A neural network approach for climate classification - Sec 7. The 533-5.2 scheme

Last section described the 533 SOM-crystal clustering methodology to obtain this scheme. In this section we will make a light walkthrough of the climate classification scheme. We will go direct because we have go through a similar illustration in Sec 4. Before that we have a little bit postprocessing to do.

Merging subtypes

By crystal clustering we get clusters of the subtypes. We don’t directly re-label the land pixels of the clustered subtypes; instead we find the new cluster centroids by a weighted average of the subtypes’ centroids. Say there’s a cluster of subtypes whose centroids are $\mathbf{c}_1, \mathbf{c}_2, …, \mathbf{c}_m$ (each is a 15-dimensional vector), and their land proportions as weights are $w_1, w_2, …, w_m$, then we have the new cluster centroid as

\[\mathbf{C} = \sum_{i=1}^m w_i \mathbf{c}_i / \sum_{i=1}^m w_i\]

In this way we get 26 new cluster centroids by crystal clustering at $T=5.2$. For each land pixel, assign the type with its nearest neighbor among the new cluster centroids.

The 533-5.2 scheme

Clusters obtained by crystal clustering at $T=5.2$ and the naming. Right-click for full image.

The MST of the new cluster centroids. The graph is plotted such that the edge lengths reflect the distances between cluster centroids. The graph shows that the methodology retains the intrinsic structure of climate data well. Right-click for full image.

The 533-5.2 scheme, 1991-2020 climate normals. Right-click for full image.

	A tropical	C subtropical	D continental	E temperate	F subarctic	Gh/Gk arctic
f humid/oceanic	Af	Cf	/	Efh/Efk	Ff	/
m moderate/semi-oceanic	Am	Cm	/	Emh/Emk	Fm	/
s seasonal	As	Cs	Ds	Es	Fsh/Fsk	/
W arid	AW	CW	DW	EW	/	/
S semi-arid	/	/	DSh/DSk	/	/	/
H highland	/	/	DH	EH	/	/

h/k denotes a warmer/cooler subtype.

The main group, as in the 322 scheme, is largely decided by winter temperature. Subtypes have more details to be specified per group.

All the judging rules mentioned below are manually induced and estimated for a better understanding; none of them is backed by the model. The machine learning model is complex and only the model itself can decide the climate type of a place based on computation.

In the group summary table, temperature is in °C and precipitation is in millimeter (mm). Summer/winter represent the consecutive three months with highest/lowest mean temperature respectively. The interval specifies the (0.05 - 0.95) quantile over the land pixels, calculation based on 1991-2020 climate normals.

The tropical group

	mean temperature	annual precipitation	summer temperature	summer precipitation	winter temperature	winter precipitation
Af	18.6 - 27.8	1327.8 - 3570.0	21.6 - 28.5	208.9 - 1003.3	13.2 - 27.2	195.3 - 1149.0
Am	17.0 - 27.7	910.9 - 2550.5	20.5 - 29.0	160.8 - 890.9	11.7 - 26.8	41.0 - 1061.9
As	20.7 - 29.2	504.8 - 2124.1	22.8 - 33.2	53.0 - 602.6	16.9 - 27.0	0.1 - 739.5
AW	20.7 - 29.5	2.8 - 141.3	27.1 - 36.4	0.0 - 44.6	11.7 - 22.8	0.0 - 55.8
Cs(AS)	17.3 - 29.9	126.5 - 656.6	21.9 - 34.6	11.0 - 315.6	11.2 - 26.6	0.0 - 188.3

Cs is listed in the table for comparison (explained below).

This group is largely warm year round (sometimes hot). According to what is learned by the model, temperature of a dry season is not important. If a place in this group has a cool winter, it must be dry, like Yunnan, China. On the other hand, likewise, in other groups featured by lower temperatures, if a place has a warm/hot summer, it must also be dry, to qualify for that cool group. The tropical title here may be well understood as warm or dry winter, corresponding to a low value of the damp-coldness feature. In general, the coldest season averages above 12°C to qualify for this group. It’s far lower than Köp­pen’s 18°C threshold, with delicate considerations regarding aridity.

