An Image-based Global Climate Sentiment Index

Abstract

As the impacts of climate change on global economic and social systems deepen, identifying and quantifying climate sentiment has become crucial for climate policy design, risk monitoring, and cross-sector governance. To address limitations in existing studies—particularly their reliance on textual data, which introduces strong language dependence and limits cross-country comparability—this study incorporates visual information as the medium for sentiment measurement. Using the Google Images search engine, we collect approximately 3.19 million climate-related images from 19 representative countries over the period 2015–2024. By detecting positive and negative sentiment embedded in visual data, we construct a globally comparable, long-horizon, multi-frequency Climate Sentiment Index. This index transcends linguistic boundaries, substantially improving the accuracy and cross-cultural robustness of sentiment identification. It provides a valuable data foundation and analytical tool for understanding public climate attitudes, tracing risk transmission channels, and analyzing the formation of market expectations, with broad applicability across economic, social, and environmental domains.

Background and Summary

Climate change has increasingly become a major source of financial risk¹. On the one hand, the growing frequency of extreme weather events and persistent global warming can directly damage infrastructure and reduce productivity^2,3, thereby affecting asset valuation, credit quality, and investment returns for firms⁴. On the other hand, low-carbon transition policies aimed at addressing climate change may lead to stranded assets and industry restructuring, introducing indirect risks and uncertainties into financial markets^5,6. Against this backdrop, market participants’ perceptions and expectations regarding climate change influence financial decision-making, while digital platforms such as Google serve as key channels for information dissemination and sentiment formation^7,8. By shaping investor sentiment and market expectations, climate sentiment can not only amplify financial markets’ sensitivity to climate risks but also steer capital flows toward green and low-carbon activities. Therefore, developing long-term, fine-grained, and reproducible measures of climate sentiment embedded in public information is crucial for understanding financial markets’ exposure to climate risks.

Existing climate sentiment indicators are largely based on social media^9,10. While social media platforms, with their real-time nature, enable individuals to express opinions at high frequency and facilitate the rapid dissemination of climate-related information and sentiment, the user base is highly uneven. As a result, the information generated tends to be strongly subjective, lacks a financial context, and is easily influenced by online noise and platform algorithmic biases, leading to pronounced sampling distortions¹¹ that limit its ability to represent mainstream societal climate sentiment. In contrast, Google—being the most widely used search platform globally—employs a relatively stable data-indexing mechanism that better reflects the perceptions of the general public. Moreover, Google’s content sources are broad, encompassing news, government reports, academic institutions, and publicly uploaded content. Because it does not rely on users actively posting, it captures the broader information environment rather than the opinions of specific user groups.

Artificial intelligence is driving an intelligent transformation in social science research paradigms through large-scale data collection and analysis. With the exponential growth of computing power and the rapid development of internet search engines, the costs of data collection, storage, and transmission have declined substantially. As a result, textual materials available for empirical analysis—such as news comments¹², social media posts⁹, annual reports of listed firms¹³, analyst reports¹⁴, and even earnings conference calls¹⁵—have become increasingly abundant¹⁶. Benefiting from major advances in computational linguistics, text mining, and machine learning, many researchers have incorporated text-analysis techniques into the fields of finance and accounting to conduct sentiment analysis on textual corpora¹⁷. In particular, event-centric frameworks that translate unstructured news into structured signals for market prediction^18,19, alongside research on social media emotion-market linkages²⁰, have established robust media-market evidence. Existing applications include the climate change news index²¹, climate physical and transition risk indices²², climate policy uncertainty index^23,24, the expressions of public views index on climate risk and the government responses index on climate risk²⁵, and the managerial climate attention index²⁶. These text-based indicators have also demonstrated significant predictive power for carbon and broader financial markets^27,28. However, these indices suffer from two fundamental limitations. First, textual content is highly dependent on language and narrative conventions, which vary substantially across countries, resulting in limited cross-country comparability of sentiment measures. Second, textual content is often shaped by editorial positions, platform rules, and discourse frameworks, making text-based sentiment identification more subjective. This inherent subjectivity underscores a critical identification challenge: the imperative to extract genuinely unique information rather than merely repackaging the same news cycle variations already captured by existing text or search proxies.

