Classification of Mosquito Habitat in Very High Resolution (VHR) Imagery | Elise’s Remote Sensing and Land Use Portfolio

The Lemon Fair Insect Control District has historically sprayed large swaths of land with insecticide to kill mosquito larvae. In addition to being a general annoyance, mosquitos can also carry disease that poses risk to humans and animals. However, the over-spraying of such chemicals can also have negative side effects, so if specific areas of mosquito habitat are identified, then targeted spraying can be used instead to both effectively kill mosquito larvae and minimize the extent of applied insecticide. This analysis aims to use land cover classification of remote sensed imagery to delineate such potential habitats within a study area on the Lemon Fair River.

I gathered imagery of the study area from the four sources in Table 1. All images were taken in April around the greatest flooded extent of the river. The drone imagery is high resolution but did not cover the entire study area, while the VT Orthoimagery and PlanetScope imagery had a lower resolution but complete spatial extent.

Image Source	Date	Resolution	Spatial Extent
VT Orthoimagery	April 2017	0.3m	Full study area
PlanetScope	April 2021	3m	Full study area
Drone	April 2021	A few cm	North stretch of river
Drone	April 2022	~1.5cm	Several acres

Table 1. Comparison of the 4 images used in the analysis.

In both the VT Ortho and PlanetScope imagery, the study area spanned multiple scenes so those scenes were mosaiced together. To classify the land cover type in the images, I completed an object-oriented analysis using the Classification Wizard on ArcGIS Pro. First, I created a schema of the land cover types to classify, listed in Figure 1. My class schema contained multiple classes instead of just the one mosquito habitat I wanted to identify (wetlands) with the goal of aiding the classifier by providing more examples to base the classification on. For each image, I collected 3-5 training polygons per class and then ran the classifier, which classified the rest of the image into the 6 classes in the schema.

^{Figure 1. The six classes in the schema used by the classification wizard. My target class to identify was wetlands.}

Overall, the classification did a poor job of identifying wetlands (and other land cover classes) in the images. While the classification of the PlanetScope image most accurately identifies wetlands, this source also provides the lowest resolution imagery and there are indiscernible features such that it is hard to validate what is truly wetland (see Figure 2).

^{Figure 2. PlanetScope image shown in a color infrared band combination to increase contrast and the land cover classification of the image.}

Figure 3 shows the classification of the VT Ortho image, which frequently misclassified cultivated fields as wetlands and did not classify actual wetlands as wetlands.

^{Figure 3. The classification of the VT Ortho image. The two “wetland” area circled in red are actually cultivated fields.}

The classification of the 2021 Drone image does a similar misclassification as the VT Ortho image of classifying cultivated fields as wetlands. Looking at a zoomed in section of the image in Figure 4 reveals that an area shadowed by trees was misclassified as wetland, perhaps due to this visual alternation. Another odd feature of the classification in this figure is the straight line as a boundary between wetlands and cultivated land in the classification when no such straight boundary exists in the image.

^{Figure 4. A color infrared rendering of the 2021 Drone image and the classification of this image.}

The classification of the 2022 drone image should not be used as data because the classification repeatedly failed to identify salient features of all land cover types, including repeatedly classifying the river as reed canary grass. Perhaps a different method of classification could yield better results (i.e. it was likely a user error).

My recommendation to the Lemon Fair Insect Control district would be to use the classification data from the PlanetScope image to identify mosquito habitat of interest but to use drone imagery as a reference for an actual image.