It’s natural to have cool winter in deserts due to high continentality and dryness. With such considerations, the standalone arid group like Köp­pen’s B group can be removed and merged into thermal zones, producing a more structured scheme.

The four types in the group follow a decrease of precipitation. Generally, Af has no dry season, Am has a 3-month dry season, As’s is around 6 months, and AW is dry all year round. The type in this thermal zone with about 9 dry months is coded Cs for the integrity of group structure and avoiding confusion with As, but actually the better code may be AS in terms of climatic characteristics.

There are places classified in this tropical group which experience much cooler winter even below the mentioned 12°C threshold, e.g. the Sichuan basin in China. There are several factors accounting for this. First, the land cover scheme we use does not clearly distinguish between tropical and subtropical zones. The Evergreen broadleaf forest land cover type can respond to both tropical and subtropical forests, or even some temperate areas. Second, compared with the Yangtze River Plain, Sichuan basin has a pronounced warmer and less rainy winter. According to centroid-based cluster analysis, this characteristic locates the area closer to Am’s centroid rather than Cf’s. This clustering is based on comparison with cluster centroids rather than Köp­pen’s decision-tree like thresholding. The classification is reasonable considering these factors.

Af humid tropical (tropical rainforest) climate, land coverage 6.43%. This type has the lowest continentality over all types.

Am moderate tropical climate, land coverage 5.96%. It’s among the types with highest seasonality due to pronounced dry-wet season pattern and very good growing conditions in the long rainy season.

As tropical dry-wet season (savanna) climate, land coverage 9.88%. It’s among the types with highest seasonality due to pronounced dry-wet season pattern.

AW hot arid (hot desert) climate, land coverage 6.51%. This type has the lowest damp-coldness over all types. The seasonality is also very low due to the constantly dry condition.

The subtropical group

	mean temperature	annual precipitation	summer temperature	summer precipitation	winter temperature	winter precipitation
Cf	12.2 - 19.2	945.9 - 1837.7	19.5 - 28.0	188.3 - 743.4	1.3 - 11.5	81.4 - 407.7
Cm	12.2 - 25.0	400.2 - 960.2	18.8 - 28.7	51.4 - 398.9	2.2 - 22.4	30.1 - 279.3
Cs(AS)	17.3 - 29.9	126.5 - 656.6	21.9 - 34.6	11.0 - 315.6	11.2 - 26.6	0.0 - 188.3
CW	10.9 - 23.9	119.1 - 393.2	16.2 - 30.7	13.3 - 149.3	4.7 - 16.8	6.6 - 128.9

This group has a cool winter compared with the tropical group. The coldest season averages around 2-15°C in general; within the group, lower temperature usually goes with less precipitation in winter. Summer temperature is usually comparable to the typical tropics (above 20°C).

The structure is largely similar to the tropical group, with the exception of the Cs type, as explained above. Another notable type is Cm, which, more often than not, experiences mild aridity through the year rather than a more pronounced dry-wet season pattern as in Am.

Cf humid subtropical climate, land coverage 2.42%. Largely equivalent to Köp­pen’s Cfa.

Cm moderate subtropical climate, land coverage 3.00%. Much of the subtropical Mediterranean area (Köp­pen’s Csa) falls into this type.

Cs hot semi-arid (hot steppe) climate, land coverage 4.84%. It covers the core area of the Am type in the 322 scheme, as well as Köp­pen’s BSh. Though not truly subtropical, it does feature one 3-month wet season.

CW arid subtropical (subtropical desert) climate, land coverage 3.49%. This type dominates the deserts in the Southern Hemisphere, featured by mild temperatures (by desert standard) and slight precipitation. In the Northern Hemisphere the strong continentality makes the types of EW and DW.