Using images as a data source for measuring climate sentiment offers several distinctive advantages that directly address this challenge. First, images are closer to real-world conditions than text²⁹. The sources of Google Images are relatively decentralized³⁰, including media illustrations as well as on-site photos and schematic images uploaded by government agencies, research organizations, social groups, and individual users. Compared with text reports that undergo editorial filtering and discursive reconstruction, image-based information is, on the whole, more proximate to the underlying events^31,32. Second, humans are inherently more sensitive to visual information. Images are often vivid and emotionally charged³³, capable of simultaneously activating sensory experience^34,35 and cognitive processing systems³⁶. The “picture superiority effect”, a well-established phenomenon in psychology, indicates that images are more easily remembered and recalled by the human brain than text³⁷. Numerous studies show that images in news content attract greater attention³⁸ and that visual content is more readily recalled than textual descriptions³⁹. Images can also prompt individuals to think actively about a topic and extend such thinking to related issues³⁶. It is undeniable that a picture is worth a thousand words^40,41. Third, images overcome semantic and cultural-contextual differences. Text-based sentiment identification is influenced by factors such as linguistic context, expression style, and translation bias, whereas images convey emotional cues—through color, composition, scene selection, and symbolic elements—that text cannot precisely describe^41,42. Particularly in typical visual scenes involving extreme weather disasters, pollution emissions, and clean-energy activities, images rely on shared human visual perception to transmit sentiment, rendering their emotional content largely cross-culturally consistent⁴³.

Based on the considerations above, this study develops the Global Climate Positive Sentiment Index and Climate Negative Sentiment Index using Google Images over the period from January 1, 2015, to December 31, 2024. As the most widely used image-search engine worldwide, Google Images offers extensive coverage and strong cross-lingual retrieval capability, making it one of the most effective channels for accessing visual information on climate-related topics across multiple countries and regions globally. To systematically capture public sentiment toward climate change at the visual level, we construct a multilingual climate keywords dictionary and use it to collect, clean, and classify 3.19 million climate-related image samples from Google Images. These images are then aggregated by country and time to build a globally representative climate sentiment index. Covering 19 countries with regional representativeness and substantial carbon-emission relevance, the index exhibits long time spans and continuous high-frequency characteristics, enabling fine-grained depiction of cross-country dynamics in climate sentiment expressed on Internet platforms. This provides high-frequency, broad-coverage, large-scale image data with an emotional-semantic basis to support subsequent analyses of how climate sentiment influences policy expectations, market behavior, and risk perception.

The contributions of this study are reflected in several dimensions. First, from a methodological perspective, it is the first to employ visual information for measuring climate sentiment, extending prior research that has relied almost exclusively on textual data. By identifying three categories of climate-related sentiment—positive, neutral, and negative—from Google Images, this study enhances sentiment-recognition capability and strengthens measurement robustness across linguistic contexts. Second, in terms of spatial coverage, we construct a cross-country comparable climate sentiment database encompassing 19 major countries. Through a unified climate keywords dictionary and standardized image-processing procedures, the study produces sentiment indices with consistent definitions across multiple countries and languages, overcoming the limitations of earlier work focused on single regions or language groups and enabling comparative research on climate risks at the global scale. Third, along the temporal dimension, the study develops long-horizon, multi-frequency sentiment series. Drawing on nearly ten years of historical data, we construct climate sentiment indices at daily, weekly, monthly, and annual frequencies. These indices not only capture long-term trends in climate sentiment but also help identify its dynamic effects on financial markets across different temporal scales, providing richer analytical perspectives for understanding the complex relationships between climate sentiment and economic or societal variables.

Methods

This study aims to construct a global climate sentiment index dataset covering the period from 2015 to 2024 in order to present the characteristics and variations of global climate sentiment in terms of positive-negative orientation, temporal patterns, and spatial distribution. The construction of the index involves five major stages: data collection, data cleaning, manual annotation, model training and application, and index aggregation. The overall framework is shown in Fig. 1. First, we retrieved and collected images from the Google Images search engine using search queries generated from a climate keywords dictionary. Second, we performed data cleaning and quality filtering on the raw images and, based on manual annotation, selected samples that met the consistency criteria for training the deep learning model. Finally, we used the trained deep learning model to classify the images and aggregated the results by country and time to generate the climate sentiment index.

Fig. 1.

Research Framework.

