Introduction
This analysis is being conducted around agricultural fields between the towns of Bridport and Middlebury, in the Lemon Fair Valley towards the south of Snake Mountain. The land cover in the study area is composed of mostly agricultural lands, some forests, and fields of Reed Canary Grass (RCG), a perennial grass that grows around wet, agricultural fields and forms thick mats of vegetation.
The purpose of this analysis is to use Very High Resolution (VHR) imagery to identify areas of one land cover type that is potential mosquito habitat, something of interest to the Lemon Fair Insect Control District. The spatial detail of VHR imagery allows us to pin-point individual shrubs and grasses that could be mosquito breeding ground. This analysis compares different VHR imagery to identify RCG in the Lemon Fair Valley landscape. This grass restricts waterways and irrigation canals, forming dense vegetations and excluding native plants and animals. As read from data gathered by the Lemon Fair Insect Control District, many residents have identified this grass to host mosquito breeding and larvae growth.
Imagery Used and Comparison of Datasets
These are the different types of VHR imagery that were used in this analysis:
- Drone Imagery (April 2021) – Collected by Bill Hegman
- Drone Imagery (April 2022) – Collected by Bill Hegman
- Imagery from Vermont Open Data Portal (Vermont Orthoimagery Finder, 2018) from NAIP
- Imagery from Planet Labs (April 2021) from PlanetScope
Images from the Vermont data portal and Planet Labs were downloaded as scenes or tiles for a certain year or time and then clipped to our specific study area. Only one scene was used from each source and therefore, merging scenes together was not required.
Table 1 describes a comparison of these datasets based on criteria that would identify them to be preferable at mapping RCG.
Table 1. Comparisons of VHR imagery types
| Criteria | Drone 2021 | Drone 2022 | VT Open Data | Planet Labs |
| Spatial Resolution | 0.0937m / pixel High resolution allows for a more detailed classification. However, this could also result in overestimating RCG features. The processing time for this data would also be high. | 0.0241m / pixel High resolution allows for a more detailed classification. However, this could also result in overestimating RCG features. The processing time for this data would be even higher. | 0.6m / pixel This resolution would probably allow for sufficient detail in classification as well as lower processing than the drone data. | 3m / pixel This lower resolution would allow for much quicker processing times but more sweeping generalizations of where RCG is located. |
| Image Acquisition & Spatial extent | This spatial resolution allows for larger acquisition areas; however, the acquisition of this imagery would take a lot of time. | This spatial resolution would only allow for small extents of landcover and the acquisition of this imagery would take a lot of time. | The acquisition of this imagery from the VT OpenGeo Data portal very much depends on the kind of data available. Image can be search by years and not by specific months. Certain years may also not cover every area. The imagery acquired for this analysis was from NAIP, which acquires data when vegetation is best grown. Generally, it would be difficult to control the data you want to acquire. | Although more consistent in acquiring images or scenes of regions on a regular basis, higher resolution imagery is often not available for certain regions under this dataset. |
| Spectral Resolution | Bands 1 – 12 The wide range of bands allows for the classification to take advantage of spectral characteristics. However, all these bands may not be necessary to identify RCG. | Red, Green, Blue, NIR The NIR band provides another band to distinguish vegetation types from one another, including RCG. | Red, Green, Blue, NIR The NIR band provides another band to distinguish vegetation types from one another, including RCG. | Red, Green, Blue, NIR Although the imagery has a sufficient range of bands with which to visualize RCG, the resolution groups spectral profiles which makes for the larger generalizations in classification. |
Pre-processing and Analysis
This analysis was done using ArcGIS Pro software. All four images were manipulated to highlight RCG features using stretching of available bands, using the ‘Percent Clip’ stretch type. Below are the symbology parameters used for each image.
Table 2. Bands and stretch factors used for visualization
| Image | Red | Green | Blue |
| Drone 2021 | B5 0.12 – 0.2 | B4 0.06 – 0.17 | B3 0.08 – 0.22 |
| Drone 2022 | B1 151 – 178 | B2 120 – 190 | B3 125 – 220 |
| VT Open Data | B4 131 – 168 | B1 106 – 145 | B2 164 – 199 |
| Planet Labs | blue 157 – 972 | green 350 – 1447 | red 267 – 1927 |
Figure1. Visualized images before classification that highlight RCG. a) Drone 2021 B) Drone 2022 c) VtTile d) Planet
After setting the symbology, the images were run through the ‘Classification Wizard’ tool. The segmentation metrics for each image are in Table 3 below.
