Climate event classification based on historical meteorological records and its presentation on a Spatio-Temporal research platform

Abstract

Climate change has become a serious issue, and tracing climate events from historical records could be a solution to find a way to deal with it. This study conducted two experiments for classifying metrological text data—one unsupervised method for exploring a solution in the lack of labeled data and another supervised method for achieving high-performance classification. Both experiments took the meteorological text records as material in the early Qing Dynasty (1644 C.E. to 1795 C.E.) from the REACHES database. We also integrated the classification results to develop a Spatio-Temporal research platform with an instant response front-end interface to help humanity researchers access and analyze data according to the three dimensions of time, area, and event categories. With our Spatio-Temporal research platform, we had the ability with ease to analyze the meteorological records during 1650 C.E. to 1700 C.E., the late stage of the Little Ice Age, to investigate the phenomenon of climate change in the Qing Dynasty of China. We will continue to expand the capacity of the database and establish a mature Spatio-Temporal research platform in the future.

1 Introduction

The impacts of climate change have become more and more apparent. Understanding its cause and effect to mitigate its deleterious consequences has become an important research topic. Many clues leading to climate disasters can be traced by observing past meteorological records documented in historical materials. The East Asian Historical Climate Database (Wang et al., 2018) (simplified as REACHES database), categorized chorographies and official histories from A Compendium of Chinese Meteorological Records of the Last 3,000 Years (here we simplified as the Compendium) (Zhang, 2004) into 27 main categories under four domains—Meteorology, Hazard, Unusual phenomena, and Others. The classification system shows interactions or relationships between each category that could be further investigated, making a significant contribution to the analysis of the temporal and spatial characteristics of meteorological phenomena.

Inspired by REACHES, we conducted two experiments on classifying metrological text data–one unsupervised method for exploring a solution in the lack of labeled data, and another supervised method for achieving high-performance classification—with the aim of categorizing a large corpus of data efficiently. Both experiments took the meteorological text records in the early Qing Dynasty (1644 C.E. to 1795 C.E.) from the REACHES database. We then integrated the classification results with a map and a timeline to develop a Spatio-Temporal search interface that facilitates climatologists to analyze data according to time, area, and meteorological categories.

2 Unsupervised Methodology

The first experiment of this study was meteorological text clustering (Chinea-Rios et al., 2015; Hammouda et al., 2004; Kotlerman et al., 2012; Wang et al., 2018). The experiment was composed of three steps: pre-processing, text representation generation, and k-means clustering (MacQueen, 1967). We took 36,123 historical meteorological records from the REACHES database as our input data. As shown in Table 1, each record contains six fields. To decrease the noise of text data which seems not related to meteorological semantic information, we first pre-processed input data, including (1) Replacing characters in descriptions about GanZhi—also called Sexagenary Cycle, a traditional numbering method of time in East Asia, with a character g. Year, Month, and Number expressions were replaced by y, m, and n, respectively. (2) Removing place names and punctuation symbols because they have nothing to do with classification. (3) Some descriptions mentioned multiple categories, and some are long, so we separated each description by periods into sentences to make it possible to distinguish multiple meteorological information from each description, and decrease text information bias.

We used these meteorological text data and the Ming Record (明實錄) to train a 200-dimensional word2vec model (Mikolov et al., 2013). We decided to use one Chinese character as a word in the word2vec algorithm, converted each sentence written in Classical Chinese into embedding vectors by averaging their character embeddings, and then used the k-means algorithm to divide all embedded vectors into k groups (Qian et al., 2004; Wang et al., 2008). We constructed the validation set to find the most suitable k value of 300, and evaluated the clustering results with 9,530 labeled records. The true-positive, false-positive, and false-negative numbers of each event classification are calculated to get the total precision, recall, and F1-score, see Table 2.

From Table 3, we could see that many clusters corresponded to the same climatic category, but after careful examination, these groups were still slightly different. Take Clusters 0 (see Table 4), 9 (see Table 5), and 95 (see Table 6) as examples. Even though these three clusters could be roughly classified as flood hazards, it was found that most of the texts of Cluster 95 referred to seasons. The records of Cluster 9 mostly referred to floods killing people and damaging villages, while the records of the other two groups do not.

3 Supervised Methodology

In the first experiment, we found that it was possible to help researchers generalize the data or even collect training instances through the above demonstrated unsupervised machine learning approach. In the second experiment, we changed our target to build a high-performance classifier by training a multi-label classifier based on a supervised method using BERT–Bidirectional Encoder Representation from Transformers (Devlin et al., 2018). BERT is a two-step deep learning framework composed of pre-training and fine-tuning procedures. Since we designed the model with multi-label outputs, separating each description into sentences in advance for better clustering results became unnecessary. We utilized 9,530 labeled records without splitting each description into sentences, evenly extracted 80% from each category for training and 20% for evaluation. Because some of the main categories lacked training instances, we merged similar ones and turned them into twenty-four categories (see Table 7).

