气候变化领域集成服务门户

Increased availability and resolution of remotely-sensed (RS) imagery of Earth's surface has greatly enhanced the precision, spatial extent, and temporal frequency at which we can analyze river channel migration and width changes. Despite a body of research identifying and quantifying sources of uncertainty inherent in such data, no framework has emerged to comprehensively quantify and handle uncertainty. Herein, we summarize and evaluate present best practices, test new approaches to quantify and handle uncertainty, and provide recommendations for future work using remotely-sensed measurements of river migration and width changes. While our research focuses on river systems, the principles and approaches are applicable to research delineating boundaries or using boundaries to measure changes: glacier retreat or advance, erosion or deposition along coastlines and lakeshores, changes in wetland extent, expansion or contraction of vegetation (e.g., deforestation), cliff retreat, sea level rise due to climate change, change in aeolian depositional systems, and anthropogenic/political boundary disputes. From our results, the following conclusions and recommendations arise:

1. Planform change measurements should span spatial intervals larger than coherent units of adjustment to avoid spatial autocorrelation.

2. Uncertainty in manual riverbank delineations is dominated by arbitrary user inconsistency rather than poor image quality (i.e., resolution, colour versus grayscale, year of acquisition) or environmental conditions (i.e., shadows and vegetation cover).

3. We recommend that digitizations follow the vegetated boundary that best approximates bankfull width, whenever possible, to avoid inconsistency along ambiguous reaches.

4. Using a spatially variable level of error detection (LoD) threshold improves the quantity and quality of retained measurements relative to a uniform LoD.

5. After applying a LoD threshold, we recommend first using expert discretion to manually classify any 'nondetect' measurements that qualify as 'significant' measurements of zero (i.e., no change actually occurred).

6. Subsequently, three methods may be used for handling the remaining nondetects; Kaplan-Meier (KM) and Maximum Likelihood Estimators (MLE). The specific approach chosen for handling nondetects is contingent upon each case, but can be guided and informed by descriptions and assumptions of each method, references to external resources, and results of our river-focused analyses.

7. Finally, we encourage a focus on improving the simplicity, generalizability, and open-source opportunities of tools and packages used for calculating river planform change and spatially variable uncertainty, thereby enabling a common platform to measure and compare results.