Historical mapping of fragmented cropland in Africa: a case study in the Copperbelt region, DRC (2000-2023)

Precise and dynamic cropland maps are essential for research and practical applications, such as soil fertility assessment and crop production monitoring. In Africa, continued population growth and increasing land-use pressures make the need for reliable land cover information greater than ever. Earth observation missions provide timely, large-scale data, and recent efforts have produced high-resolution (30m or better) global and continental cropland/land use land cover (LULC) maps. However, low consensus among these maps for cropland predictions in Africa largely limits their downstream local applicability despite reported high accuracy.

Here, we conducted a case study in the Copperbelt region (DRC), where most cropland is managed by smallholders within fragmented landscapes. Our objectives were to: (i) map cropland dynamics from 2000 to 2023; (ii) evaluate the accuracy of both static maps and dynamic changes (cropland gain and loss); and (iii) compare the performance of our maps with five existing high-resolution (10/30m) cropland/LULC products. We used the Landsat Analysis Ready Data (ARD, 30m resolution) to derive eight annualized NDVI time series (aggregated every three years from 2000 to 2023) as input data. A binary random forest classifier was trained on over 6000 cropland and 12000 non-cropland reference samples collected from 2000 to 2023. Independent validation for the static map in 2020 showed an overall accuracy (OA) of 91.2%, outperforming all existing maps (OA: 60.2%–83.2%). While effective at identifying large cropland fields, most existing maps overlooked small, fragmented fields, leading to an underestimation of cropland area up to 91%. Based on our predicted maps, cropland area increased by 20% from 2000 to 2023. Two drastic short-term changes were observed: a surge from 2017 to 2020 (+57%) and a decrease from 2020 to 2023 (-37%), reflecting intense deforestation and urban expansion in the two periods. However, accuracy for detecting cropland gain (71.9%) and loss (53.3%) was limited, likely due to the 30m resolution being insufficient to separate smaller fields, particularly near suburban built-up areas where cropland is often interspersed with single houses.

In conclusion, existing global and continental cropland/LULC maps remain inadequate for regional use in Africa, where fragmented cropland is prevalent. Improving these maps requires region-specific training samples, particularly from smallholder farms. Moreover, detecting cropland changes remains challenging, and higher-resolution imagery may present an opportunity to better monitor the dynamic landscapes.