The scale of the rescue effort is remarkable. During two field campaigns in 2023, a team led by Derrick Muheki of the Vrije Universiteit Brussel first spent two weeks compiling a comprehensive inventory of the archives, cataloguing station names, codes, coordinates and altitudes, and verifying that each bundle of sheets was correctly assigned to its station, year and month. They then spent two months photographing every available sheet using a digital camera mounted on a tripod inside a portable photo box with uniform lighting. Because the archive site has no electricity, the entire setup ran on rechargeable batteries. In total, 9885 sheets from 37 weather stations across the Democratic Republic of the Congo were imaged, covering records that mostly span from the 1960s to the mid-1990s, with a few extending into the early 2000s. The urgency was real: a fire had damaged one wing of the Yangambi archive just months before the digitization campaign began.

Photographing the sheets was only the first step. Reading them posed a far greater challenge, since every value was handwritten, in styles that varied from observer to observer and decade to decade, on paper of wildly varying quality and condition. To solve this, the team applied an improved version of MeteoSaver, an open-source software package they developed that uses machine learning to transcribe handwritten tabular data. The software processes each image through a pipeline of modules: it enhances the image with adaptive threshold binarization, detects the table structure and individual cells using computer vision algorithms, recognizes the handwritten values with a trained Tesseract optical character recognition model, and then subjects every transcribed value to a battery of automated quality control checks before formatting the results in the standardized Station Exchange Format prescribed by the Copernicus Data Rescue Service.

The quality control system is where the technical ingenuity really shows. Each transcribed value is checked against region-specific temperature thresholds, logical consistency rules and inter-variable relationships. Because the original observers calculated five-day and monthly totals and averages on every sheet, the software can recalculate these summaries from the transcribed daily values and compare them with the handwritten ones, confirming entire blocks of data when they agree within a 0.2 degree Celsius margin that accounts for historical rounding. The sheets also record dry bulb temperature, wet bulb temperature, actual vapour pressure, relative humidity and vapour pressure deficit at 06:00, 15:00 and 18:00 local time, and these variables are physically linked by well-known atmospheric equations. The software exploits those relationships, using a modified Magnus-Tetens formula and the psychrometric equation from the Smithsonian meteorological tables, to cross-validate the transcribed values and flag or correct inconsistencies. This iterative approach paid off dramatically: after five consecutive runs over the archive, in which newly confirmed transcriptions were fed back into the training data, the number of quality-controlled confirmed values rose from about 321,000 to more than 1,019,000, a 215 percent increase.

Validation against independently hand-transcribed records from five stations showed that the automated pipeline achieved match rates between 81.2 and 92.3 percent, with a median of 87.4 percent, and a median mean absolute error of 0.35 degrees Celsius. That error is of the same order as the observational uncertainty of historical thermometer measurements in the tropics, and part of the disagreement likely stems from errors in the manual reference data itself rather than the machine. Most mismatches were not misread digits but one-day shifts caused by row misalignment in table detection, a problem the authors identify as a priority for future improvement. Combined with previously transcribed datasets, including the COBECORE project’s pre-1960 records from the Belgian State Archives and a manually transcribed series for Yangambi-Km.5 extending to 2020, the new dataset spans 1908 to 2023 and forms the largest in situ temperature and precipitation dataset ever assembled for the Congo Basin.

What the rescued records reveal is unambiguous. Of the 21 stations with sufficient data for trend analysis, 20 show warming in at least one of daily maximum, minimum or average temperature. The median trends are 0.22, 0.10 and 0.15 degrees Celsius per decade respectively, corresponding to roughly 0.7, 0.3 and 0.5 degrees Celsius of warming over a 30-year period. These magnitudes are consistent with previous regional assessments that estimated 0.75 to 1.05 degrees Celsius of warming in Central Africa since 1960, but they now rest on direct, station-level evidence from within the basin itself. The analysis, based on the Theil-Sen estimator, a robust method for detecting trends in noisy time series, also shows that the temperature distributions themselves are shifting: with each successive decade between 1961 and 1990, the probability distributions of daily maximum, minimum and average temperatures move progressively toward warmer values, with modal temperatures and 95th percentile extremes rising by nearly a degree Celsius between the first and last decades at several stations.