Data collection

This study employs Google Images as the primary source for image data collection. The sampling period is set from January 1, 2015, to December 31, 2024, covering key stages in global climate governance since the adoption of the Paris Agreement. Regarding country selection, this study covers 19 countries within the G20 that hold regional representativeness and substantial carbon-emission influence, including China, the United States, the United Kingdom, Australia, Brazil, Canada, France, Germany, Italy, India, Indonesia, Japan, South Korea, Mexico, Russia, Saudi Arabia, South Africa, Argentina, and Turkey.

Drawing on existing literature and authoritative reports^12,22, we constructed a keyword dictionary that comprehensively covers the core topics of climate issues. The dictionary contains 73 keywords in total (the complete list is provided in Figshare⁴⁴) and encompasses five major dimensions. The first is Climate Change and Carbon Emissions, which includes fundamental concepts such as global warming and greenhouse gas emissions. The second is Climate Policy and Action, which covers the names of international organizations and agreements related to climate change, as well as terms associated with regional climate policies. The third is Energy Transition, including terms related to renewable energy and fossil energy. The fourth is Extreme Weather and Disasters, consisting of extreme weather events and natural disasters that are closely linked to climate change. The fifth is Climate Finance and Investment, which includes terms associated with green financial instruments and corporate disclosure.

During the data collection process, we construct retrieval strategies based on the developed keyword dictionary and perform country-by-country and day-by-day searches in Google Images. To ensure compliance with copyright requirements, the “Usage rights” filter provided by Google Images is applied during retrieval to restrict results to images labelled for reuse (e.g., Creative Commons licences). All available images were captured for each country-date pair within our search scope. Through this strategy, we obtain a total of 3,194,289 publicly accessible images within the specified period, amounting to approximately 25 GB of data. This yields a large-scale, diverse, and temporally continuous visual dataset, providing a solid foundation for visual content analysis and the identification of public climate sentiment.

Data cleaning

To ensure data quality and the effectiveness of subsequent sentiment identification, we conducted data cleaning and quality filtering on the original image set. All downloaded images were subjected to readability checks, during which corrupted files, images that could not be opened, and samples with abnormal formats (e.g., non-standard JPG/PNG files or empty files containing only metadata) were removed. We then excluded extreme samples with excessively high or low pixel resolutions, as well as monochromatic images without meaningful visual content, in order to avoid potential distortions in model outcomes.

Following common practice and to meet the input requirements of deep learning models, all qualified images were resized proportionally by setting the shorter side to 256 pixels and subsequently center-cropped to a fixed size of 224 × 224 pixels. This procedure ensures uniform image specifications while preserving the core visual content of each image.

Sentiment classification

Drawing on insights from prior studies and reports, we classify climate-related images into three primary sentiment categories^45–48—positive, neutral, and negative—so as to characterize the emotional orientation embedded in public visual information and the attitudes it conveys toward climate issues. To further enhance classification rigor, we additionally include a separate “non-climate-related” category, which allows us to filter out illustrations on webpages that are unrelated to climate change and prevent them from interfering with subsequent sentiment indices calculations. Specifically, positive images capture optimistic or constructive emotional responses to climate action and often depict scenes such as renewable energy facilities, green technological innovations, ecological restoration outcomes, or international climate cooperation. These images convey a tone of optimism, progress, and constructive engagement. Neutral images, though relevant to climate themes, do not exhibit clear emotional direction; they mainly present factual or contextual information, such as meeting records, scientific processes, data charts, or general natural landscapes. Such images provide background or factual context without eliciting strong emotional reactions. Negative images highlight the adverse impacts of climate change, portraying scenes involving extreme weather events, wildfires, glacier melting, ecological degradation, or infrastructure damage. These images convey warning signals and a sense of crisis, often triggering public concern and risk awareness. Given that webpages containing climate-related content may also embed images unrelated to climate issues, we include the non-climate-related category to capture visual content without substantive relevance to climate topics, such as everyday life scenes, commercial advertisements, or portraits of individuals. Examples of image classification categories are shown in Step 3 of Fig. 1.

