The first layer surface reflectance (SR) image of Landsat Collection 1 (C1) is selected, including all available Landsat 5–8 images from 2000 to 2020. These images, which have undergone atmospheric correction using the LaSRC algorithm (Skakun et al., 2021), were directly obtained from the GEE platform (https://developers.google.com/earth-engine/datasets/catalog/landsat last access: 10 May 2025). We preprocessed global images based on the quality assessment (QA) bands of Landsat SR, including removing shadows and cloud interference. These data are used to calculate CCDC to supplement the current missing CCDC dataset in some forest disturbance areas. At the same time, high-quality Landsat images are also the main dataset for us to visually select sample points.

Previous studies have shown that global forest disturbances have a high degree of spatial clustering (Hansen et al., 2013). For example, shifting cultivation is mainly distributed in the tropics (Chen et al., 2023a); Forest fires are mainly distributed in northern forests (Scholten et al., 2021). Adding spatial distribution information of forests is beneficial for improving the accuracy of forest disturbance identification. Considering the regional differences in the distribution of global forest disturbances, we defined four major regions, Africa, Southeast Asia and Australia, Central America and South America, and the Northern Forest Region. Therefore, we added latitude and longitude information as well as partition information when extracting the spatial attributes of the sample points. A large number of non-forest areas and undisturbed regions around the world do not require calculation. In order to reduce the consumption of computing power, we use Hansen's global forest change dataset to identify disturbed forest areas worldwide from 2000 to 2020 (Hansen et al., 2013).

To assess the mapping accuracy of global forest change areas, this study incorporated multiple authoritative land cover and forest cover products as reference datasets, including: (1) ESA WorldCover 2020, (2) Global Forest Change dataset (Hansen et al., 2013), and (3) the Global Forest Cover Change dataset (Roffe et al., 2022). These datasets served three primary functions in our analysis: delineating the global forest extent in 2000, identifying 2020 forest cover distribution, and classifying non-forest land cover types in 2020. The detailed information of these datasets is systematically documented in Table 3.

The CCDC algorithm is commonly used to monitor surface cover disturbances (Zhu and Woodcock, 2014). It fits the model to the Normalized Difference Vegetation Index (NDVI) of spectral observations and can reflect three types of pixel changes: seasonal changes (such as phenology), slow changes (such as vegetation growth or degradation), and rapid changes (such as deforestation, fires) (Zhu and Woodcock, 2014). CCDC uses a powerful iterative reweighted least squares method (RIRLS) (Burrus et al., 1994) to fit the observed geomorphological features and trends that reflect phenology (Tollerud et al., 2023; Hwang et al., 2022). The expression for the mathematical fitting line is as follows: 1NDVI^(x)RIRLS=a0+a1cos⁡2πTx+b1sin⁡2πTx+a2cos⁡2πNTx+b2sin⁡2πNTx Where x represents day of year; T represents the number of days per year; N represents the number of years of Landsat. a0 represents the benchmark value of NDVI; a1 and b1 represent the annual changes in NDVI; a2 and b2 represent the interannual changes in NDVI. NDVI^(x)RIRLS represents the predicted value of NDVI based on RIRLS fitting corresponding to the xth day of year.

We collected Google Global Landsat based CCDC segments (1999–2019). The dataset was created from the Landsat 5, 7, and 8 Collection-1, Tier-1, surface reflectance time series, using all daytime images between 1 January 1999 and 31 December 2019. Each image was preprocessed to mask pixels identified as cloud, shadow, or snow (according to the “pixel_qa” band), saturated pixels, and pixels with an atmospheric opacity > 300 (as identified by the “sr_atmos_opacity” and “sr_aerosol” bands). We have removed duplicate pixels in the overlapping scenes between the north and south. The results were output in 2° tiles for all landmasses between -60 and +85° latitude. We can directly call this dataset [ee.ImageCollection(“GOOGLE/GLOBAL_CCDC/V1”)] in GEE. The dataset provides extensive coverage of global forest areas, but small number of missing areas occur along the edges of some images, accounting for approximately 6 % of the total global forest area. For the missing areas in the dataset, the CCDC algorithm is used to complete them, thereby obtaining vegetation change characteristics covering all forest areas worldwide. Based on the segmented fitting results of these features, we extracted the OC, PDC, and PDP of each pixel separately (Fig. 2). In fact, not all the pixels will perfectly fit into this assumed model, which would consequently affect the fitting performance of CCDC and therefore the subsequent disturbance mapping. To address this issue, we also added feature indicators such as DP and LUC (Table 2) before machine learning training and classification to improve the robustness of the classification model.