The changing face of temperature extremes is perhaps the most striking finding. Across the aggregated station network, the average number of hot days per year, defined as days exceeding the 95th percentile of daily maximum temperature, nearly doubled between the earliest and most recent decades of the common observation window, rising from about three days per year above 33.3 degrees Celsius to about five. Days with average temperatures above 26.9 degrees Celsius doubled from roughly three to six per year. Meanwhile, cold days, those falling below the 5th percentile, declined from around seven to five per year, with cold nights showing a similar decline. At individual stations with longer records, such as Luki-Plateau, Mulungu-Mulohe and Yangambi-Km.5, the number of hot days more than tripled over the period 1961 to 2000. Because the IPCC’s Sixth Assessment Report could not assess changes in hot extremes in Central Africa at all, these observations fill a genuine hole in the global picture of climate change, providing the raw material that future detection and attribution studies will need.

The precipitation story is more tentative. Only three stations had more than 30 years of quality-controlled daily rainfall data, and at two of them, Yangambi-Km.5 and Mulungu-Mulohe, heavy precipitation events became markedly more frequent in the later half of the record. At Mulungu-Mulohe, a daily rainfall of 63 millimetres, which had a return period of about two years in the earlier period, became roughly a once-a-year event in the later period. At Yangambi-Km.5, an 83-millimetre day likewise became twice as frequent. The third station, Gimbi-Plateau, showed no substantial change. With such limited spatial coverage, the authors are careful to present these as station-specific examples rather than a representative picture of rainfall change across the Democratic Republic of the Congo, a caution that reflects the harder quality control problem posed by precipitation, for which fewer automated consistency checks are available.

The team also compared their station trends with the ERA5-Land reanalysis, a state-of-the-art gridded product from the European Centre for Medium-Range Weather Forecasts. The reanalysis captures the seasonal cycle and the broad warming trend from the 1960s onward reasonably well, but shows no systematic pattern of agreement or disagreement with the stations, sometimes underestimating and sometimes overestimating local trends. More troubling, ERA5-Land exhibits a pronounced artificial jump of about 4.5 to 5 degrees Celsius in its temperature series during the early 1950s that appears in none of the station records from that era, a sign that the reanalysis is weakly constrained in this region during its early period when few observations were available to anchor it. The rescued dataset, now uploaded to open-access repositories including the Copernicus Global Land and Marine Observations Database and Zenodo, offers exactly the kind of ground truth needed to fix such products.

The implications reach well beyond Central Africa. Most station records end in the mid-1990s, when the First and Second Congo Wars disrupted observations nationwide, meaning the full acceleration of warming since the 1990s is likely not captured in these data. Roughly 300,000 additional temperature records and 200,000 precipitation records were transcribed but did not yet pass quality control, and 225 hand-drawn sheets were excluded entirely, all of which could be recovered with improved software or citizen science verification. Meanwhile, the COBECORE project has imaged records from 575 pre-1960 stations that could add millions more observations if transcribed. The authors present their workflow, from inventory to imaging to machine-learning transcription to quality control, as a replicable template for other data-sparse regions where millions of weather observations still moulder on paper. For a basin that holds the world’s second-largest river and rainforest, and some of its most climate-vulnerable populations, turning those fragile pages into hard numbers is not just an archival triumph; it is the foundation for understanding, and preparing for, the region’s warming future.

Subject of Research: Climate data rescue and observed warming trends in the Congo Basin

Article Title: From paper to proof: revealing Congo Basin warming through rescued climate archives

Article References: Muheki, D., Hufkens, K., Jacobsen, K., Verbeeck, H., Boeckx, P., Kankonde Ntumba, D., Kapalay Moulasa, O., Vercruysse, B., Birkholz, J. M., Verbruggen, C., Hawkins, E., Lampe, S., Yakusu, E. K., Imwangana, F. M., Mbifo, J., Likwela, T. B., Meunier, F., Dewitte, O., Thorne, P., & Thiery, W. (2026). From paper to proof: revealing Congo Basin warming through rescued climate archives. Earth System Dynamics, 17(5), 1455-1511. https://doi.org/10.5194/esd-17-1455-2026

