processed, Surface Reflectance imagery.
Pre-Processing of Imagery
The data that is available on GEE is the “PlanetScope Surface Reflectance Mosaics (Analysis Ready).” This dataset is tailored for the use of scientific or quantitative monitoring and interpreting of satellite imagery. The pre-processing steps done with the NICFI imagery are much more extensive than those done with the Landsat or Sentinel imagery. While Landsat and Sentinel imagery provide tools with which one can independently remove cloud cover, the NCFI imagery performs this removal and masking of both clouds and cloud shadows before delivering the imagery. This higher quality preprocessed dataset would, therefore, be advantageous to work with when creating tree cover basemaps.
Temporal Resolution of Imagery
The NICFI dataset includes “Archive” or “Historical Basemaps” that have been collected biannually (one basemap every six months) from December 2015 to August 2020. From September 2020 onwards, imagery has been recorded on a monthly basis, termed as “Monitoring Basemaps.” The program is expected to run till August 2022 with the possible addition of two years.
The NICFI imagery has lower temporal depth, which would make it difficult to identify long-term forest change trends. Given the differences in collection frequency from 2015 to 2020, it may also be difficult to identify or determine seasonal trends in forest cover. With the possibility of the program itself ending very soon, it is unlikely that tree cover basemaps will be able to be updated using the same imagery. In comparison, the Landsat and Sentinel datasets have begun in 2013 and 2017 respectively; however, these datasets can still pull from other satellite missions in their programs whose data remain relatively consistent. These datasets are also updated much more frequently, Landsat being updated daily and Sentinel being updated every 10 days, allowing for more in-depth temporal analyses.
For example, seasonal changes in forest cover could be identified with Landsat and Sentinel imagery that pull from a larger dataset over the years as the images taken are more frequent. However, in NICFI, due to the lower frequency of collection, the same analysis may be full of cloud masks that would not be able to be filled in with other data when selecting a range of time periods over the years.
Spatial and Spectral Resolution
The spatial resolution of NICFI imagery is 4.77 meters, which is greater than Landsat (medium-resolution at 30 m) or Sentinel (high-resolution, multispectral at 10 m) imagery. However, the spectral resolution of NICFI imagery is lower, containing only four bands: blue, green, red, and infrared. Landsat (7 bands) and Sentinel (13 bands) imagery contains a much greater range in bands, allowing for the possibility of more indices to be calculated, such as the bare soil index (BSI).
Although these other bands may be helpful in analyses depending on what one wants to investigate, the presence of the NIR band may be sufficient to create baseline tree cove maps and identify and highlight trends governing tree cover. The spatial resolution of NICFI imagery may be more advantageous in more accurately identifying differences in landcover; however, this may also be a tradeoff with spectral resolution depending on what one is hoping to achieve with the analysis.
Radiometric Characteristics
Both the NICFI and Landsat imagery have a radiometric resolution of 16 bit while Sentinel has 12 bit. Coupled with the high spatial resolution, the NICFI imagery would be useful for identifying trends in the data through finer detection of the reflection of different features.
Visual Comparison of Imagery
The differences in spatial resolution are evident in Figure 1. A higher resolution would allow for more detailed tree cover basemaps to more clearly distinguish between trees and fields. This would prove more difficult with the Landsat imagery. A higher resolution would also allow for identifying trends governing tree cover, such as types of agricultural activities. The Sentinel imagery may also allow for this but it has a lower radiometric depth than NICFI, which would be less advantageous in distinguishing features.
Recommendations
Overall, the NICFI imagery has its benefits and drawbacks. Its extensive pre-processing, high resolution, and high radiometric depth would allow for the creation of detailed tree cover basemaps. However, its limited temporal depth may be disadvantageous in identifying yearly trends and updating basemap data. Its limited spectral resolution may also limit the identification of these trends. However, one would need to evaluate the tradeoff depending on the analysis one wishes to conduct.
Classification and Training Points
When classifying land cover in Rwanda, here is the link to the training points I had used to identify different features: https://code.earthengine.google.com/?asset=users/geog310/trainingpts2_P2. Below is also the class schema that classifies and distinguishes these different features.
Table 1. Classification schema
| Class | Screenshot example | Define the class | Why are you including this class? |
| Forest |  | Forests in this classification would be patches (small or extensive) of tree cover. They are distinguished by their typically darker green shades. | This class is the desired category for classification. |
| Dry Fields |  | Dry fields include grassland and tilled agricultural fields. They are distinguished by the brown and yellow shades of dried or sparse vegetation. | The landscape has a mix of different kinds of fields/plains. Dry fields have a different spectral profile from green fields and would need a separate classification. |
| Green Fields |  | Green fields include grassland and tilled agricultural fields. They are distinguished by green vegetation that includes minimal tree cover. | The landscape has a mix of different kinds of fields/plains. Green fields have a different spectral profile from dry fields and forests (they do not have a similar dark green colour or pattern) and would need a separate classification. |
| Water |  | Water is distinguished by its characteristic dark blue shades. In this classification, it refers mostly to large bodies of lakes. | Water bodies, especially lakes in this classification are distinctly different from other features and therefore, would need their own class. |
| Urban Areas |  | Urban areas are dense settlement areas as opposed to more scattered rural settlement (for the purposes of this analysis). It includes building, roads, and other man-made infrastructure. | This class has a distinctly different spectral profile from other features and should not be confused with dry fields and should be distinguished from forests. Therefore, it requires its own classification. |