Forest disturbance has strong disturbance sources. Therefore, using existing disturbance source datasets to assist in identifying typical forest disturbance types can effectively improve the accuracy of mapping results. Generally, a high spatial consistency is typically observed between disturbance types such as forest fires and plantation expansion and global fire and plantation distribution. This study uses global fire distribution datasets, artificial plantation distribution datasets, oil palm datasets, and other auxiliary methods to identify forest disturbance types. Meanwhile, there is a high correlation between population distribution and forest disturbance. This study collected a forest disturbance potential dataset from three aspects: population density, forest fire distribution, and spatial distribution of oil palms.

We constructed the sampling framework based on the Hansen Global Forest Change dataset and auxiliary disturbance distribution maps. The global forest region was stratified into potential disturbance layers, each corresponding to a likely disturbance type based on auxiliary data. This stratification ensured that even rare disturbance types were adequately represented in the sample pool.

To prevent overrepresentation of dominant disturbance types, we defined minimum target sample sizes for each class (Table A1). Within each stratum, candidate sample points were generated using random coordinate sampling, ensuring that each class met or exceeded its target sample size. The final sample sizes reflect only the visually interpretable and high-quality points retained after rigorous screening, which removed points affected by cloud, shadows, or poor image quality. This quality control step resulted in minor variations (< 10 % per class) from the target sample sizes. This approach ensured balanced representation across all disturbance types while maintaining spatial randomness within strata. Note that the class-balanced sampling and stratified design described above pertain to the training sample. The validation sample was selected under the same stratification but without class balancing, following a probability sampling approach to ensure unbiased accuracy estimation.

For 8 elementary disturbance types, 12 well-trained forest remote sensing professionals visually identified 57 000 forest disturbance type sample points for over 300 h based on Landsat time-series imagery (2000–2020) and high-resolution google imagery. The labeling process was completely independent of the initial stratification information. To ensure global consistency in sample points selection, 8 individuals were uniquely responsible for selecting 8 types, while an additional 4 individuals conducted secondary confirmation of the selected sample points. Undisturbed sample points were randomly selected from global forest cover areas and underwent the same visual verification. A point was labeled as “undisturbed” only if no detectable disturbance was observed throughout the study period in both Landsat time-series and high-resolution imagery. The sample point are evenly distributed in the global forest disturbance area (Fig. B1).

For the challenging distinction of “shifting cultivation”, its identification relied on detecting unique cyclical patterns in the time series. Interpreters were trained to confirm three key characteristics within the high-resolution historical imagery. (1) Clear cyclical boundaries: evidence of alternating phases of forest (fallow), clearing/burning (clearance), and crops (cultivation) on the same parcel of land over multiple years; (2) Short-cycle land cover change: a complete cycle typically lasts a few years, distinguishing it from permanent deforestation for agriculture; and (3) Small-scale and fragmented spatial patterns: shifting cultivation plots are usually small, irregularly shaped, and interspersed with patches of mature forest. Sample points were only designated as shifting cultivation if they met multiple of these criteria simultaneously to ensure accuracy. For the forest weak disturbance types caused by drought disturbance (16) and pest disturbance (17), their sample point selection needs to refer to high-resolution long-term remote sensing images. Meanwhile, weak disturbances in forest cover are highly time-bound. For example, the decline in vegetation index caused by a period of drought will quickly recover due to an increase in precipitation. At the global scale, it is currently limited by the availability of remote sensing images. We are unable to select relevant sample points through Landsat imagery. Therefore, this study did not consider these two weak disturbance types of drought disturbance and pest disturbance. This will be an independent topic for further research.

Considering that the core of this study is forest disturbance type classification, the Classification and Regression Tree (CART) classification regression tree algorithm, which has a high advantage in convenient classification, was chosen. In this article, we only discuss CART used for classification. For classification trees, CART uses the Gini coefficient minimization criterion for feature selection to generate binary trees. The specific process consists of two steps: decision tree generation and pruning. During the decision tree generation phase, a tree is constructed from the training dataset and is grown to its maximum possible size. Subsequently, pruning is performed using the validation dataset to select the optimal subtree, with the minimization of the loss function serving as the criterion for pruning.

Considering the significant differences in disturbances in different regions of the world, four CART sub decision trees were trained for four typical disturbance regions, namely Africa, Southeast Asia Australia, Central America South America, and the Northern Forest region, to form a larger global forest disturbance classification tree model. The entire process is completed interactively on three platforms: Python, GEE, and ArcGIS. During the training and validation process of the model, the sample points we selected were divided into a training set and a validation set according to a 7:3 pattern. 40 000 sample points are used for model training, and another 17 000 randomly selected sample points are used to validate the model training results (Stehman, 2014; Olofsson et al., 2014).

