Case study on climate change effects and food security in Southeast Asia – Scientific Reports

The agricultural productivity of croplands is influenced by a combination of climatic and anthropogenic factors, including temperature, precipitation, pesticide/herbicide application, pollution, fertilizer usage, pH regulation, tillage practices, and others^33,34,35. Given the complexity and interdependence of these elements, our study adopts a holistic approach by focusing on the primary determinants of agricultural output: climate and fertilizers. By developing models to examine the cropland dynamics and project changes in fertilizer consumption, we aim to predict shifts in rice production, taking into account potential shifts in arable lands, yield variations and fertilizers usage. The methodology of our research is illustrated in Fig. 3, which also outlines the data sources we have utilized. Detailed information on the datasets can be found in the “Data and preprocessing” section.

Over the past 150 years, the application of fertilizers has significantly enhanced crop yields³⁶. Figure 4 demonstrates the agricultural consumption of three major types of fertilizers in the years 1966–2021, along with projections for the next decade, based on our autoregression model (detailed in the “Rice yield model” section). The data reveal distinct consumption patterns for each country, reflecting varying fertilizer usage dynamics. These patterns suggest that appropriate fertilizer use could potentially mitigate the negative effects of climate change on crop production.

Historical and projected consumption of three primary fertilizers types: (a) Nitrogen, (b) Phosphorus, (c) Potassium. The solid line represents consumption data from 1966 to 2021, while the dashed line indicates projections from 2022 to 2031, measured per unit of arable area (This figure was created in Python 3.10 and the Jupyter Notebook programming interface).

Our research aims to account for the influence of climate on cropland status, incorporating social factors for a more comprehensive analysis. In actual agricultural practice, lands suitable for crop cultivation are selected based on their potential, either as a conscious decision of farmers or processes analogously to natural selection, which identifies the most suitable combinations of land characteristics and other variables. This makes cropland suitability dependent on both social and climate conditions. Therefore, we consider past agricultural land as a predictive factor together with climate conditions to capture complex interactions between socioeconomic and climatic patterns.

Cropland suitability

To model cropland suitability, we integrated climate and elevation data to develop a range of machine learning models, aiming to effectively incorporate social factors. These models range from conventional algorithms to more advanced neural network architectures, including Multilayer Perceptrons (MLP) and Convolutional Neural Networks (CNN). CNNs have garnered significant attention in climate and weather forecasting owing to the spatial nature of the data^25,37,38. Recent research extends beyond these methods to consider recurrent neural networks and deep neural networks for capturing intricate spatio-temporal relationships in the data^19,39,40,41 for regression problems in weather forecasting and climate modeling. Additionally, we used bioclimatic variables, which are climate indices updated annually, to analyze trends in temperature and precipitation. These variables offer insights into the current climate conditions (for more details, see “Data and preprocessing” section). These indexes are crucial for identifying and understanding climatic patterns.

In order to conduct a fair evaluation of the models, it is essential to select appropriate metrics for the task in question. Balanced accuracy is a performance metric that measures the percentage of correct predictions with respect to the share of each class, making it particularly useful when dealing with imbalanced classes where one class is underrepresented compared to the other. In this study, we use balanced accuracy to evaluate the performance of our classifier in distinguishing between the presence and absence of crops (denoted as class 1 and class 0, respectively). We estimate the precision (the tendency not to predict false croplands) and recall (the ability not to predict false non-croplands) using the optimal threshold based on maximizing the F-Measure—the harmonic mean of precision and recall. Among the models tested, the XGBClassifier⁴² outperforms its counterparts (see Table 1). While slightly behind CNN, CatBoost, and Random Forest, this model has the highest Balanced accuracy, Recall, and ROC-AUC values, indicating its superiority in cropland modeling based solely on climate conditions.

Having identified the XGBClassifier as the top-performing model for cropland suitability classification, we trained it on both climate conditions and social factors. For the social factors, we considered land usage over the 7 years preceding the prediction date. This period reflects the socioeconomic influences on agricultural land use. A feature importance analysis conducted using the SHAP tool, which applies Shapley values from game theory to explain model outputs⁴³, confirms that prior land use is the major factor contributing to the superior performance of the resulting model (Fig. 5).

