The problem addressed in this article has attracted sustained attention over the past decade, yet a number of practical obstacles remain unresolved. This work sets out an approach that addresses those obstacles directly and evaluates it against established benchmarks.
Prior work falls broadly into two strands. The first treats the problem as one of representation; the second as one of optimisation. Our contribution sits between the two, borrowing the representational insight of the former while retaining the tractability of the latter.
Our approach proceeds in three stages: preprocessing, model construction and evaluation. Each is described in sufficient detail for independent reproduction, and our implementation is available on request.
Across all evaluation conditions the proposed method performs at least as well as the strongest baseline, and substantially better in the low-data regime that motivated this work. Performance degrades gracefully as noise increases, which we attribute to the regularisation strategy described in section 3.2.
Two limitations warrant emphasis. The evaluation is confined to a single geographic context, and generalisation beyond it is untested. The method also assumes a data volume that smaller institutions may not have.
We have presented and evaluated an approach to Civil & Environmental Engineering problems that improves on established baselines while remaining computationally tractable. Future work will extend the evaluation to additional contexts.