We adopted the pre-trained Chinese BERT language model to encode the text descriptions and formed the labels in training instances into a series of binary-decision problems as in Table 8. Set the batch size to thirty-two, the learning rate to 1e-5, and the max length of each sentence no longer than 255 characters with character-based segmentation.

We evaluated the classification model in every twenty epochs to see the results in Table 9, calculated both the micro and macro metrics for every event label, using precision, recall, and F1-measure. Finally received a high performance of 96.7% micro F1-score.

4 Front-End Interface

We presented meteorological events on a map interface based on the year and location of the climate events in the historical meteorological records, providing a platform for researchers to easily find desired data (see Fig. 1).¹ The main features of the interface include a scrolling timeline, a pop-up condition selection window, and an instant response map. When a user selects the conditions, the map will correspondingly display the records satisfying the conditions. If the cursor hovers over a location tag on the map, the map on the page below will show all meteorological records of the location within the timeline interval.

5 Case Study

To show the usage of our platform, we looked into meteorological records from 1650 C.E. to 1700 C.E. in our database, which is the late stage of the Little Ice Age, to investigate the phenomenon of climate change in the Qing Dynasty of China (Chu, 1973; Solomon et al., 2007). First, we chose the Temperature event category. Among these records, there are sixty-eight extreme cold climate records. We could see that this kind of phenomenon was located from the tropical zone to the tepid zone in China (see Fig. 2). Therefore, we can conclude that extreme cold records appear in the middle- to high-latitude areas and lower latitude areas.

Extreme cold weather usually comes with disasters. We included the Rain category as our variables in Step 2. Among these records, there were 379 records related to Snowfall. To deeply look into the records, besides directly collecting the data of meteorological phenomena of Snowfall, we further collected the indirect meteorological phenomena data, such as ‘three days in a row’, ‘frost-damaged trees’, etc. (see Table 10). From Table 10, we can see that during the Little Ice Age (1650 C.E. to 1700 C.E.), the freezing climate led to disasters more frequently than that during the Non-Little Ice Age (1745 C.E. to 1795 C.E.).

When investigating the meteorological phenomena during the Little Ice Age, we found uncommon snowfall records in Taiwan. As shown in Table 11, these records were in terrain areas, such as Chia-Yi county and Tainan City, which could be seen as solid evidence of the extreme climate during the Little Ice Age.

With the help of our platform, users can investigate the relationships between different climate events. Through keywords such as ‘locusts’ and ‘grasshoppers’, we had conducted statistics on the Locust Plague events, and compared them with the Drought events in Shandong Province. Because the frequency has an amplification effect, the intensity of ten times is actually not ten times stronger than that of one time. Therefore, we follow the way used for information retrieval and take the logarithm of frequency+1 to calculate the Pearson correlation coefficient between the log frequency of Locust Plague and that of Drought as shown in Fig. 3. The result is that the correlation coefficient value is 0.4 and significant, indicating that the two have a moderately positive correlation. Judging from the degree of graph overlap, the peaks and troughs almost correspond to each other (peak to peak, trough to trough). We can observe a certain degree of correlation between the occurrence of Drought and Locust Plague based on the number of records in historical texts alone, and it can also echo the saying that ‘the drought followed by the locust plague’. However, the number of locusts is not simply affected by drought. According to the existing references, the combined results of temperature, summer rainfall, and winter length must also be considered. Therefore, we are also actively collecting data such as precipitation and temperature for more detailed analysis of the causes of locust plagues.

6 Conclusion

In order to understand the impact of climate disasters and find a way to deal with them, tracing the historical climate event could be a solution. This study established a principle to classify meteorological phenomena from China's Three Thousand Years of Meteorological Records based on machine learning strategies, developed a Spatio-Temporal research platform, and built an instant response front-end interface. Since language model techniques are developing rapidly, getting good classification results becomes easier with well-prepared training instances. However, producing high-quality training instances is still costly for personal research. Therefore, we proposed a two-stage process of utilizing an unsupervised method to help collect training instances through briefly browsing our text material in the first stage, and then, in the second stage, use a supervised method to get better training performance.

In the climate case study, we collected users’ feedback, improved the front-end interface of our platform, and enhanced the precision of a mass of meteorological data analytics. Although the research platform only contains the meteorological data from 1647 C.E. to 1795 C.E. for now, we hope to expand the capacity of the database and establish a mature Spatio-Temporal research platform in the future.

Funding

This work was supported by the Center for GIS, RCHSS, Acdemia Sinica, Development of Big Historical Text Information Extraction Techniques for Compiling Research-oriented Knowledge Bases under grant AS-ASCDC-110-101, Automatic Wikipedia Article Generation Based on Articles in Other Languages under grant MOST 109-2221-E-008 -058 -MY3.

References

Endnotes