Massive 40-Year Satellite Dataset Maps Every Forest Change Across China

China’s forests have been quietly rewriting their own story for four decades, and until now scientists have lacked a trustworthy ground truth to read it. A team of researchers led by Rong Shang and Jing M. Chen of Fujian Normal University, working with colleagues at institutions including the University of Toronto, has unveiled a reference dataset that documents where Chinese forests remained untouched, where they were disturbed, and where new forest took root between 1986 and 2025. The work, published as a preprint under review in the journal Earth System Science Data, delivers 12,991 geographically representative samples at a resolution of 30 meters, each one painstakingly verified by human experts rather than left to the judgment of an algorithm alone.

The dataset, formally called the Reference sample dataset of Stable, Disturbed and Forestated forests, or RSDF, was assembled using stratified random sampling to ensure that every major forest region of China contributed its fair share of samples. This sampling strategy matters because forest change in China is anything but uniform. According to the analysis, 79.61 percent of the samples represent forest that stayed stable across the entire 40-year window, while 15.99 percent recorded a single disturbance or forestation event. The remaining 4.40 percent of sites experienced two or more distinct events, a reminder that some landscapes have been cut, burned, or replanted more than once within a single human generation.

What sets this effort apart from earlier mapping campaigns is the sheer rigor of its annotation process. Every sample was labeled through hierarchical visual interpretation that combined Landsat time-series data, which stretches back to the mid-1980s, with multi-source high-resolution satellite imagery and, where available, official forest inventory records. Interpreters did not simply glance at a picture and guess. Instead, they examined the full spectral trajectory of each site, tracking how six individual spectral bands and two vegetation indices shifted over time, and compared matched pre-change and post-change RGB composite images side by side. This standardized visual diagnostic evidence allowed analysts to distinguish a timber harvest from a wildfire, and a plantation from natural regrowth, with a level of confidence that automated classifiers have struggled to match.

Quality control did not stop at the first annotation. The team implemented a multi-expert blind review framework in which interpreters assessed samples without knowing what their colleagues had concluded, combined with a sequential consensus mechanism that escalated ambiguous cases through additional rounds of scrutiny. The results speak for themselves: more than 95 percent of the final samples received high-confidence ratings, the mean inter-expert pairwise agreement reached 0.83 with a standard deviation of 0.07, and the average Cohen’s Kappa coefficient, a statistical measure of agreement that corrects for chance, came in at 0.65 with a standard deviation of 0.13. In the world of reference data collection, where sloppy labels can silently corrupt an entire generation of machine-learning models, these numbers represent a benchmark that other national datasets will now be measured against.

The spatial patterns embedded in the dataset tell a vivid story about how and where China’s forests have changed. Forest disturbances were concentrated most heavily in East and South China, the densely populated and economically dynamic regions where timber demand, agricultural expansion, and urban pressure have historically pressed hardest on wooded land. By contrast, disturbance activity was comparatively weak across Northwest, North, and Northeast China, where forests are often more remote, drier, or subject to different management regimes. This east-west and north-south asymmetry provides exactly the kind of regional texture that coarse national statistics tend to smooth away, and it gives carbon modelers a realistic map of where ecosystem carbon stocks were most likely to fluctuate.

Perhaps the most striking single finding is the dominance of harvest as the leading cause of forest disturbance nationwide. Fires, other disturbance types, and forestation events each accounted for only a small proportion of the recorded changes. That result carries real weight for climate policy, because a harvested forest and a burned forest follow very different recovery trajectories and release or reabsorb carbon at different rates. Knowing that human extraction, rather than natural catastrophe, has been the primary driver of forest loss across China allows researchers to tune disturbance-recovery models, refine estimates of the national carbon budget, and evaluate whether reforestation programs are genuinely offsetting the wood and land that continue to be removed.

The absence of a long-term national reference sample collection has been a persistent bottleneck for Chinese forest science. Machine-learning approaches to disturbance detection, change-detection algorithms, and satellite-derived forest products all require labeled examples to train on and independent samples to validate against. Without them, products generated for China could not be rigorously checked, and disagreements between competing maps could not be adjudicated. The RSDF dataset directly fills this gap, and the authors emphasize that its applications extend to algorithm calibration, model training, and product validation, effectively providing the shared yardstick that the community has been missing for forty years.

The temporal reach of the dataset is equally significant. Starting in 1986, shortly after the Landsat archive became dense enough to support consistent annual monitoring, the dataset spans the entire arc of China’s modern forest transformation, from the early years of large-scale afforestation campaigns through the recent era of ecological restoration programs. Because the record extends to near present, it captures not only historical baselines but also the most recent dynamics, allowing scientists to assess whether current policies are bending the curve of forest change in the intended direction. For carbon budget assessment in particular, a continuous four-decade record is the difference between estimating a snapshot and reconstructing a full trajectory of emissions and uptake.

The dataset has been made publicly available through the Figshare repository, reflecting a growing conviction in the earth-science community that reference data are infrastructure, not private property. Open access means that research groups anywhere in the world can use the samples to benchmark their own algorithms against Chinese forest dynamics, and that future updates can extend the record as new satellite imagery arrives. The preprint is currently open for community discussion, with peer review underway at Earth System Science Data, a journal that specializes in publishing datasets of exceptional value to the broader research community.

For anyone tracking the health of the planet’s forests, the significance of this release is hard to overstate. Forests are among the most important terrestrial carbon reservoirs, and disturbances to them, whether from logging, fire, or land conversion, ripple through the global carbon cycle, biodiversity patterns, and regional climate. By anchoring four decades of Chinese forest change in a rigorously verified, expert-annotated sample collection, the RSDF dataset transforms what was once a patchwork of uncertain estimates into a coherent, testable record. It is the kind of unglamorous, meticulous data work that rarely makes headlines on its own, yet it underpins nearly every headline that follows about forests, carbon, and the pace of environmental change in one of the world’s most consequential landscapes.

Subject of Research: A 40-year reference sample dataset of stable, disturbed and forestated forests across China for forest change monitoring and carbon assessment

Article Title: A reference sample dataset of stable, disturbed and forestated forests over China from 1986 to near present

Article References: Shang, R., Xu, S., Yang, Z., Lin, X., Fan, L., Fang, K., Xu, M., Ding, A., Yan, Y., Liang, Y., Song, C., Chen, W., Qian, J., Zhang, A., Heng, J., Chen, S., Zhang, X., Liu, L., Li, W., … Chen, J. M. (2026). A reference sample dataset of stable, disturbed and forestated forests over China from 1986 to near present. https://doi.org/10.5194/essd-2026-705

Image Credits: AI Generated

DOI: 10.5194/essd-2026-705

Keywords: forest disturbance, forestation, China, remote sensing, Landsat, reference dataset, carbon cycle, satellite imagery, Earth System Science Data, forest monitoring, sample annotation, carbon budget