Flood impacts on global crop production: advances and limitations

We used a 0.5^∘-resolution global dataset of historical yields for 1981–2016 (GDHY, Iizumi and Sakai 2020). Annual time series of yield Y (metric tons per hectare, t ha⁻¹) generally have a temporal trend T (t ha⁻¹) which represents yield increases associated with improved technology and management. To include the association with floods, we calculated yield change Yʹ—the ratio relative to the yield trend—in which T was removed from the yield time series as

$$\begin{equation} Y^{^{\prime}}_{i,j,c}=\frac{1}{n}\sum_{k=1}^{n}\frac{Y_{i,j,k}-T_{i,j,k}}{T_{i,j,k}}, \end{equation} \tag{ 1 }$$

where i, j, and k are indices of year, grid cell, and cropping season of a crop c, respectively. The n is the total number of the cropping seasons—the main season and the second season for maize and rice, the spring and the winter for wheat, and just one major season for soy (Sacks et al 2010). The time series of $T_{i,j,k}$ was estimated using a locally weighted regression model (Lu et al 2017) set to the smoother span f = 0.14, corresponding to using a 5-year interval for smoothing of 35-year data. We note here the limitations of crop data inherited from GDHY, which were model estimates and not free from uncertainties due to imperfect modeling, inaccurate inputs, misreporting in agricultural census statistics, and the use of time-constant information in 2000, such as the crop calendar (Sacks et al 2010) and the harvested area (Monfreda et al 2008) used for GDHY. It is particularly important to improve the gaps in crop distributions in the near future, especially when compared to the study by Ray et al (2012), Kim et al (2021), and Grogan et al (2022).

We calculated the return period R with inundation area fraction I during 1951–2016. The daily I data were obtained from the Global Drought and Flood Catalogue (He et al 2020 https://prep-next.github.io/data/GDFC/index.html)—the I was the simulation output of the global river and inundation model CaMa-Flood (Yamazaki et al 2011), with soil moisture and runoff simulated by the enhanced variable infiltration capacity (Liang and Lettenmaier 1994) land surface model with the enhanced global meteorological forcings (He et al 2020). Next, the daily I was spatially aggregated from the original grid size of 0.25^∘ to 0.5^∘, to be consistent with the grid size of the crop yield maps presented in the previous section. The harvested crop area A in 2000 (Monfreda et al 2008) was used as the weight in this spatial aggregation: $I^{0.5^{\circ}} = \left(\sum I^{0.25^{\circ}}A^{0.25^{\circ}}\right)/\left(\sum A^{0.25^{\circ}}\right)$. Then the maximum value of daily I for the 90 day interval d just before the harvest day H—which was obtained from the global crop calendars (Sacks et al 2010)—was given for a yearly I as

$$\begin{equation} I_{i,j,c}=\frac{1}{n}\sum_{k=1}^{n}\max_{1\leqslant d\leqslant 90}I_{H_{j,k}-d}. \end{equation} \tag{ 2 }$$

Finally, the return period R of I was estimated with the Weibull formula as

$$\begin{equation} R_r=\frac{N-1}{r}, \end{equation} \tag{ 3 }$$

where r is the rank on the flood index, which is arranged in descending order, and N is the total number of years. The R was calculated by year i, grid cell j, and crop c associated with yield change Yʹ.

We calculated each grid cell's mean yield changes for flood years and non-flood years. The years with an inundation area's return period of $\lt10$ years were classified as non-flood years, whereas those with a return period of $\geqslant10$ years were classified as flood years—this threshold of a return period presented the largest significant flood-affected areas among all return periods. When a significant difference in the mean yield change between flood and non-flood years was detected based on Student's two-tailed t test (p < 0.1), the grid cell's yield change was recorded as the flood-induced yield change and was used for the subsequent analysis. The grid cells that showed significant difference were also recorded in order to estimate the flood-affected crop areas. For comparison, the other flood-induced yield changes were estimated using precipitation instead of inundation area. The precipitation data were obtained from the WFDEI meteorological forcing data (Weedon et al 2014)—water and global change (WATCH) forcing data, which was based on European centre for medium-range weather forecasts re-analysis interim data corrected using Global Precipitation Climatology Centre (GPCC) data.

