A neural network approach for climate classification - Sec 4. The 322 scheme

This section will showcase and interpret the climate types which are the result of simple SOM clustering as discussed in the last section. I would like to name this scheme as the 322 scheme.

The 322 scheme, 1991-2020 climate normals. Right-click for full image.

The borders function by Chad Greene is used to plot country borders. It’s merely for the legibility of the maps and the plotted borders do not carry any political meanings. Locations of some cities and borders may be inaccurate. Use of this code does not mean the author’s acknowledgement of the information.

The 322 scheme

Based on our analysis of the three principal climate features in Sec 2, it’s not hard to interpret the climate types. The first principal axis is the so-called damp-coldness, which, basically put, stands for (winter) temperature level. So the scheme defines three temperature levels: warm (tropical and subtropical), medium (temperate), cold (arctic and subarctic). The continentality-seasonality plane is a bit more intricate. The same position on the plane can have different ecological significance due to different temperature levels. For example, a high level of seasonality almost always favors plant growth in the cold zone because low seasonality means coldness all year round there. However, in the warm zone, a high level of seasonality leads to a dry-wet season pattern which is not the best, compared with a low seasonality combined with a low continentality, which means a rainforest climate. On the other hand, a low seasonality combined with a high continentality means a hot desert climate in the warm zone. Despite this intricacy, we conclude four climate subtypes for each temperature level:

• f subtype: the best growing condition on the temperature level.
• m subtype: moderate, not markably seasonal and modest growing condition overall.
• s subtype: seasonal and modest growing condition overall.
• x subtype: the worst growing condition on the temperature level, due to extreme aridity or coldness.

The table below gives a set of names and abbreviations that is easy to understand for the climate types:

	f best	m moderate	s seasonal	x worst
A warm	Af rainforest	Am warm semi-arid	As dry-wet season	Ax hot desert
D medium	Df temperate	Dm oceanic	Ds highland	Dx dry continental
F cold	Ff continental monsoonal	Fm humid continental	Fs tundra	Fx icecap

The diamond on the map locates the place whose climate is closest to the cluster centroid. The climate chart in the lower left corner illustrates the climate of that place as a good representative of the climate type. The yellow bars in the climate chart stand for the monthly potential evapotranspiration (PET). The elevation data source is GMTED2010. Elevation is labelled for reference only; the model does not use elevation data in calculation. Right-click for full image.

The location information of the diamonds is queried from OpenStreetMap. Again, it’s merely for legibility and the information may be inaccurate. It does not mean our acknowledgement or endorsement of the location information.

The land coverages calculated below are based on our dataset with 66501 land pixels at $0.5° \times 0.5°$ resolution, 1991-2020 climate normals, Antarctica excluded.

The warm group

Af rainforest climate, land coverage 10.78%. Aside from the traditional Af/Am Köppen counterparts, it also covers the southeastern parts of large continents, like southeastern US and China. These regions are classified as the subtropical Cfa in the Köppen scheme. Despite a cooler season, there’s abundant precipitation all year round.

Am warm semi-arid climate, land coverage 8.39%. Largely equivalent to Köppen’s BSh, with some cooler BSk area included.

As dry-wet season/savanna climate, land coverage 13.35%. Largely equivalent to Köppen’s Aw, with some cooler Cwa/Cwb/Csa/Csb area included.

Ax hot desert climate, land coverage 9.99%. Largely equivalent to Köppen’s BWh, with some cooler BWk area included.

The medium group

Df temperate climate, land coverage 6.82%. This type does not have a Köppen counterpart, but it’s easy to see that it covers the world’s most human livable areas. Most areas have distinct four seasons. The summer and winter are both not unbearable and the humidity varies within a year, while there’s no pronounced dry season. Some have a slightly dryer winter, some have a slightly dryer summer - our model does not explicitly distinguish that.

Dm oceanic climate, land coverage 4.08%. This type is not equivalent to Köppen’s oceanic Cfb, but more like Cfc. It’s featured by a long, cold and very snowy winter. The precipitation is very high in winter while the temperature constantly hovering around 0°C. Temperatures are heavily moderated all year round. Winter is not extreme and summer is usually mild.

Ds highland climate, land coverage 2.18%. With no knowledge of altitude, the model successfully recognizes the highland climate type. This is probably due to the NN’s capturing the exceptionally high evaporation combined with low temperature which is a unique feature of highland. Note that this type only includes the dry continental highland where altitude is mostly above 3000m (e.g. the Tibetan Plateau and the Andes). The wet and snowy windward slopes like the Alps and Caucasus are classified as the oceanic Dm.

Dx dry continental climate, land coverage 8.14%. Largely equivalent to Köppen’s BWk/BSk, excluding the non-continental areas in Africa and Australia.

The cold group

Ff continental monsoonal climate, land coverage 11.20%. The more continental part of the boreal. Overall precipitation is low but more concentrated in summer, and the annual temperature range is very large. Winter is extremely cold and dry, while the summer can sometimes be hot.

Fm humid continental climate, land coverage 11.58%. The ocean-influenced part of the boreal. Precipitation is moderate and evenly spread over the year, and temperature is milder than Ff.

Fs tundra climate, land coverage 9.34%. Köppen’s ET in the high North, the oceanic and highland (alpine) induced areas excluded. It’s like a harsher version of Ff.

Fx icecap climate, land coverage 4.12%. Largely equivalent to Köppen’s EF.

It can be seen that the 322 scheme is very clear, concise, structurized and easy to interpret and understand. There’s no need to have a standalone arid group B as in the Köppen scheme. However it omits a lot of details - some types are very broad, and the subtropical and subarctic climate types are not distinguished. We are going to derive a more detailed climate classification scheme later.

Comparison with $k$-means clustering

Before going on, let’s make a comparison with the result of $k$-means clustering. $k$-means clustering was done on the same set of climate features, where $k=12$ and it ran 200 iterations. The result gives a new type similar to Köppen’s tropical and subtropical monsoonal Am/Cwa, and the oceanic type Dm is merged into the temperate type Df. This scheme is less structurized and thus not preferred. SOM shows its advantage in this problem.

The kmeans-12 scheme, 1991-2020 climate normals. Right-click for full image.

The 322 scheme, 1991-2020 climate normals. This map is for comparison here. Right-click for full image.