Feature importance evaluated using the SHAP tool for models with different feature sets. Features: lc—land class 7 years prior; bio1 to bio12—see notations in Table 5. (a) XGBClassifier via climate features, (b) XGBClassifier via climate features and previous land usage (This figure was created in Python 3.10 and the Jupyter Notebook programming interface).

These findings highlight the importance of considering past agricultural land use when predicting the future status of croplands. This observation can be interpreted from various perspectives that are not mutually exclusive. First, it suggests that landowners may rely heavily on traditional farming methods instead of adapting new techniques that account for shifts in climate conditions. Second, it indicates that landowners might consider local factors not captured in our climatic and agricultural datasets. Third, landowners employing effective practices, whether based on empirical evidence or not, may benefit from a positive feedback loop, gaining better access to resources like fertilizers or financial assets. Overall, this feature highlights the critical socioeconomic conditions influencing cropland usage.

Projected changes in cropland area by 2028 relative to 2021, forecasted by the climate model for (a) Cambodia, (b) Lao PDR, (c) Thailand, (d) Philippines, (e) Viet Nam, (f) Myanmar. The horizontal axis represents longitude and the vertical axis latitude. (Maps were created using Rasterio version 1.3.9⁵¹ and Python version 3.10).

Variations of land arability and rice production due to the climate change

Country-level agricultural rice production depends on the total area dedicated to arable lands and fertilizer use. Thus, we model potential changes in cropland area pixel-wise for Southeast Asia and fertilizer usage for each country.

Climate change leads to variations in cropland suitability, which we model using XGBClassifier, considering socioeconomic and climate change factors. Figure 6 shows the map with marked pixels having a high probability of crop status changes in the year 2028 compared to the year 2021, according to our model. Green color highlights the pixels with arable lands potential to expand, while red color—those with the potential to degrade.

In Cambodia, one can observe that the area near the Mekong River at the intersection of Kampong Cham, Kampong Thom, and Kratié provinces is to experience the most intense risk for arable lands in 2028, further, are to the South-East from the Tonle Sap Lake is under moderate risk. However, the North-Easten half of the country becomes potentially opportune for arability. Moreover, South-Western regions, such as areas near Kampong Chhnang city and Pursat province that are close to Lake Tonlé Sap and areas in the Aural District, West of the capital Phnom Penh, close to the Cardamom Mountains, are also potentially favorable for arability.

Considering Lao PDR, the proximities of the Mekong River near Nen Ngam Reservoir are projected to be more suitable for arability with minor exclusions. Overall, the southern and Western parts of the country demonstrate moderate potential for arability.

Khorat Plateau in Thailand is projected to be under moderate arability risk, along with areas near Ping River in the middle part of the country. Moreover, arable lands in the Chachoengsao district, on the border with Cambodia, and some lands in the North near Myanmar are at high risk. However, North-Easter lands near Lao PDR are demonstrating high potential prior to 2028.

In the Philippines, Mindanao islands are projected to have a favorable environment for arable lands. However, the central parts of Luzon islands are considered to be under moderate risk with minor high-risk fields.

Viet Nam’s arable lands are projected to be among the most at risk in Southeast Asia. Arable lands around Hanoi, South-West towards Thanh Hoa, and down the Ca river near Vinh city are under arability loss risk in 2028. Also, Tay Ninh, An Giang, Kien Giang provinces, Tan Hung and Vinh Hung districts are under high cropland suitability loss risks, together with areas near Song Ray Lake and areas South of Can Tho city. On the other hand, areas North of Hanoi and a few areas on the intersection of Phu Yen, Dak Lak, and Bihn Dihn provinces show some potential for new arable lands.

Finally, Myanmar has no significant loss of arable lands. It has moderately risky areas in Ayeyarwady, Bago, and Magway provinces and arability-gain areas dispersed among the Sagaing, Kachin, and Shan provinces. On the other hand, the vast majority of the country has moderate potential for arability gain in the Eastern and Southeast provinces.

Table 2 demonstrates the results of rice yield modeling utilizing the cropland suitability model and fertilizer usage. See section “Rice yield model” for the modeling details. Here, yield is estimated per modeling grid cell of \(49 \text { km}^2\) area. A negative yield or production percentage value indicates a decrease, whereas a positive value represents an increase. The combined yield model shows \(R^2 = 0.97\) and mean absolute percentage error \(MAPE=4.2\%\) on test data.