The continental group

	mean temperature	annual precipitation	summer temperature	summer precipitation	winter temperature	winter precipitation
Ds	-5.3 - 9.1	335.6 - 879.9	13.1 - 22.3	166.0 - 426.3	-27.0 - -5.0	9.3 - 81.6
DSh	0.1 - 11.6	166.2 - 597.0	13.4 - 23.9	37.0 - 345.7	-18.3 - 0.2	4.7 - 148.8
DSk	-2.7 - 9.7	169.0 - 488.7	15.5 - 24.7	42.9 - 165.6	-22.0 - -4.6	29.9 - 119.5
DW	4.1 - 13.7	26.7 - 221.8	18.7 - 28.1	9.9 - 94.1	-12.8 - -0.5	1.2 - 58.8
DH	-6.0 - 5.8	64.6 - 664.6	5.0 - 11.2	43.7 - 364.5	-19.0 - 1.9	1.3 - 65.3
Emk(Df)	-3.4 - 10.1	693.3 - 1302.2	11.5 - 21.7	196.4 - 413.3	-20.5 - -1.8	90.6 - 318.2

Emk is listed in the table for comparison (explained in the temperate group).

This group is more true of continental compared with Köp­pen’s D group, distinguished from the temperate but less continental zone, thanks to the continentality climate feature learned by the model. Dryness and large annual temperature variance dominate the vast mid-latitude interior continent. No f or m types are literally defined in this group. Winter is usually colder (coldest season averaging minus teens) than the E group, but the code D is adopted to follow Köp­pen’s convention.

Ds continental monsoonal climate, land coverage 3.35%. It’s among the types with highest seasonality. The cold and dry winter combining the warm and wet summer makes the growing conditions swinging greatly in a year. Covers the core area of Köp­pen’s Dwa/Dwb.

DSh semi-arid warm continental (cool steppe) climate, land coverage 3.26%. The condition is dryer/worse comparing with Ds due to warmer temperature or a less summer-concentrated precipitation pattern. Covers the core area of Köp­pen’s BSk.

DSk semi-arid cool/cold continental (cold steppe) climate, land coverage 2.25%. Little precipitation spreads almost evenly through the year. Huge annual temperature range with very cold winters and warm to hot summers marks the high continentality. This type may be called the Kazakhstan climate due to its predominance there.

DW arid continental (continental desert) climate, land coverage 1.99%. Covers the core area of Köp­pen’s BWk. This type has the highest continentality overall, though the highest single place is in Kazakhstan labelled as DSk.

DH continental highland climate, land coverage 1.69%. Covers the core area of Ds in the 322 scheme. This type mostly occurs on the very high plateau over 3000m above sea level in the mid-latitude area, e.g. the Tibetan Plateau and the southern Andes, thus it may also be called the Plateau climate. The snowy windward slopes like the Alps and Caucasus are classified into (semi-)oceanic types such as Efk, Emk, Ff, Fm.

Without a warm summer, the DH type sufficiently qualifies as subarctic, and is classified as such in other schemes like Köp­pen. Observing the mid-latitude continental location of the area, it appears more appropriate to name it as continental highland rather than highland subarctic. The location contributes to the high evaporation which is not observed in the typical high-latitude subarctic zone of similar thermal level. This is a unique feature of this climate type.

The temperate group

	mean temperature	annual precipitation	summer temperature	summer precipitation	winter temperature	winter precipitation
Efh	4.1 - 15.6	595.3 - 1873.8	12.7 - 24.4	15.0 - 142.4	-6.0 - 9.1	212.3 - 816.5
Efk	3.7 - 11.9	1126.7 - 3268.3	8.8 - 19.5	204.8 - 751.9	-3.0 - 7.4	288.6 - 954.8
Emh	6.5 - 12.7	471.7 - 1026.7	13.9 - 23.3	118.9 - 299.0	-3.8 - 5.6	90.2 - 255.4
Emk(Df)	-3.4 - 10.1	693.3 - 1302.2	11.5 - 21.7	196.4 - 413.3	-20.5 - -1.8	90.6 - 318.2
Es	9.3 - 18.3	250.4 - 691.4	18.6 - 29.5	1.3 - 169.7	-1.6 - 9.8	29.7 - 343.7
EW	13.4 - 23.7	80.0 - 278.6	22.8 - 34.8	0.0 - 43.6	1.5 - 12.3	25.3 - 130.0
EH	2.4 - 13.3	365.8 - 1137.6	5.9 - 22.6	207.3 - 609.8	-5.3 - 9.2	5.4 - 68.9

