Gain and loss: Human and environmental wellbeing – drivers of Kilimanjaro’s decreasing biodiversity

PLOS One | https://doi.org/10.1371/journal.pone.0334184 October 29, 2025 6 / 23 single town or portion of a bigger town (Fig 1). The first available Landsat image for Kilimanjaro was from 1976, and the first eligible national population census was from 1978. The other censuses were from 1988, 1998, 2012 and 2022. Due to these limitations (including the availability of cloud-free satellite imagery), we analysed the years 1976 (using the 1978 census data), 2002 and 2022. As corresponding basis for analysing the historical maps from 1911 we used the 1913 census by mission statistics [37]. We used this data to calculate population density (number of inhabitants per km2) for analyses. In addition to population, we applied house building activity as a proxy for economic growth, as no data on per capita income or other indicators of economic development, such as ownership of superior goods (e.g., the number of motor vehicles), were available for the entire period. We used a global satellite-derived built-up area statistic called GHS-BUILT-S from the Global Human Settlement Layer (GHSL) project led by the Joint Research Centre of the European Commission [38]. In parallel, we used our land cover classification to delineate built-up areas (Table 1). For visualisation of these trends we digitized groups of buildings using time series of Google Earth images and the four 1:50,000 Ordnance Survey topographic sheets from the early 1990s (Sheet No: Moshi 56/4, Ol Molog 42/3, Rombo 57/1, Himo 57/3). Land cover

We used the LULC maps for 1976 and 2000 of Hemp [20,23] based on Landsat MSS imagery from 1976, and Landsat ETM+ imagery from 2000. To update the 2000 map to correspond to the 2002 census year, we made adjustments by visually inspecting various very high resolution (VHR) satellite imagery available at Google Earth and ESRI World Imagery platforms, correcting boundaries of agriculture/savanna and other classes. Similarly, we updated the physiographic map by Hemp et al. [28] to reflect the situation as of 2022. Due to the lower spatial, spectral, and radiometric resolution of the Multi Spectral Scanner (MSS) on Landsat 1 satellite compared to the more advanced ETM+ and operational land imager (OLI) sensors on later Landsat satellites, the number of LULC classes had to be reduced to match what was realistically achievable to map from the 1976 Landsat 1 image. This led to a very rough categorisation of the main land cover and ecosystem types (Table 1): Fig 3. Cultivated lowlands on the southern slope of Kilimanjaro. In the foreground, the sugar cane plantations of TPC, surrounded by smallholder fields and – further up – rice fields, March 2025. https://doi.org/10.1371/journal.pone.0334184.g003