Analyzing the results indicated in Table 2, we emphasize that under a negative scenario of no action to reach for utilizing potentially arable lands (Overall Croplands, bold column), Viet Nam, Philippines, Philippines, Lao PDR and Indonesia are expected to loss significant number of total arable lands, meanwhile Thailand, Myanmar and Cambodia are expected to have moderate losses. On the other hand, having reached for potential lands, all countries may not even mitigate severe losses but increase the total area of arable lands, except Indonesia (Overall Croplands, parenthesized column). A similar picture is for paddy rice fields only (Paddy Rice Croplands column).

Due to the climatic conditions in 2028 and forecasts of fertilizers usage (see Fig. 4), per area yield is considered to fall for each country, except for the Philippines. Note that for Cambodia, Myanmar, and Viet Nam, it is larger than \(10\%\). Despite the fact that some countries like Myanmar, Thailand, and Cambodia are not to experience a dramatic drop in the total area of croplands and paddy rice fields, they are to experience a yield drop. This is caused by a drop in fertilizer usage for these countries; see Fig. 4. For Cambodia there is a low usage of Potassium, for Myanmar it is both Potassium and Phosphorus and for Thailand it is Nitrogen, Phosphorus and Potassium.

Next, we estimate the total Production Change for the countries’ areas in the last column. According to our Cropland Suitability XGBClassifier, in a negative course, i.e., no reach for potentially arable lands and loss of current ones, the overall picture is similar or more dramatic for all countries. However, if some countries, e.g., Lao PDR, Thailand, and Malaysia, reach for potentially arable lands, they might not just mitigate losses but even increase their production significantly.

To sum up, this study focuses on identifying potential risks rather than proposing development strategies. According to our findings, Cambodia and Viet Nam face severe threats in rice production, while the Philippines is expected to experience growth. Moreover, if countries take an opportunity to utilize potential paddy rice croplands, they might mitigate production drop risks and even increase their rice output. These findings highlight potential risks and emerging opportunities for policymakers that define agricultural strategy.

Comparative analysis with existing research

Comparing our study’s results with those of existing research, statistical analyses, and projections provides valuable insights. We categorize these comparisons into two main areas: machine/deep learning modeling of cropland suitability and rice-related projections.

In the field of rice yield projections, the IPCC Sixth Assessment report provides⁵² projections of agricultural production up to 2040 and 2080 for Southeast Asia countries. This report indicates similar results to those of our research for Thailand, Cambodia, and the Philippines, which support the validity of our research and align it with credible IPCC organizations. However, our research provides a more detailed analysis by countries and a closer projection timeline.

Moreover, research⁵³ evaluates production potential, net exports, and yield gap projections by 2040 in Southeast Asia. The authors took into account current crop management methods via the results of questionnaires of agronomists collected by countries. Yield potential projection was conducted using ORYZA v3. This is a plant growth simulation software that is a valuable tool for understanding the genetics of a specific rice plant and improving local agricultural practices, but it is not built for country-scale analysis. Furthermore, their analysis of yield gaps and rice demand-supply relied on data from local reference weather stations. Authors in⁵⁴ developed a probabilistic framework for predicting the Vegetation Health Index (VHI) up to 3 months in advance, using a Quantile Random Forest model that correlates VHI with rice price shocks⁵⁴. Our approach extends this by linking predicted cropland suitability directly to rice yield, providing a more direct connection to cropland status than just economic indicators like price shocks.

Regarding machine and deep learning, studies^23,25 propose neural network frameworks for predicting yields of rice, soybean, and corn, leveraging recurrent and convolutional architectures. Authors consider the USA corn belt and China, respectively, which are already well-studied areas. These studies focus on the performance and analysis of neural networks tested solely, without applying them further to obtain insightful results for practical applications.

In contrast to the studies and reports above, our study, first, allows for a more transparent and focused modeling methodology and tools. Specifically, by using CMIP5 climate projections, analysis of different predictor spaces—bioclimatic variables with or without socioeconomic factors and testing a wider range of learning models. Second, we build cropland suitability prediction and use it to project rice production up to 8 years ahead. Finally, our study includes risk analysis for rice production in Southeast Asia, which is understudied in the research field.