This group features an average temperature of -5-9°C for the coldest season, with the exception of Emk. Even though Emk experiences a much colder winter sometimes comparable to the continental group, the winter is with heavy precipitation from ocean airmass. Also it’s not cold enough in many areas to qualify for a subarctic definition. Due to the maritime moderation and locations in higher latitudes, most areas in this group enjoy a milder summer compared with the subtropical group, with the exception of Es and EW. Es sits in the transition zone between Efh/Cm and DSh and EW is a transitional type between CW and DW.

Efh warm oceanic (wet-winter Mediterranean) climate, land coverage 0.68%. Largely equivalent to Köp­pen’s Csb. This type has the highest seasonality over all types, featured by a distinct dry-summer wet-winter paradigm.

Efk cool oceanic (temperate rainforest) climate, land coverage 1.37%. Covers the core area of Köp­pen’s Cfc. With heavily moderated temperature and abundant precipitation all year round, this type has a similarly lowest continentality after Af. It’s largely a cool version of Af.

Emh warm semi-oceanic climate, land coverage 3.40%. It can be seen as a transitional type between Efk and Fsh, which is the reason why the m subtype can be deemed as semi-oceanic. The mild continentality is expressed by the modest precipitation while the oceanic moderation of temperature is preserved. The type may also be called the European climate due to its predominance on the European continent.

Emk cool/cold semi-oceanic climate, land coverage 3.17%. On the west side of the Atlantic coast, the condition is drastically different, with long, cold and very snowy winters. As opposed to Europe, the oceanic characteristic is expressed by abundant precipitation through the year while the wintry coldness is inherited from continental climates. In some climate classification schemes it’s named the Laurentian climate due to its predominance over the Laurentia bioregion (northeast US and southeast Canada, around the Great Lakes and the St. Lawrence River). It may also be regarded as a humid continental climate (Dfb/Dfc as in Köp­pen), but personally I don’t favor this way. The exceptionally high precipitation disqualifies the definition of a continental climate despite the low winter temperature.

Es temperate dry-wet season (dry-summer Mediterranean) climate, land coverage 1.32%. The climate cannot be called temperate in most cases, just the winter thermal level falls into the warm temperate zone. Most areas feature a hot and dry summer and a cool and rainy/snowy winter, except for the North China Plain. The counter-Mediterranean Es variant experiences a brief hot and rainy season in July and August, and mostly dry in other months of the year. This variant did not exist decades before and the local climate belonged to DSh; it emerged due to the warming winter in this region as shown in the case study in Sec 5. The cool/cold temperate counterpart is Ds. Due to the coldness Ds’s dry season is not pronounced compared with Es.

EW arid temperate (semi-continental desert) climate, land coverage 1.58%. The climate is not truly temperate, just the winter thermal level falls into the warm temperate zone. Similar to Es, most areas feature a hot and dry summer and a cool and rainy/snowy winter, but precipitation is more scarce. Exceptions are in the CW-locked area in Argentina, where some areas have a slightly summer-concentrated rainfall pattern. The EW type in Argentina is closely bordering on CW, just slightly more continental due to inland location or higher elevation.

EH temperate highland climate, land coverage 0.57%. This is a rare type occuring on low-latitude high plateau like the northern Andes or continental low-elevation (1000-3000m) highlands like Lesotho and Qinba mountain range in China. Basically it’s a summer-rain version of Emh.