To obtain sentiment classifications for individual images, this study employs the ConvNeXt model⁴⁹ from the deep learning domain to perform four-class prediction. The ConvNeXt architecture incorporates key design principles from the Vision Transformer (ViT) family and modernizes the architecture of traditional CNNs. This includes the use of larger convolutional kernels, deeper network structures, and more contemporary components such as LayerNorm, enabling it to achieve performance comparable to or exceeding that of ViT models while retaining the computational efficiency of convolutional networks. ConvNeXt-Tiny is the lightweight variant within the ConvNeXt family, featuring relatively low computational overhead and making it suitable for training and deployment under limited computational resources and large-scale image datasets. In this study, we adopt ConvNeXt-Tiny for the classification task. It was initially pre-trained on the ImageNet-1K dataset, which contains 1,281,167 images, giving it strong generalization capability. However, the task of recognizing sentiment in climate-related images differs from general image recognition and therefore cannot be accomplished directly using the pre-trained model. To adapt it to our image classification task, we fine-tune the model through transfer learning. Transfer learning allows effective utilization of the domain knowledge embedded in a pre-trained model. By modifying only the final fully connected layer and fine-tuning the parameters using a relatively small set of labeled samples, the model can achieve strong task-specific performance.

To this end, we randomly sampled 10,000 images from the full dataset and recruited 8 graduate students specializing in climate change economics to annotate their sentiment categories. Each image could be assigned to only one sentiment class. To further improve the quality of the training samples, each image was independently annotated by 4 annotators, followed by an internal consistency validation in which only samples with complete agreement among all annotators were retained. Ultimately, we obtained 4,912 images with perfectly consistent labels to construct the fine-tuning dataset, ensuring the reliability of the annotations. The fine-tuning dataset was then divided into training, validation, and test sets using a 7:1.5:1.5 ratio.

During the model fine-tuning stage, we first replaced the original classification head with a new classifier containing four sentiment categories. We then adopted a two-stage fine-tuning strategy. In the first stage, all network parameters except the classification head were frozen, and the model was trained for a small number of epochs to quickly adapt to the structural requirements of the target task. In the second stage, all parameters were unfrozen, and the model was trained for multiple epochs using the AdamW optimizer with a relatively low learning rate. An early-stopping mechanism was applied such that training automatically terminated when the classification accuracy on the validation set failed to improve for 5 consecutive epochs, preventing overfitting and enhancing generalization. The fine-tuned ConvNeXt-Tiny model outputs a sentiment classification for each image, and we apply it to classify the sentiment of the full dataset.

Indicator construction

Given that positive and negative sentiments are not necessarily inversely related—they may move in opposite directions or increase simultaneously—and that each may affect investor sentiment differently, this study constructs two separate indices based on the image sentiment classification results: a Climate Positive Sentiment Index and a Climate Negative Sentiment Index.

Climate Positive Sentiment Index (CPSI)

We aggregate the image-based sentiment classification results by country and time to construct the CPSI. This index reflects the societal level of positive climate sentiment by measuring the proportion of positive-sentiment images among all climate-related images. It offers advantages such as intuitive interpretation, strong computability, and ease of cross-regional comparison, and it can be produced at daily, weekly, monthly, and annual frequencies. For country $i$ in period $t$, the CPSI is defined as:

$${\mathit{CPSI}}_{i,t}=\frac{N_{i,t}^{\mathit{pos}}}{N_{i,t}^{\mathit{pos}}+N_{i,t}^{\mathit{neu}}+N_{i,t}^{\mathit{neg}}}$$ 1

where $N_{i,t}^{\mathit{pos}}$, $N_{i,t}^{\mathit{neu}}$ and $N_{i,t}^{\mathit{neg}}$ represent the numbers of images classified as positive, neutral, and negative, respectively, for country $i$ in period $t$.

To capture the overall global sentiment trend, this study further constructs a global-level CPSI. Given that the 19 selected countries exhibit strong regional representativeness and substantial influence on carbon emissions, we combine their samples and treat them as a proxy dataset for global climate sentiment. For the global level, the CPSI in period $t$ is defined as:

$${\mathit{CPSI}}_{\mathit{golbal},t}=\frac{N_{\mathit{golbal},t}^{\mathit{pos}}}{N_{\mathit{golbal},t}^{\mathit{pos}}+N_{\mathit{golbal},t}^{\mathit{neu}}+N_{\mathit{golbal},t}^{\mathit{neg}}}$$ 2

where $N_{\mathit{golbal},t}^{\mathit{pos}}$, $N_{\mathit{golbal},t}^{\mathit{neu}}$ and $N_{\mathit{golbal},t}^{\mathit{neg}}$ represent the numbers of images classified as positive, neutral, and negative, respectively, at the global level in period $t$.

