As seen in our results section, our current model produces results that fall into one of three categories: successful thresholding, failure due to missing data, or failure due to noise. Most of our data can be successfully thresholded with our current methods, however there are some processed NDWI images that were generated with missing data thus implying water where there should not be. Another failed scenario occurs when the images captured by LANDSAT are muddied by clouds or other noise that cause specific bands to be nearly unusable. We also found a general upward trend of water surface area until about 2019, where there was a sharp drop-off. However we do not have much confidence in the validity of these results as the amount of image data varies from year to year, which would ultimately affect our surface area estimates.
Our results are similar to other granular analysis of small bodies of water up until 2020, as an analysis of inland water in Sri Lanka [2] found a similar upwards trend in water surface area. However, our results require more image data to increase the confidence level of our predictions. Our approach is limited upon the availability of image data for Lake Oroville, and the time limitation for this quarter. Our approach can be improved upon by 1. obtaining more image data and 2. preparing the images through a pre-processing pipeline in order to mitigate noise and other issues as much as possible. More in depth analysis on the channels of water from Lake Oroville as well as the intra-annual changes and patterns would also be an interesting topic to further explore.