The South Atlantic must be partitioned into sub-regions within which the relationship between salinity and temperature is relatively homogeneous.
The TS plots indicate that salinity can depend linearly or quadratically on temperature and might also depend on longitude and/or latitude. The relatively small changes in the TS plots between adjacent sub-regions suggests that sparsely sampled regions can be modeled by grouping relatively widely separated data.
To prevent abrupt discontinuities across boundaries between sub-regions used for modeling, overlapping sub-regions (shingles) can be used:
To allow the model for each shingle to reflect most strongly the relationship near its center, the regression can be weighted inversely with the square of the distance of the data from the center.