Image Credits: AI Generated

DOI: 10.5194/esd-17-1455-2026

Keywords: Congo Basin, climate data rescue, Democratic Republic of the Congo, temperature trends, extreme heat, machine learning, MeteoSaver, historical weather records, ERA5-Land, precipitation extremes, IPCC, Earth System Dynamics

The rescue operation began in earnest in 2023, when Derrick Muheki, a PhD researcher at VUB and lead author of the study, joined two field campaigns in the Democratic Republic of the Congo. Working alongside INERA staff and colleagues from Ghent University, the team photographed more than 9,000 handwritten monthly sheets of weather observations collected at 37 stations distributed across the country. Some of these records were also recovered from the State Archives of Belgium, a legacy of the colonial-era meteorological networks that once operated across Central Africa. The physical condition of the documents added urgency to the effort. Part of the INERA archive had been damaged by fire only months before Muheki’s visit, a near-loss that underscored how vulnerable such irreplaceable records remain.

Photographing the sheets was only the first step. Converting more than 9,000 images of cursive handwriting into usable numbers would have taken years of manual transcription, so the team relied on MeteoSaver, an open-source software tool developed by Muheki at VUB in collaboration with the Ghent Centre for Digital Humanities. The software automates the transcription of digitized weather records into digital spreadsheets, bridging the gap between historical scans and quantitative climate analysis. Its performance has drawn attention well beyond the climate community: TIME recognized MeteoSaver among its Best Inventions of 2025 as a special mention, highlighting how data rescue has become a technological frontier in its own right.

The resulting database, published openly on Zenodo, contains more than one million daily and sometimes sub-daily weather observations. Together with previously rescued records, it represents one of the largest climate datasets ever assembled for the Congo Basin, a region that had until now been largely overlooked by the international research community. The archive covers observations from 1908 to 2023 and includes daily maximum, minimum and average temperatures, daily precipitation, and dry- and wet-bulb temperatures measured three times a day at the 37 stations. Although each station covers a different period, most records begin in the 1960s and end in the 1990s, an abrupt truncation that reflects the Congo wars and the nationwide institutional breakdown that followed.

Once the transcribed data were quality-controlled and analyzed, a consistent warming signal emerged across the basin. Between 1961 and 1990, daily maximum temperatures rose by about 0.7 degrees Celsius, minimum temperatures by 0.3 degrees Celsius, and average temperatures by 0.5 degrees Celsius, corresponding to warming of tenths of a degree per decade. Crucially, the trend did not stop when the INERA records end. Using weather stations with longer continuous records, the researchers show that the warming of the Congo Basin continues into the 2000s, confirming that the signal detected in the rescued archives is part of an ongoing, accelerating pattern rather than a relic of the twentieth century.

The most consequential findings concern extremes. Each successive decade shows an increase in the number of hot days per year, and the hot days themselves are getting hotter, while the number of cold days decreases, mirroring patterns observed across the globe. On average, the number of hot days per year nearly doubled between the 1960s and the 1980s. For the millions of people living in the Congo Basin, many of whom depend on rain-fed agriculture and have limited access to cooling or healthcare, such a shift in the frequency of extreme heat carries direct consequences for labor productivity, food security and human health.

Rainfall extremes show a similar intensification. Examining precipitation records from three weather stations, the authors found that the number of heavy rainfall events doubled at two of these stations in the 1980s compared with the 1960s. The observed increase in both the frequency and intensity of weather extremes in the DRC is consistent with global trends linked to rising temperatures driven by continued fossil fuel emissions. In a region where infrastructure is often fragile and flood defenses sparse, a doubling of heavy rainfall events translates into heightened risks of flooding, landslides and disruption to river transport along the Congo River and its tributaries.

The scientific significance of the study extends beyond the regional numbers. Central Africa remains one of the few regions where the most recent Intergovernmental Panel on Climate Change report could not confirm an increase in hot extremes, largely because of the absence of past observations and scientific studies. Wim Thiery, senior author of the study and professor at VUB, emphasized that the new observational evidence fills a critical gap in a data-sparse yet climatically significant region. The Congo Basin is home to the world’s second-largest tropical rainforest, a major carbon sink, and its climate dynamics influence rainfall patterns far beyond its borders. Reliable historical observations are therefore essential both for understanding regional change and for evaluating and improving the climate models used to project its future.