Table 3. Segmentation metrics for imagery used
| Image | Spectral Detail | Spatial Detail | Min. segment size |
| Drone 2021 | 10 | 12 | 10 |
| Drone 2022 | 9 | 8 | 10 |
| VT Open Data | 12 | 15 | 10 |
| Planet Labs | 15 | 15 | 10 |
Depending on the variety of features in each image, training samples for the classification were selected by the following classes listed in Table 4. Examples of these classes are provided in Figure 2.
Table 4. Classes used for collecting training areas to classify the images
| Class | Description |
| Reed Canary Grass | Grass that is distinctively thickly hatched and around agricultural fields and wet areas. This presence of this grass has great variability as it may be cleared from agricultural fields during certain years, making it a dynamic feature |
| Water | Still or flowing water parts of rivers or ponds |
| Forest | Areas of both conifer and deciduous forest cover including shadows |
| Barren | Areas of bare soil that could either be agricultural land or uncultivated open soil fields |
| Planted/Cultivated | Agricultural lands with some vegetation |
| Shrubland | Areas that are not cultivated or Reed Canary grass. Areas that include willows and cattails and other types of grasses |
After the classification process under the Classification Wizard tool, the resulting classified images were run through the ‘Reclassify’ tool that reclassed RCG to a value of 1 and all other classes combined to a value of 0.
Figure3. Classified images with purple areas identifying RCG. a) Drone 2021 B) Drone 2022 c) VtTile d) Planet
Interpretation and Evaluation
The drone imagery from 2021 accurately captures any existing RCG areas. However, it seems to overestimate certain areas, such as within agricultural fields or shrubland. The training samples used may have captured areas of RCG with similar spectral profiles. As the Classification Wizard classifies the image based on the provided visualization, certain yellow areas have been classified as RCG, as seen in Figure 4. However, these areas seem to be large swathes of land with a range of light to dark features (as brought out with symbology), as seen in Figure 4, that make this classification somewhat unreliable.
The drone imagery from 2022 appears to be more accurate than that from the previous year. This is perhaps because fewer degrees of higher resolution may result in more detailed classification and less spectral confusion among features. There was one portion of agricultural fields that has been mistakenly identified as RCG, as seen in Figure 5, but this is perhaps because a training sample was accidentally placed there. The higher resolution also provides greater difference in texture among the vegetation types that allows for better differentiation than in the Drone 2021 imagery. The classification captures a lot of detail; however, there are portions that are slightly over captured, such as very small areas within forests and shrubland.
The Vermont Tile data seems to also capture all areas accurately where RCG is present. There are much fewer places where over capturing seems to be an issue and the classification seems to capture enough detail necessary to identify general areas of RCG. as seen in Figure 6. Although the time of the year, based on the looks of the vegetation, was around the summer or early fall, the spectral properties of the RCG were able to be brought out through the symbology with the infrared band.
The Planet imagery seems to identify general areas of RCG, but these generalizations seem to be large. as seen in Figure 7. This seems to be a result of the lower resolution of imagery at 3 meters/pixel. To identify mosquito habitat, one would probably want more detailed a classification.
Recommendations
The best classifications seem to have come from the Drone 2022 imagery and the Vermont Tile imagery. Both of these data sets have their benefits and drawbacks. While the Drone 2022 imagery accurately captures areas of RCG, it does tend to over capture some areas. However, these areas a relatively small and can easily be distinguished as noise. The higher resolution of this imagery results in issues in acquisition as well as processing. Without the technology, it would be difficult to acquire this resolution of imagery. However, even so, it would cover a limited spatial extent as well as take a large amount of time to process.
On the other hand, the Vermont Tile imagery at 0.6 m/pixel also accurately captures areas of RCG. However, the availability of data is not as consistent as other sources and therefore, one would have to work with whatever is available on the portal. It does not allow for much spatial and temporal control in choice. Therefore, either of these two datasets would be preferable in identifying RCG, and therefore, mosquito breeding ground, depending on the resources one has and the spatial and temporal expectations.