The flood impact on crop production loss $\Pi^-$ in United States (US) dollars ($) was estimated with

$$\begin{equation} \Pi_{c,z}^-= \sum_x\sum_y \min(0,Y^{^{\prime}}_{x,y,c})T_{x,y,c}A_{y,c}P_{c,z}, \end{equation} \tag{ 4 }$$

where A and P are the harvest area (ha) and producer price ($ t⁻¹), respectively. The subscripts x, y, and z are indices of the flood year based on the return period of $\geqslant 10$ years, flood grid cells chosen in the previous section, and country. The A was from Monfreda et al (2008) and P was the average from 1998 to 2002 obtained from the FAO statistical database (www.fao.org/faostat/en/#data/PP). Actually, it is ideal to apply historical A and P datasets for this calculation because $\Pi^-$ was the total loss during the 1982–2016 period. Therefore, these datasets were studied in advance. The other case, the crop production gain $\Pi^+$ is calculated with $\max(0, Y^{^{\prime}}_{x,y,c})$ instead of $\min(0, Y^{^{\prime}}_{x,y,c})$ in equation (4). Lastly, we use the statistically significant Yʹ described in the previous subsection.

Using the ROC curve, annual counts of national flood grid cells were tested with binarized time series of national flood events during 1982–2016. The binary label of the time series as a flood indicator was assigned according to whether the annual count of national flood events was ${\geqslant}1$ or 0 based on the emergency events database (EM-DAT, CRED accessed 7 February 2023). The target score was the annual time series, which consisted of the annual total counts of flood grid cells described in the section 2.3.

The global medians of the annual grid yield changes were estimated to be significant losses of 4% for soy, 3% for rice, 2% for wheat, and 1% for maize at the return period of over ten years (figure 1(A)), while most of the yield changes for return periods under ten years were not losses. The tendency that a longer return period of inundation area leads to a larger yield loss therefore seems reasonable. The differences in yield loss between the crops imply that the crop vulnerability to floods varies with crops' morphological characteristics and growing field environment. The vulnerability related to the former includes that crops with relatively lower heights such as rice, soy, and wheat were more susceptible to water lodging than taller crops, such as maize, though the inundations were of equal depth. Related to the latter, especially rice has traditionally been cultivated in paddies and is sensitive to floods (Bailey-Serres et al 2012): rice fields located along with river channels or in lowlands were exposed to floods more frequently than upland crop fields. This result is consistent with the result that the effect of drought on rice appeared to be relatively small compared to those on other crops (Kim et al 2019).

Zoom In Zoom Out Reset image size

Flood-induced yield gains were detected for maize and soy at a three year return period (green and yellow in figure 1(A)). This result was interpreted as meaning that inundated areas for this return period were too small to inundate the fields of these crops, whereas more water than normal years resulted in yield gains for them. This tendency was shown at return periods shorter than ten years for maize, soy, and wheat that were cultivated in dry fields. Li et al (2019) indicated that excess rainfall in warm states of the US led to a positive maize yield caused by fulfilling water demand and mitigating heat stress. Chen et al (2019) also pointed out the positive response to floods in China under certain conditions where water shortages were alleviated. The same effects would cause our result in flood-induced yield gains for a return period shorter than ten years.

The global total production loss during 1982–2016 reached $5.5B and its gain was $1.9B: $3.1B and $0.6B for rice, $1.1B and $0.6B for maize, $0.8B and $0.4B for wheat, and $0.4B and $0.2B for soy, respectively. Totals of the top ten countries indicated a 83% loss and a 67% gain for these global totals (figure 2(A)). China was the most flood-stricken country with 56% of the global total production loss, whereas Argentina and Brazil were flood-benefited countries in terms of total crop production. FAO reported $20B of cereal loss due to natural disasters for the least-developed and lower middle-income countries in Africa, Asia, and Latin America during 2008–2018 (FAO 2021). Taking about 19% of this loss as having been caused by floods (FAO 2021), $3.8B was the total flood-induced crop loss calculated for these countries. This value is comparable with our result, $1.2B loss during approximately the same period.

Zoom In Zoom Out Reset image size

Most of the annual events by country that were captured in this study as flood events (figure 2(B)) were comparable with those in EM-DAT (The International Disaster Database, https://public.emdat.be/data) and other reports mentioned previously. However, a quantitative relationship was not found between them because the financial damages estimated by the EM-DAT and the reports included the damages of not just the agricultural sector but also other sectors. Noticeably, just a few flood events devastated crops in most of the countries (figure 2(B)).

