Measuring disaster-related crop losses: How far can household surveys take us, and what new research is needed?

Natural hazards are striking agricultural systems more often and at greater cost. In 2024, recorded disasters exceeded the twenty-year average, at a direct cost of around US$242 billion (CRED, 2025). Direct damages, though, are only the costs that get counted. Once losses to health, education, livelihoods, ecosystems, and supply chains are added, the global toll rises to closer to US$ 2.3 trillion per year; UNDRR, 2025). A large share of that damage lands on farms: FAO estimates that disasters destroyed US$3.26 trillion in crop and livestock production between 1991 and 2023, an average of US$99 billion a year, or 4 percent of global agricultural GDP, with Africa bearing the heaviest relative burden at 7.4 percent (FAO, 2025). For the 3.8 billion people whose livelihoods depend on agrifood systems — especially for smallholders who rely on crops for both income and food — these losses are a recurring threat to consumption, nutrition, and long-term development. Yet a meaningful share of these damages remains undercounted. Global inventories such as EM-DAT track major disasters well but apply reporting thresholds that exclude the smaller, more frequent shocks. These are not marginal: UNDRR reports an analysis of 104 countries finding that small and medium, localized and frequent disasters caused 68 percent of all economic losses between 2005 and 2017. In agriculture the same gap shows up directly: comparing survey microdata against EM-DAT, recent research finds that surveys capture USD 5.1 billion more in drought and flood damage across the same countries and years, affecting over four times as many people (Wollburg et al., 2024). Why survey microdata matters Household and agricultural surveys close this gap by measuring impacts directly, plot by plot, and linking them to farmers’ circumstances. That granularity reveals not just the size of the losses, but who suffers them and under what conditions. Evidence from six African countries from the Living Standards Measurement Study – Integrated Surveys on Agriculture (LSMS-ISA) finds that around 35 percent of plots were impacted by disasters in a given season, with affected farmers losing over half their potential harvest. Disaster did not strike uniformly – women-managed and poorer households were impacted both more often and harder. Surveys also complement satellites, which cover wide areas continuously but often cannot detect the boundaries of smallholders’ small, irregular plots or catch non-weather losses such as pests and disease – nor link damage estimates to household outcomes (Wollburg et al., 2024). Estimating a harvest that never happened A new review note by the 50x2030 Initiative synthesizes evidence from three analytical papers into practical guidance on a deceptively hard question: how do you measure something that never happened? To know how much a farmer lost, you must compare what they actually harvested with what they would have harvested had the disaster not struck. That second figure — the attainable yield — is never observed, and different approaches to reconstructing that missing number exist, each with benefits and challenges. The most practical is the land area and crop damage approach: farmers are asked what share of their plot they managed to harvest, and how badly the rest was damaged. It needs nothing beyond the farmer’s own recall, works plot by plot, and scales to a national survey at almost no extra cost — which is why the note treats it as the natural anchor of any crop loss module. Its weakness is human: percentages are hard to estimate, answers cluster on round numbers, and in a severe, memorable drought year farmers tend to overstate how much they lost (Markhof et al 2022). Three other approaches — a trained enumerator’s independent assessment, a comparison with the harvest the farmer expected at planting, and a benchmark from past or neighboring yields — each pin down the same number in a different way, and each has its own blind spot. No single approach is best practice in isolation — and none has yet been fully validated against an objective “gold standard” that could show which comes closest to the truth. For now, combining them is what adds value: because each approach has a different weakness, agreement across methods signals confidence, while divergence flags the observations that deserve a closer look. Building that missing benchmark is itself one of the priorities for future research. Knowing which disaster to blame Measuring how much was lost is only the first part. The other part is attribution — establishing what or which disaster or shock caused the loss. Here the evidence is harder to read than it looks. Farmers in the same village often disagree on whether they were hit by drought, and their reports match satellite rainfall data only in the most severe seasons. The tempting reading is that farmers are simply misreporting. But exposure and ability to cope differs from plot to plot and from farmer to farmer — a soil that holds water better can buffer a poor rainy season — so in a mild drought some farmers may lose a harvest while their neighbors don't. When drought is severe, differences in perception and in plot characteristics both matter less, and agreement between farmers rises. The difficulty is that survey data alone cannot separate the two: disagreement may reflect differences in perception or real differences in highly localized losses (Markhof, Ponzini & Wollburg, 2022). Where the research goes next Two frontiers look especially promising for improving the measurement of crop losses. The first is systematic integration of geospatial data — rainfall indices, flood maps, vegetation-stress indicators — to anchor farmers’ shock reports against objective conditions and to feed models that predict attainable yields. The second, largely untested, is image-based AI: photographs of damaged plants, processed by computer-vision models trained on crop-cut data, could one day estimate loss severity and tell drought stress from pest or flood damage without relying on recall. These technologies do not eliminate the need for household surveys. Rather, they make survey data more valuable. Image-based AI assessment, for instance, could run through the survey itself — farmers photographing their plots or individual plants when a loss occurs, making the survey multimodal. Surveys also provide the ground-level information needed to validate new methods — and remain uniquely able to connect a disaster’s physical effects with the people who bear its costs. Improving disaster-loss measurement, in the end, means not choosing between surveys, satellites, and new technologies, but learning how to combine them. And crop losses are only one part of the toll: damages to livestock and stored assets remain even less well measured, and extending the same survey-based approach to them is work that still lies ahead. Main sources Markhof, Y.V., Ponzini, G. & Wollburg, P.R. (2022). Measuring Disaster Crop Production Losses Using Survey Microdata: Evidence from Sub-Saharan Africa. World Bank Policy Research Working Paper No. 9968. doi.org/10.1596/1813-9450-9968 Ponzini, G. & Wollburg, P. R. (2026). Measuring Disaster-Related Crop Production Losses Using Survey Micro-Data: A Review of Approaches, Evidence, and Survey Design Considerations. 50x2030 Initiative to Close the Agricultural Data Gap. Link Wollburg, P., Markhof, Y., Bentze, T. & Ponzini, G. (2024). Substantial impacts of climate shocks in African smallholder agriculture. Nature Sustainability, 7(11), 1525–1534. https://doi.org/10.1038/s41893-024-01411-w CRED (2025). 2024 Disasters in Numbers. Centre for Research on the Epidemiology of Disasters, Brussels. Link UNDRR (2025). Global Assessment Report on Disaster Risk Reduction 2025. United Nations Office for Disaster Risk Reduction, Geneva. Link