We assessed the accuracy of the rubber map presented in Wang et al. [1] using established validation practices [2]. These practices estimate area and associated accuracy metrics for mapped land cover classes within a defined region, in this case, ‘rubber’ in Southeast Asia. Although Wang et al. claimed to apply such practices, their failure to use random sampling led to invalid estimates, statistics, and conclusions.

We used stratified random sampling to allocate 300 validation points to each of the three classes: ‘non-forest’, ‘forest’, and ‘rubber’. The ‘forest’ and ‘non-forest’ classifications were derived from an existing land cover map as used by Wang et al. and the ‘rubber’ class was obtained from their classification model.

For each of the 900 validation points, we inspected a 10x10 meter square centred on the point’s location using Google Earth and, where necessary, Sentinel-2 time series. We classified points as ‘forest’ or ‘non-forest’ based on tree cover observed in the sub-meter resolution images. ‘Rubber’ was identified by the distinctive form of rubber trees in high-resolution images. If the identification of rubber trees was unclear, as it was in a few cases, it was verified through assessing phenology in Sentinel-2 time series, as also explained and applied by Wang et al..

We further categorized rubber into rubber ‘monoculture’ (uniform closed-canopy plantations) and ‘jungle’ rubber (isolated or clustered trees within secondary forests or mixed species stands). This allowed evaluation of the Wang et al. map when considering just monoculture rubber, and with regard to both monoculture and jungle rubber. For our monoculture-only assessment, jungle rubber points were classified as ‘forest’.

We calculated three accuracy metrics: Overall accuracy—The proportion of correctly classified pixels across all classes. Producer’s accuracy—The percentage of validation points for a given class correctly classified on the map (omission errors). User’s accuracy—The likelihood that a pixel classified as a specific class on the map is correct on the ground (commission errors).

We estimated the true area planted with rubber following established methods which account for classification errors and also recognise uncertainty by providing confidence intervals [2]. To estimate the deforestation associated with rubber, we constrained our validation to areas classified as old growth humid forest in 2001 (see following paragraph) and estimated the area of rubber within this. We supplemented the original validation points with additional points to ensure 300 per class within the old growth humid forest area. This approach assumes that the 2021 rubber area was forest in 2001 and that this lacks any marked classification errors. If an area was classified as forest in 2001 and later as rubber, we assume that the area was deforested and converted to rubber, but as the mapped rubber areas includes classification errors, we generated a corrected estimate, with confidence intervals, of deforestation associated with rubber expansion.

We assessed the commission error in Wang et al.’s map [1] using previously validated data for planted oil palm and pulpwood [3,4,5]. Based on these we calculated the overlapping area between the rubber extent map and the oil palm and pulpwood maps.

Our ‘old growth humid forest’ layer was derived from Turubanova et al. [6] who call it ‘primary forest’. These authors define this forest ‘as mature natural humid tropical forest cover that has not been completely cleared and regrown in recent history’. These authors further clarify that these forests have ‘no detectable signs of human caused alteration or fragmentation, as delineated per the intact forest landscape (IFL) method’ but ‘include mature humid tropical forests that have experienced partial canopy loss at the 30 m mapped spatial resolution, or are located in proximity to forest loss, or that exist as fragments that do not meet IFL criteria’ [6]. These criteria may include selectively logged forests that remains largely intact, older regrowth forests but exclude predominantly deciduous forests. The exclusion of predominantly deciduous forests suggests any significant deforestation of old growth deciduous forest, for example, in Cambodia, Laos and China, is likely underestimated. Such forests are absent in the main equatorial regions (Sumatra and Borneo) where most rubber-related deforestation was identified by Wang et al. In Java, where deciduous forests could occur, and rubber is locally cultivated, there is little natural forest left outside protected areas [7] and deforestation is likely to be replanting of tree plantations, including rubber.

We also checked our assertion that Wang et al.’s numbers would be misunderstood in media coverage. Cases that illustrate our concerns were readily found. For example, Reuters cited the study as follow [8] ‘The researchers, in a paper published by Nature, said that previous data suggested rubber was a relatively minor problem when it comes to deforestation, compared with commodities like soy and palm oil. But high-resolution satellite data, which helped identify more plantations run by smallholders, suggested that forest losses “greatly exceed” previous estimates. More than 4 million hectares of forest have been lost to rubber plantations since 1993, with two thirds of it in Indonesia, Thailand and Malaysia, they said.’