The global totals of the flood-affected crop areas—the harvest areas in grid cells of which flood-induced yield changes were statistically significant—were 13, 8, 8, and 4 million hectares (Mha) for rice, maize, wheat, and soy, which were 9%, 6%, 4% and 6% of each global harvest area, respectively (figure 3(A)). Of these global affected areas, two-thirds had yield losses for rice, maize, and soy, and half had losses for wheat (dark red). These yield losses were predominantly in the upper Songhua, the lower Yangtze, and the Hongshui in China, the middle and the lower Mississippi in the US, the lower Parana and the Uruguay in Argentina and Uruguay, and the Indus in Pakistan (figure 3(A)). The crop losses in Northeast Asia were conspicuous. All identified crop loss areas (red in figure 3(A)) in this study were consistent with reported local flood-prone regions in the river basins of China (Li et al 2016, Yang et al 2016, Chen et al 2018, 2019, Wang et al 2022), the US (Kaur et al 2020, Davenport et al 2021, Lazin et al 2021, Shirzaei et al 2021), India (Samantaray et al 2014, Amarasinghe et al 2020), Argentina (Seiler et al 2002, Ferrazzoli et al 2010, Abelen et al 2015), and Pakistan (WFP 2010, Sayama et al 2012). For crop-specific information, the Mississippi in the US for wheat, the southern branches of the Amazon in Brazil for maize and rice, the middle Godavari and Krishna in India for rice and soy, and the Loire and Seine in France for wheat were identified. Here, we should not overlook the crop loss areas that were not statistically significant (light red in figure 3(A)), because those areas could be damaged by floods in the near future when a wetter world arrives caused by the global warming (Allen and Ingram 2013, Donat et al 2016). Those potential areas are concentrated in global food baskets such as eastern China, the central US, and southern Brazil, and make up about 18% to 35% of the global harvest areas for the examined crops (figure 3(A)).

Zoom In Zoom Out Reset image size

Yield gains from floods have been reported for local studies (Chen et al 2019, Li et al 2019), and were detected in our results for some crops and regions of the world as well. One reason for the yield gains in flood years could be that inundation or soil saturation levels were not critical to yield loss under a dry conditions—high temperature and low soil moisture make floods beneficial to crops by mitigating their heat stress and fulfilling water demand (Li et al 2019). This yield gain would also be prominent in dry regions where evaporation was high and water was the primary limiting factor of crop cultivation. In our results, the yield gains were detected in the lower Rio Grande in the US for maize and wheat and the upper Niger in Mali for maize and rice (figure 3(A)). Therefore, we infer that flooding might not lead to crop loss under the wetter-than-normal conditions in a dry region because the return period is the relative value of temporal variations in an inundation area. This limitation should be noted for future research. Flood-induced yield gains were also detected in the lower Parana in Argentina for maize and soy, the lower Amazon in Brazil for rice, and the Krishna in India for rice. The reason given earlier for yield gains does not well explain these results, since there were no dry regions in this case. As another reason for yield gains, we suggest that there was a mismatch between cultivated and inundation areas according to localized spatial variation within a grid cell as reported by Chen et al (2019). Unfortunately, we could not determine any relationship of our yield changes with geographical slope and also could not obtain global data to define their relationship with the flow accumulation area and the distance from a nature freshwater resource. Therefore, it is probably necessary to investigate flood-induced yield changes with fine spatial resolution data to precisely define flood-induced crop losses. Fine temporal resolution data are also required because duration of inundation is one of the significant factors used to estimate yield loss (Kaur et al 2020).

It was difficult to directly validate global flood effects on crop yields because of data limitations. Nonetheless, ROC analysis of an annual count of flood grids with a binarized flood event compiled from the EM-DAT has been a meaningful validation of the relationship between a flood area and crop yield (Chen et al 2019, Shirzaei et al 2021, Wang et al 2022). Figure 4(A) shows that the area under the ROC curve (AUC) for all crops in China and Argentina is ${\simeq}0.7$, which means excellent estimation (Yang and Berdine 2017). In addition, more than half of crops in all ($17/32$) had AUC higher than 0.6 which was statistically acceptable. Thus, the annual flood grid count was able to describe whether floods occurred in areas with national crops. The crops in some countries with an AUC $\lt 0.6$ were detected, and these were inappropriate for estimating flood effects on crop yields, which was a limitation of this study. Nonetheless, our estimation is helpful for understanding global or national flood effects on crop production because the major countries and crops affected by floods were analyzed.

Zoom In Zoom Out Reset image size

The ROC threshold of the annual flood grid count, which describes the presence and absence of floods in terms of a spatial domain, was differed by country and crops (figure 4(B)). A higher threshold implied a more immense flood event, namely, a rarer event than a lower threshold. In the case of China, the flood-induced yield changes of soy were relatively rarer events because floods occupied larger area were rare. Otherwise, those of rice occurred more frequently, and then the losses were greater than soy (figure 2(A)). This result imply that the reason of yield losses did not cause magnitude of flood event but frequency of that.