Based on the preliminary identification of 8 types of forest disturbances and the global forest change range by CART, considering the dynamic changes of flood inundation areas from 2001 to 2020, the forest areas that have been submerged by floods are determined, and then flood disaster disturbances are classified (20). Meanwhile, based on the distribution of oil palm in the plantation, the interference of oil palm since 2000 has been extracted (21). In addition, by overlaying the forest coverage in 2000 and the quantity coverage in 2020, the newly added forest areas in 2020 were extracted from the non-forest areas in the early stage (22). Combine the 8 core forest disturbance types identified by the CART with these 3 supplementary forest disturbance types to form a complete GFD map. Finally, this study generated a global forest disturbance type dataset containing 11 forest disturbance types. The final GFD map, except for the newly added forest, uses Hansen's global forest change dataset as a mask to remove pixels that have not undergone forest changes.

The error matrix is a simple cross tabulation of class labels assigned by remote sensing data classification and reference data of sample point locations. The error matrix organizes the collected sample data in a way that summarizes key results and helps quantify accuracy of various types of forest disturbance. The main diagonal of the error matrix highlights the correct classification, while the diagonal elements show omissions and errors. The cell entries and marginal values of the error matrix are the basis for accuracy evaluation estimation. In the matrix, nij represents the number of sample points whose map label is class i and whose reference label is class j. Let ni⋅=∑jnij be the total number of sample points in map class i, and n=∑ini⋅ be the total number of validation sample points.

Because the validation sample is obtained by stratified random sampling with strata defined by map classes, the contribution of each map class to the overall accuracy should be weighted by its mapped area. Let Wi=Ni/N denote the proportion of mapped area (or pixel count) belonging to map class i within the assessment domain, where Ni is the number of pixels (or area) mapped as class i and N=∑iNi.

Overall accuracy (OA) is the proportion of area correctly classified. It is computed as the area-weighted sum of the within-class correct proportions: 2OA=∑iWiniini⋅ User's accuracy for Class i (Ui) (the number of sample points mapped to class i with reference to class i): 3Ui=niini⋅ Producer's accuracy for class j is the probability that a reference sample of class j is correctly mapped as class j. Under the stratified design, the total proportion of area belonging to reference class j is estimated by combining all map classes using their area weights, i.e., ∑iWinijni⋅. Therefore: 4Pj=Wjnjjnj⋅∑iWinijni⋅

The sampling variability associated with the accuracy estimates should be quantified by reporting standard errors. The variance estimators are provided below, and taking the square root of the estimated variance results in the standard error of the estimator. For overall accuracy, note that OA=∑iWiOAi, where OAi=niini⋅. The estimated variance is: 5VOA=∑iWi2OAi1-OAini⋅-1 For user's accuracy of map class i, the estimated variance is 6VUi=Ui1-Uini⋅-1

For producer's accuracy of reference class j, let Pj=AjBj, where Aj=Wjnjjnj⋅Bj=∑iWinijni⋅. Using a first-order approximation, the estimated variance is 7VPj=1Bj2[VAj+Pj2VBj-2PjCov(Aj,Bj)] with VAj=Wj2njjnj⋅1-njjnj⋅nj⋅-1,VBj=∑iWi2nijni⋅1-nijni⋅ni⋅-1 and Cov(AjBj)=VAj because the i=j component of Bj contains Aj, while the remaining components of Bj come from other strata. The standard error of each estimator is computed as SE =V(⋅).

We compared GFD map with these other datasets to calculate the degree of agreement on typical forest disturbance types. At the global scale, the currently available dataset, “Classifying drivers of global forest loss (CDGFL)”, has a relatively low resolution (10 km) (Curtis et al., 2018), and the concept of forest loss differs from that of forest disturbance in this study. Therefore, the absolute area of forest loss under different drivers cannot be directly compared with the forest area under different disturbance types. Consequently, we only compared the proportional characteristics of forest cover under the same drivers and disturbance types across different global regions to validate the agreement of our global spatial distribution of forest disturbances.

Here, we divided the global continental area into 200 grids (n) with spatial location information, each measuring 10° × 10°. First, we calculated the proportion of forest loss area under different drivers within each grid relative to the global forest loss area, denoted as pi, where i represents the driver type. Second, we calculated the proportion of forest area under different disturbance types within each grid relative to the global forest disturbance area, denoted as qi, where i represents the disturbance type. The agreement of this study was determined by examining the spatial consistency between these two proportions. The primary metrics used for validation included the coefficient of determination (R2) from linear regression (James et al., 2013), mean absolute error (MAE), and root mean square error (RMSE). 8R2=1-∑inqi-pi2∑inqi-pi¯29MAE=1n∑inpi-qi10RMSE=1n∑in(pi-qi)2