Skip to content
Joseph Smith

Marsh Change Mapper

Draw a study area, place spatially balanced sample points, classify cover at each point on historical and recent aerial imagery, and test for change.

Your work stays in this browser. Points and classes are not sent to this website. Imagery streams from the New Jersey Office of GIS (NJOGIS) or Esri. Download the CSV as a backup before clearing browser data.

Imagery
Study area

No study area yet.

Sample points

Points use Balanced Acceptance Sampling. Spare points continue the same sequence: if a point can't be classified, replace it with the next spare, in order.

Methods

Sampling. Points are a Balanced Acceptance Sample (Robertson et al. 2013, Biometrics 69:776–784). A two-dimensional Halton sequence with a random start covers the study area’s bounding box, and points inside the study area are kept in sequence order. The sample spreads evenly in space, like GRTS. Any leading run of points is also balanced, so spare points can replace unusable ones in order. The seed reproduces the sample.

Change. Each point has a class at both times, so the data form a paired transition table. Counts are modelled as Multinomial with a Dirichlet prior. The prior is weak by default: a total weight equal to one point, spread evenly over cells. Each posterior draw of the transition probabilities gives the cover of every class at both times and their difference. Results are posterior medians with 95% equal-tailed credible intervals, from 10,000 draws. P(increase) is the posterior probability that a class’s share went up. Because both times come from the same points, the comparison accounts for the pairing. Two separate proportion tests would not.

Swap. Following Pontius & Santacruz (2014, Remote Sensing Letters 5:270–277), a class’s gross loss is its share at time 1 minus its persistence. Gross gain is its share at time 2 minus its persistence. Net change is gain minus loss. Swap is twice the smaller of gain and loss: loss in one place matched by gain somewhere else. Overall change splits into quantity (net) and swap.

Interior and edge marsh. Each class is marked as marsh, interior non-marsh (ponds) or edge non-marsh (open water). Interior loss is the share of the study area that went from marsh to an interior class, and interior gain is the reverse. Edge loss and gain are the same for edge classes. Net change at each setting is gain minus loss, computed in every posterior draw, so its credible interval and P(net loss) carry the full uncertainty. Changes among marsh classes, or among non-marsh classes (for example pond to open water), are not marsh loss or gain.

Area. If a study area is given, proportions convert to hectares. The area comes from a spherical approximation and is suited to marsh-scale polygons, not survey work.

Imagery. New Jersey imagery is streamed from NJOGIS web map services (img.nj.gov). Please credit the NJ Office of GIS and the data steward (for 1970 and 1977, NJDEP) in products that use them. Historical photos are georeferenced scans and may be offset by several metres. Check alignment against stable features before classifying.