The work also carries a pointed message about global scientific equity. Thiery noted that the findings highlight the urgent need for more data rescue projects, especially in data-sparse regions, to close knowledge gaps on climate trends. Regions in the Global South are least responsible for the greenhouse gas emissions driving climate change yet suffer the brunt of its impacts, and robust historical records are a prerequisite for informed adaptation planning and a just transition. Muheki, who witnessed the fragility of the archives firsthand, described the urgency of the work: seeing how vulnerable the paper records were, and knowing that part of the archive had already burned, made clear that decades of irreplaceable data could have been lost. Accessing more of these records, he said, will help improve reconstructions of climate history and projections of future climate change in a part of the world where adaptation to climate impacts is already a necessity.

The study, authored by an international team spanning VUB’s Water and Climate Department, Ghent University’s Congo Basin Centre of expertise and Ghent Centre for Digital Humanities, BlueGreen Labs, INERA in Yangambi and Kinshasa, the University of Reading, the Royal Museum for Central Africa, Université de Kisangani, Université de Kinshasa, Maynooth University, the Royal Library of Belgium and the Center for Geological and Mining Research in Kinshasa, demonstrates what patient archival work combined with modern machine-assisted transcription can achieve. The two field campaigns were documented in a Nature News article in September 2025, and the open-access database now invites researchers worldwide to build on the rescued observations. For a region that has been nearly invisible in global climate datasets, the Congo Basin finally has a documented warming history, and it was written by hand.

Subject of Research: Historical weather data rescue and warming trends in the Congo Basin

Article Title: Hot days become twice as common between 1960 and 1990 in the Congo Basin: Proofs rescued from paper archives

Article References: Hot days become twice as common between 1960 and 1990 in the Congo Basin: Proofs rescued from paper archives. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: Congo Basin, climate change, data rescue, heat extremes, Democratic Republic of the Congo, INERA, MeteoSaver, Vrije Universiteit Brussel, Earth System Dynamics, historical weather records, heavy rainfall, IPCC

The study, accepted for publication in the journal Weather ahead of the release of the feature film Pressure, which dramatises Stagg’s dilemma with Andrew Scott in the title role and Brendan Fraser as General Eisenhower, generated fifty separate forecasts of the June 1944 storm using machine-learning prediction techniques. Rather than producing a single deterministic answer, the researchers examined how confident the AI system was about the conditions that Stagg had to assess for 5 and 6 June. The results were strikingly mixed. For 5 June, the original invasion date, the model placed the storm in the correct position and correctly anticipated poor conditions over Normandy, in agreement with Stagg’s own reasoning. Yet for 6 June, the day the landings actually took place, the model made a consequential error, predicting that the storm would sweep away towards Norway far more quickly than it truly did and that the invasion day would be calmer and clearer than reality.

Professor Andrew Charlton-Perez, lead author of the study, framed the historical stakes in unambiguous terms. Stagg, he noted, held one of the hardest forecasting jobs in history: get the prediction wrong and the invasion could have failed, with consequences for the entire course of the Second World War. Some Allied meteorologists, after weeks of predominantly hot and settled weather, were insistent that conditions would remain fine for the invasion to proceed as planned on 5 June. The new research suggests that Stagg, had he been handed the AI forecast alongside his own charts, would likely have reached the same conclusion he did historically. The model pointed to a genuine risk of dangerous winds on 5 June, a signal strong enough to justify recommending the one-day postponement that ultimately allowed the landings to proceed under a brief lull in the weather.