Climate Negative Sentiment Index (CNSI)

The same method is applied to construct the Climate Negative Sentiment Index, ensuring comparability between the CPSI and the CNSI. For country $i$ in period $t$, the CNSI is defined as:

$${\mathit{CNSI}}_{i,t}=\frac{N_{i,t}^{\mathit{neg}}}{N_{i,t}^{\mathit{pos}}+N_{i,t}^{\mathit{neu}}+N_{i,t}^{\mathit{neg}}}$$ 3

For the global level, the Climate Negative Sentiment Index in period $t$ is defined as:

$${\mathit{CNSI}}_{\mathit{golbal},t}=\frac{N_{\mathit{golbal},t}^{\mathit{neg}}}{N_{\mathit{golbal},t}^{\mathit{pos}}+N_{\mathit{golbal},t}^{\mathit{neu}}+N_{\mathit{golbal},t}^{\mathit{neg}}}$$ 4

Data Records

The Climate Sentiment Index dataset is publicly available on Figshare⁴⁴. The repository contains 4 projects, with the detailed file list provided in Table 1. Figure 2 presents the monthly global and national-level climate sentiment indices for the 19 countries, while statistical plots for other frequency indices are detailed in the CSI_dataset folder of the Figshare repository⁴⁴. The image metadata files, which include the URLs of the original images underlying the index construction, are stored in the source_code/data directory of the Figshare repository⁴⁴.

• Global-level Climate Positive Sentiment Index: This project provides the global image-based Climate Positive Sentiment Index from 2015 to 2024, including daily, weekly, monthly, and annual indices. The index is provided as four separate CSV files corresponding to the four temporal resolutions on Figshare⁴⁴.

• Global-level Climate Negative Sentiment Index: This project provides the global image-based Climate Negative Sentiment Index from 2015 to 2024, including daily, weekly, monthly, and annual indices. The index is provided as four separate CSV files corresponding to the four temporal resolutions on Figshare⁴⁴.

• National-level Climate Positive Sentiment Index: This project provides the image-based Climate Positive Sentiment Index for 19 countries from 2015 to 2024, including daily, weekly, monthly, and annual indices. The index is provided as four separate CSV files corresponding to the four temporal resolutions on Figshare⁴⁴.

• National-level Climate Negative Sentiment Index: This project provides the image-based Climate Negative Sentiment Index for 19 countries from 2015 to 2024, including daily, weekly, monthly, and annual indices. The index is provided as four separate CSV files corresponding to the four temporal resolutions on Figshare⁴⁴.

Table 1.

Summary of the dataset.

Level	Item	Frequency	Filename
Global	Global climate positive sentiment index	Daily	global_cpsi_daily.csv
		Weekly	global_cpsi_weekly.csv
		Monthly	global_cpsi_monthly.csv
		Annual	global_cpsi_annual.csv
	Global climate negative sentiment index	Daily	global_cnsi_daily.csv
		Weekly	global_cnsi_weekly.csv
		Monthly	global_cnsi_monthly.csv
		Annual	global_cnsi_annual.csv
National (19 Countries)	Climate positive sentiment index of 19 representative countries	Daily	national_cpsi_daily.csv
		Weekly	national_cpsi_weekly.csv
		Monthly	national_cpsi_monthly.csv
		Annual	national_cpsi_annual.csv
	Climate negative sentiment index of 19 representative countries	Daily	national_cnsi_daily.csv
		Weekly	national_cnsi_weekly.csv
		Monthly	national_cnsi_monthly.csv
		Annual	national_cnsi_annual.csv

Fig. 2.

Monthly Climate Sentiment Indices.

Technical Validation

The technical validation of this study consists of 3 components: comparison with manual verification, comparison with existing datasets, and limitations analysis. This section demonstrates the validity and advantages of the index from multiple perspectives.

Comparison with manual verification

We randomly selected 15% of the annotated images that passed the internal consistency validation—739 images in total—as the test set. This test set was excluded from all training procedures and used solely for evaluating model performance. The evaluation metrics include precision, recall, and f1 score. Precision refers to the proportion of correctly classified samples within all samples predicted to belong to a given category. Recall refers to the proportion of correctly classified samples within all samples that actually belong to that category. The f1 score, defined as the harmonic mean of precision and recall, provides a combined assessment of the model’s accuracy and recall performance. The results indicate that the deep learning method achieves a level of performance broadly consistent with manual annotation, as detailed in Table 2.

Table 2.

Performance Comparison of Three Coders’ Verification Results and Models’ Predictions on the Test Dataset.