The subarctic group

	mean temperature	annual precipitation	summer temperature	summer precipitation	winter temperature	winter precipitation
Ff	-14.8 - 3.8	713.5 - 1689.0	-4.6 - 11.6	136.0 - 379.4	-23.7 - -1.4	164.3 - 491.8
Fm	-10.5 - 1.1	338.5 - 707.1	5.6 - 12.4	113.0 - 253.5	-27.5 - -8.6	49.9 - 152.3
Fsh	-10.5 - 5.5	300.4 - 677.2	12.1 - 18.3	133.2 - 255.8	-35.3 - -6.3	36.5 - 131.8
Fsk	-16.2 - -4.1	195.0 - 529.9	7.7 - 13.1	84.4 - 291.5	-43.5 - -20.0	16.3 - 59.9

Like Köp­pen’s definition for the subarctic types (the c subtype, at most 3 months averaging above 10°C), the condition in summer starts to play a role. Aside from a similar or colder winter of the continental group, temperature in the brief summer usually does not go over 12°C on average. The exception is Fsh, which may not be deemed subarctic in many cases. Fsh is basically a non summer-rain Ds, rendering worse growing conditions.

Many regions, e.g. the southern tip of South America and majority of Iceland, fall into Köp­pen’s arctic tundra type, but they are literally maritime moderated, not resembling the typical harsh arctic climate. The climate and vegetation in Ushuaia, Argentina and Iqaluit, Canada are vastly different. Our scheme delineates them (some of these regions even fall into the temperate zone in our scheme).

Ff subarctic oceanic climate, land coverage 1.52%. Compared with Efk, the temperature steps down further and the precipitation is not as much (but still well above evaporation). This type may be called the Icelandic climate due to its predominance there.

Fm subarctic semi-oceanic climate, land coverage 4.20%. A transitional type between Emk and Gh on the thermal axis. Also transitional between Ff and Fsk on the continentality spectrum. This is the type with the highest damp-coldness.

Fsh cold humid continental (taiga) climate, land coverage 10.89%. Compared with Emk, not only colder in general, it also exhibits a larger annual temperature variation and less precipitation. The type may also be called the boreal climate due to its predominance over the boreal Siberia and northern Canada.

Fsk subarctic continental climate, land coverage 5.81%. It’s Fsh transitioning towards Gh with continentality enhanced. This type has a similarly high damp-coldness with Fm due to the extreme coldness in winter.

The arctic group

	mean temperature	annual precipitation	summer temperature	summer precipitation	winter temperature	winter precipitation
Gh	-19.2 - -7.6	58.6 - 343.1	0.3 - 7.9	30.4 - 115.8	-35.9 - -19.3	2.2 - 76.3
Gk	-29.3 - -9.8	3.9 - 624.3	-13.7 - 0.3	1.5 - 136.6	-41.2 - -18.3	0.4 - 184.4

Generally at most three months averaging above 0°C, this group is more true of arctic than Köp­pen’s definition of no month averaging above 10°C.

Gh warm-summer arctic (tundra) climate, land coverage 6.97%. This is the type with a brief period above 0°C.

Gk cold arctic (icecap) climate, land coverage 3.49%. Generally no month averaging above 0°C.

It can be seen many climate types of interesting distinct characteristics are derived, especially shown in the fine temperate zone E. But the scheme is more complicated and harder to grasp compared with the 322 scheme. Both schemes can have their use under different scenarios.

The reason we choose $T=5.2$ is that, on one hand, we feel no need for more climate types other than the above 26 (The result remains the same for $T=5.3, 5.4, 5.5$. If we set a higher $T=5.6$, Fsk will break into two types, which we think no good). On the other hand, if we set a lower $T=5.1$, EH will be merged with DH. We consider EH as a meaningful type so $T=5.1$ or below is not favored. The choice of $T$ is case-specific and depends on the target. Clustering quality metrics may also help to decide the optimal value of $T$.

Comparison with $k$-means clustering

The kmeans-26 scheme, 1991-2020 climate normals.

Above shows the result of $k$-means clustering on the original climate features dataset. As was done in Sec 4, we set $k=26$ and the algorithm ran 200 iterations. The result is obviously poor, with Greenland and the Sahara Desert fractionated. It’s because the $k$-means method favors a clustering with even-sized clusters and the two regions, despite their largely homogeneous climate conditions within, are large in size. Also, $k$-means does not retain any structural information among the clusters.