The story of 6 June, however, reveals the uncomfortable limits of even the most advanced machine-learning systems. Eisenhower’s famous order, “OK, we’ll go,” rested on Stagg’s belief that conditions on 6 June would be fair enough to launch. In reality, soldiers landing on the Normandy coast encountered high winds and stormy seas more treacherous than the reassuring picture painted by the model. The AI system showed the storm clearing away from Britain too quickly, incorrectly suggesting that 6 June would be calm and bright. It also predicted clearer skies than actually occurred, compounding the false sense of security. In other words, a modern forecasting system trained on decades of global atmospheric data, with access to vastly more observations than Stagg ever had, would still have underestimated the very storm that made the D-Day landings so hazardous.

Perhaps the most technically revealing finding lies beneath the model’s overall confidence. Hidden within its aggregated forecast was a wind statistic that told a very different story: the model still assigned a 30 percent probability that winds on 6 June would exceed the safety limit for the landing craft. That significant residual risk was masked by the model’s reassuring average picture, illustrating a subtle challenge in how probabilistic ensemble forecasts communicate danger. An overall forecast that reads as benign can conceal a meaningful tail of adverse outcomes, and it takes a trained human forecaster, scrutinising the distribution rather than the mean, to extract that signal. Stagg himself, weighing conflicting advice from competing forecasting teams in the manner dramatised in the film, performed precisely this kind of synthesis with the tools of 1944: surface observations, limited upper-air soundings, and hard-won experience of Atlantic weather behaviour.

The methodology behind the study is as noteworthy as its conclusions. To judge how well the model performed, the researchers compared its forecasts against what actually happened over the Channel and the Normandy beaches, drawing on modern weather records from the Copernicus Climate Change Service, which reconstructs historical atmospheric conditions. Because weather records from 1944 were sparse, the team also turned to an unexpected archival treasure: newly rediscovered paper weather logs from the Faroe Islands, held in an archive at the Met Office. The Faroe Islands sit close to the path the storm followed as it passed Britain, so the readings recorded there in 1944 captured the storm at a critical point that the main observational networks missed. The comparison offered an independent check on the system’s position and intensity that would otherwise have been impossible.

Converting those handwritten logs into digital data allowed the researchers to assess the storm’s strength more closely, and the exercise produced a small but meaningful revision to history: the June 1944 storm may have been slightly stronger than current reconstructions suggest. This finding has implications beyond historical curiosity. Reconstructions of past weather underpin how scientists validate forecasting systems against real events, and even modest corrections to storm intensity can change how a model’s errors are interpreted. The researchers argue that digitising more historic paper logs from archives across Europe would allow storms like the one that threatened the D-Day landings to be tested with even greater precision in the future, enriching the observational record precisely in the oceanic and remote regions where modern reanalyses remain least constrained.

The wider lesson the authors draw speaks directly to the current enthusiasm for artificial intelligence in weather prediction. Machine-learning forecasts can process enormous volumes of data in seconds, producing global outlooks at computational speeds that traditional numerical models, which simulate atmospheric physics on supercomputers for hours at a time, cannot match. Yet the D-Day replay demonstrates that these systems can still make mistakes, particularly in unusual or extreme situations that sit outside the statistical patterns they learned from training data. A historical storm from a data-sparse era is an especially demanding test case, and the model’s errors on 6 June, its over-rapid clearing and its excessively optimistic skies, are exactly the kind of failure that could prove costly if trusted uncritically.

Charlton-Perez’s conclusion is that the human forecaster remains as essential today as in 1944. Having a trained meteorologist make sense of forecast data, interrogate the probabilities, and apply knowledge and experience to the final call is, in his words, just as critical now as it was for the D-Day landings. The study stands as a rare natural experiment: a chance to pit a state-of-the-art AI system against a decision that altered world history, using the same deadline pressure and far better tools. Stagg passed the test with paper charts and courage. The machine, for all its speed and scale, got half of it wrong, and a thirty percent hidden wind risk reminds us why someone still needs to read between the lines of every forecast before the order is given to go.

Subject of Research: Testing modern AI weather forecasting models against the historical D-Day storm forecast of June 1944.

Article Title: AI D-Day weather forecast falters under pressure

Article References: AI D-Day weather forecast falters under pressure. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: AI weather forecasting, D-Day, James Stagg, machine learning, Normandy invasion, storm forecasting, Weather journal, University of Reading, Copernicus Climate Change Service, ensemble forecasts, historical weather records, meteorology