Declining groundwater levels due to the absence of a planning system makes aquifers vulnerable to subsidence. This paper investigates possible hotspots in terms of Subsidence Vulnerability Indices (SVI) by applying the ALPRIFT framework, introduced recently by the authors by mirroring the procedure for the DRASTIC framework. ALPRIFT is suitable to cases, where data is sparse, and is the acronym of seven data layers to be presented in due course. It is a scoring technique, in which each data layer bears an aspect of land subsidence and is prescribed with rates to account for local variability, and with prescribed weights to account for relative significance of the data layer. The inherent subjectivity in prescribed weights is treated in this paper by learning their values from site-specific data by the strategy of using artificial intelligence to learn from multiple models (AIMM). The strategy has two levels: (i) at Level 1, three fuzzy models are used to learn weight values from the local data and from observed target data, and (ii) at Level 2, genetic expression algorithm (GEP) is used to learn further, in which the outputs of the models at Level 1 are reused as its inputs and observed data as its target values. The results show that (i) the Nash-Sutcliff Efficiency (NSE) coefficient for ALPRIFT with measured land subsidence values is approx. 0.21; (ii) NSE is improved to 0.88 by learning the weights at Level 1 using fuzzy logic, and (iii) NSE is further improved to 0.94 by further learning at Level 2 using GEP.