	Precision	Recall	F1-score	Support
Positive	0.8796	0.8120	0.8444	117
Neutral	0.8477	0.8312	0.8393	154
Negative	0.8533	0.9078	0.8797	141
Non-Climate-Related	0.9636	0.9725	0.9680	327
Accuracy	—	—	0.9053	739
Macro Average	0.8861	0.8809	0.8829	739
Weighted Average	0.9051	0.9053	0.9048	739

Comparison with existing datasets

Due to the lack of publicly available image-based climate sentiment datasets, direct comparison with existing datasets is not feasible. This study is among the earliest to analyze climate sentiment using image data. Existing datasets are derived from social media, news reports, and search indices. Table 3 summarizes these datasets and provides detailed information on their coverage areas, highest frequencies, and years.

Table 3.

The Existing Dataset Overview.

Data Sources	Database Name	Geographical Coverage	Maximum Frequency	Coverage Period	Authors	Data Availability
Images	Climate Sentiment Index	Worldwide	Daily	2015–2024	Our research	—
Social media	Twitter Climate Sentiment	UK & Spain	Daily	2019	Loureiro and Alló50	NO
	Investor Climate Sentiment	US	Monthly	2010–2019	Santi9	YES
News reports	WSJ Climate News Index	Worldwide	Monthly	1984–2017	Engle, et al.21	YES
	Reuters Climate Risks Index	US	Daily	2000–2018	Faccini, et al.22	YES
	Carbon Credit Sentiment Index	Worldwide	Daily	2014–2021	Rogmann and Schreiber51	NO
	Climate Policy Uncertainty	US	Monthly	1985–2025	Gavriilidis23	YES
	Global Climate Policy Uncertainty	Worldwide	Daily	2000–2025	Ma, et al.54	YES
Search indices	Global Warming Google Index	Worldwide	Monthly	2004–2017	Choi, et al.52	YES
	Weather Risk Fear Index	Worldwide	Weekly	2004–2018	Fernandez-Perez, et al.53	NO

Currently, research on climate sentiment is limited, and the methodologies employed are relatively simple. Some scholars have extracted public sentiment toward climate-related topics from social media texts using methods such as word frequency analysis. Loureiro and Alló⁵⁰ collected climate change-related content from Twitter in the United Kingdom and Spain and used the NRC Emotion Lexicon to perform word frequency analysis to construct an index. Santi⁹ calculated sentiment scores for climate-related posts on the StockTwits platform using R language packages, using these scores as indicators of climate sentiment. Other scholars have constructed indices based on climate content in news reports. Engle, et al.²¹ developed a climate change news index based on historical Wall Street Journal articles using the TF-IDF method. By analyzing climate change news texts from 2000 to 2018, Faccini, et al.²² employed the LDA topic model to construct four risk indices corresponding to different themes: natural disasters, global warming, climate policy, and international summits. Web search indices also reflect climate sentiment and perception. Rogmann and Schreiber⁵¹ constructed a carbon-credit sentiment index by extracting daily media and social-media emotions linked to the term ‘carbon credit’ from MarketPsych’s global text-analytics database. Choi, et al.⁵² used Google search volumes for “global warming” to represent public attention to climate change. Fernandez-Perez, et al.⁵³ examined the role of weather-related risk aversion sentiment in commodity futures pricing through Internet searches based on 149 search terms.

Based on this, the present study introduces “images” as a previously overlooked key source of information for constructing climate sentiment. Compared with the high noise and subjective bias commonly found in social media and news texts, images offer a more direct and stable reflection of public emotional responses to climate-related events, thereby reducing the uncertainty associated with textual sources. Meanwhile, this study expands climate perception beyond traditional attention-based indicators to include both positive and negative sentiment dimensions, substantially enriching the characterization of climate sentiment. In the future, this index dataset offers considerable potential for further research.

Limitations and future work

Although the constructed climate sentiment indices demonstrate robustness and validity in testing and can effectively capture the dynamic changes in public sentiment toward climate issues, several limitations of this study merit attention.

First, during the data collection phase, as the narrative of climate discussions continuously evolves, the keyword dictionary constructed based on existing literature inevitably suffers from a lag in semantic evolution, potentially failing to fully encompass certain emerging specific technical terms.

