The concept of seasons is something that is part of our daily life but defining it can be challenging. We associate seasons with everything that shows a certain periodicity, which we call seasonality. However, if this periodicity is the seasonality, what is the formal definition of seasons? This lack of clarity seems not to be confined to terminology and leads to an ambiguity that can become limiting when we wonder what will happen to the seasons in the future.
Meteorological seasons are more of a heuristic concept than well-defined entities. This is because it is not possible to give a globally valid definition of meteorological seasons. On the other hand, they are widely used in the evaluation of climate change. Furthermore, the seasonal patterns have great impact on many other sectors, such as agriculture, tourism, and epidemiology.
The division into seasons that is used nowadays is always used when analyzing the future projection obtained by climate models. However, there is evidence that in the last decades, a wide range of seasonal patterns has changed, and we can assume that the division into seasons should be constantly verified and updated, especially in a context of climate changes.
The first purpose of this work is to develop a methodology for the identification of meteorological seasons in climatic datasets, trying to minimize the arbitrary assumptions.
The second purpose of the work is to develop a methodology for evaluating how the seasons detected are represented in different climate datasets. This would allow us to study how the seasons we are experiencing nowadays are expected to change in the future under the influence of climate change.
We have developed and tested a set of machine learning algorithms for fulfilling these tasks. Thus, we have applied them to a case-study region, the Hindu Kush – Karakoram – Himalaya (HKKH). This region is particularly interesting because it shows two different seasonal patterns, a four seasons pattern in the Hindu Kush – Karakoram subregion and a two seasons pattern in the Himalaya subregion.
We found that these tasks can be adequately satisfied by machine learning algorithms. For our first purpose, we found that a Radially Constrained Clustering algorithm can recognize a partition of seasonal patterns into the meteorological seasons which is physically meaningful and well understandable, without the need of human supervision.
For our second purpose, we found that a SoftMax perceptron can easily recognize the meteorological seasons in future projections. This led to other interesting results, for example a correction of expected seasonal precipitation trends in the HKKH.