What are the advantages and disadvantages of using stratified random sampling for accuracy assessment?

1 Why stratify?

Stratified random sampling offers several benefits over simple random sampling, which does not consider the spatial distribution or the thematic variation of your map. By stratifying your map, you can reduce the sampling error and sample size, as you can allocate more samples to strata with higher variability or importance. You can also ensure a representative and balanced sample of all classes and categories in your map by guaranteeing a minimum number of samples per stratum. Additionally, you can incorporate prior knowledge or auxiliary data into your sampling design, as you can define the strata based on criteria such as land cover type, elevation, slope, or spectral characteristics.

2 How to stratify?

Depending on the type and scale of your data, as well as the objectives and resources of your project, there are various methods for stratifying a map. These include equal-area stratification, which divides the map into equally sized regions; equal-frequency stratification, which divides the map into regions with equal number of pixels or polygons; and optimal stratification, which divides the map into regions that minimize the within-stratum variance and maximize the between-stratum variance. Equal-area stratification is simple and easy to implement, but may not capture heterogeneity or importance of different regions. Equal-frequency stratification is more sensitive to thematic variation, but may create irregular or fragmented regions. Optimal stratification is more efficient and accurate, but requires more computational power and data availability.

3 What are the challenges?

Stratified random sampling can be challenging, and you should be aware of the potential drawbacks. For instance, choosing the right number and size of strata is important, as it can affect accuracy, precision, feasibility, and cost. Having too few or too large strata can reduce the benefits of stratification and introduce bias or uncertainty. Additionally, dealing with mixed or unclear pixels or polygons may cause confusion during sampling or validation. Rules or techniques like majority rule, fuzzy logic, or spectral unmixing may be required to assign them to a stratum. Finally, spatial autocorrelation or dependence of pixels or polygons within or across strata may violate the assumption of independence and influence statistical inference and error estimation. To address this issue, methods such as spatial statistics, spatial weights, or spatial bootstrapping can be used.

4 How to improve it?

Stratified random sampling is a useful and flexible tool for accuracy assessment, but it should be tailored to the characteristics and objectives of a project. Reviewing the literature and standards for accuracy assessment, exploring and analyzing data and maps, planning and testing the sampling design, and documenting and reporting the results are all important steps to ensure success. Additionally, sufficient information and justification should be provided for any choices or assumptions made.

5 Here’s what else to consider

This is a space to share examples, stories, or insights that don’t fit into any of the previous sections. What else would you like to add?