Second, during the model training phase, to establish a low-noise, clean baseline for the ConvNeXt model, this study retained only those images with unanimous annotations. While this screening strategy facilitates the reliable extraction of sentiment shocks capable of triggering societal resonance, it also introduces a risk of selection bias, potentially overlooking visually ambiguous yet economically relevant content.

Third, the sentiment indices in this study are based on all images identified as climate-related, without further fine-grained thematic classification. This means that, in the current version, the heterogeneity of sentiment across different types of climate topics—such as policy negotiations, extreme weather events, technological innovations, and corporate emission reductions—is not explicitly distinguished. For example, news related to climate disasters is often accompanied by clearly negative sentiment, whereas reports on technological breakthroughs or the development of green industries may carry a more positive tone. Future work could consider incorporating topic modeling or event classification at the semantic level, pairing sentiment changes with specific climate subtopics, to more accurately reveal emotional responses in different contexts and their impact pathways on financial markets.

Last, this study currently relies on a single modality—images—to determine sentiment for each data category, without integrating textual and visual information. Future work could consider employing advanced multimodal sentiment analysis models to achieve joint analysis, thereby further improving the accuracy of sentiment recognition and better reflecting actual public perception.

Considering the potential directions for extension mentioned above, although the current climate sentiment index already demonstrates considerable practicality and generalizability, there remains significant room for improvement in thematic precision and information integration. Our future research will continue to explore model refinement and data structure optimization to further enhance the application value of the sentiment index in green finance and climate communication studies.

Usage Notes

This dataset provides a novel foundation for behavioral finance research related to climate change and has broad applicability in areas such as asset pricing, risk management, and policy evaluation.

First, the CPSI and CNSI, as two complementary climate sentiment factors, can be applied to the pricing of financial assets such as stocks and futures. For quantitative investors, they can serve as systemic state variables in multi-factor asset pricing models, capturing changes in market expectations of climate transition opportunities and climate-related risks. This enables the assessment of whether exposure to positive and negative sentiment is priced by the market, thereby identifying differences in expected returns between green and high-carbon assets and potential climate sentiment premia. For regulators, long-term monitoring of the relationship between these two types of climate sentiment and asset prices can help detect early signs of potential bubbles and support the dynamic adjustment of macroprudential policies.

Second, the climate sentiment indices can contribute to the assessment of systemic financial risk. As climate-related risks are increasingly incorporated into financial regulatory agendas, market sentiment has become an important non-traditional factor affecting system stability. In stress testing and scenario analysis, using historical sentiment extremes as sentiment shocks can help construct extreme scenarios that more closely reflect actual market reactions. In cross-market and cross-sector risk-transmission models, the indices can also serve as exogenous shock sources, enabling the identification of how and at what speed climate sentiment propagates within the financial system through channels such as capital flows, risk reassessment, or policy-expectation adjustments.

Finally, the climate sentiment indices can function as quantitative tools for evaluating policy sentiment effects. On one hand, aligning the timing of major climate policies—such as the introduction of the EU Carbon Border Adjustment Mechanism, the passage of the U.S. Inflation Reduction Act, or the signing of the Paris Agreement—with movements in the sentiment indices allows for the assessment of whether these policies effectively enhanced positive sentiment, mitigated risk sentiment, and whether such sentiment responses exhibit cross-country heterogeneity. On the other hand, the indices can reveal sentiment-based shortcomings in policy transmission or communication, thereby providing evidence for improving policy design and enhancing policy acceptance.

In sum, the climate sentiment indices developed in this study not only fill an important gap in climate public-opinion data but also offer a unified and flexible analytical tool for interdisciplinary exploration of societal responses to climate change.

Author contributions

X.Z.: Research design, data processing, experimental implementation, manuscript writing. Y.Z.: Data collection, results analysis, figure preparation, manuscript writing. S.Z.: Data collection, figure refinement, manuscript revision. K.G.: Supervision, results analysis, scientific discussion. D.Z.: Supervision, results analysis, scientific discussion, funding acquisition. Q.J.: Supervision, results analysis, manuscript revision, scientific discussion, funding acquisition.

Data availability

The complete dataset is available in the Figshare repository at https://doi.org/10.6084/m9.figshare.30689201. All files are distributed under a CC-BY 4.0 license, which permits unrestricted use, distribution, and reproduction in any medium, provided that appropriate credit is given to the original authors and the source.

Code availability

The codes that were used to generate the image-based climate sentiment indices are available in Figshare⁴⁴.

Competing interests

The authors declare no competing interests.