We observed interannual changes in the temperature and salinity of the surface layer of the Adriatic Sea when measured during the period 2005–2020. We observed non-stationarity and a positive linear trend in the series of mixed layer depth, heat storage, and potential energy anomalies. This non-stationarity was related to the climate regime that prevailed between 2011 and 2017. We observed significant changes in the interannual variability of salinity above and below the mixed layer depth and a positive difference in the surface barrier layer. In an effort to reconstruct the cause of this phenomenon, a multi-stage investigation was conducted. The first suspected culprit was the change in wind regime over the Mediterranean and Northeast Atlantic regions in September. Using the growing neural gas algorithm, September wind fields over the past 40 years were classified into nine distinct patterns. Further analysis of the CTD data indicated an increase in heat storage, a physical property of the Adriatic Sea known to be strongly influenced by the inflow of warm water masses controlled by the bimodal oscillating system (BiOS). The observed increase in salinity confirmed the assumption that BiOS activity affects heat storage. Unexpectedly, this analysis showed that an inverse vertical salinity profile was present during the summer months of 2015, 2017, and 2020, which can only be explained by salinity changes being a dominant factor. In addition, the aforementioned wind regime caused an increase in energy loss through latent energy dissipation, contributing to an even larger increase in salinity. While changes in the depth of the mixed layer in the Adriatic are usually due to temperature changes, this phenomenon was primarily caused by abrupt changes in salinity due to a combination of BiOS and local factors. This is the first record of such an event.

Measuring data efficiently in the framework of geosciences has proven to be more cumbersome than expected, despite technological advances. While remote sensing techniques, such as satellite observations, provide extraordinary spatial coverage, they still lack the fine spatial and temporal resolution of in situ measuring techniques. Naturally, the level of coverage obtained by remote sensing techniques could be replicated with physical measuringstations and devices, however, the financial cost would be immense. Therefore, if we are to broaden the spatial coverage while retaining both resolutions and minimising cost, we need to strategically deploy as few sensors as possible. In order to tackle this problem, we haveutilised three unsupervised learning (clustering) methods not only to demonstrate how a smaller subset of sensors can provide significant measurement accuracy, but also to show that there exists an optimal sensor placement (as opposed to random placement). Data used for thisdemonstration is ERA5 wind components at 10m height from 1979 to 2019 over the Mediterranean sea, at a spatial resolution of 0.5° × 0.5° every 6 hours.Clustering methods used are K-means clustering, Self-Organising Maps (SOM) and Growing Neural Gas (GNG). We have clustered the data into 5, 10, 20, 50, 100, 200 and 500 groups and treated the median centers of the resulting domains as the optimal placement for sensors. After the clustering was completed, we have attempted to reconstruct the missing data using two regression models: linear and K-Nearest Neighbours. Reconstructed data was compared (in both size and angle) to original data, and the results show that with just 5 points (out of a grand total of 1244 wet points), reconstruction accuracies are as follows: 65.6, 65 and 62.5% for linear regression reconstruction and 71.6, 71.2 and 70.5% for KNN reconstruction, when applied to GNG, K-means and SOM respectively. Increasing the number of points has diminishing returns (especially in excess of 100 points), with linear regression reconstruction accuracy peaking at ≈ 95% and KNN reconstruction remaining in the high 70%. As demonstrated, GNG and K-means performed slightly better than SOM, due to the nature of SOM’s rigid algorithm.This work has been supported by Croatian Science Foundation under the project UIP-2019-04-1737.