Comparison of ​​cropland area according to land cover data and ARMSTAT data

A comparison with ARMSTAT data on cultivated areas was conducted for four land cover datasets – ESRI, ESA, GLAD, and Dynamic World (GLC_FCS30D was excluded from the analysis as it differs most significantly from the Government-reported data). For comparison, we also used 2022 Government data on the area of cultivated land in Armenia.

Land cover datasets and Government-reported data were compared with three ARMSTAT indicators for the same year as the land cover data:
1) Arable land (Arable in graphs) , that is, an area intended for cultivation, but not necessarily used every year. 
2) Annually cultivated area (Cultivated in graphs), that is the sum of annually plowed area, the area of fruit and berry plantations (including greenhouses, hothouses and inter-row fruit-bearing plantations), and vineyards.
3) Annually plowed area (Plowed in graphs) that is plantations of grains and leguminous crops, potatoes, vegetables and melons.

According to ESRI, ESA, and GLAD datasets, the cropland area in most marzes is smaller than the area of arable land but larger than annually cultivated area reported by ARMSTAT. The cropland area identified by DW exceeds the arable land reported by ARMSTAT in almost all marzes, except for marzes Lori and Tavush (Figure 1). The cultivated area reported in the 2022 Government data exceeds the arable land area in all marzes (GOV (A) in Fig.1). If the cultivated area within settlements is excluded, the difference with the ARMSTAT data becomes smaller (GOV (B) in Fig.1).

The cropland areas identified by all datasets exceed the annually cultivated area reported by ARMSTAT, except for the GLAD data in marzes Ararat and Armavir.

Figure 2 provides a more detailed breakdown by marz.

Figure 1. Difference between areas of croplands in tested datasets and ARMSTAT data on arable lands, annually cultivated, and annually plowed areas (dataset data minus ARMSTAT data)


Figure 2. Difference between areas of croplands in tested datasets and ARMSTAT data on arable lands, annually cultivated, and annually plowed areas (dataset data minus ARMSTAT data) across marzes

The fact that in ESRI, ESA, and GLAD datasets the cropland area is smaller than the area of arable land but larger than annually cultivated indicates that these datasets classify a part of lands designated for cultivation but not cultivated during the reference year as croplands. The area of land designated for cultivation that was left uncultivated in the given year is equal to Astat-Cstat, where Cstat is cultivated area in ARMSTAT data; Astat is arable area in ARMSTAT data. Thus, the share of uncultivated fields that are identified in ESRI, ESA, and GLAD datasets as croplands can be defined as U=(C-Cstat)/(Astat-Cstat), where C is cropland area in a dataset. Across the marzes, this figure varies between 0% and 100% (Fig. 3). In cases where the cropland area from land cover datasets exceeds arable land area reported by ARMSTAT, this indicator exceeds 100%. This is most evident in the ESA and ESRI data for the Ararat and Armavir marzes, where these datasets estimate the cropland area to be 20–40% larger than the arable land area reported by ARMSTAT, while approximately 90% of the arable land in these marzes is annually cultivated. The cropland area in all datasets exceeds the annually plowed area. The Government data exceed both annually cultivated and annually plowed area reported by ARMSTAT.

Figure 3. The share (%) of uncultivated arable land that is classified as cropland by the land cover datasets

Similar to the comparisons with Government-reported data, a preliminary overall indicator for assessing land cover data accuracy can be the total discrepancy between cropland areas in datasets and ARMSTAT data which is the sum of absolute area discrepancies (by modulus, regardless of sign) across marzes (Figure 4). Overall, ESRI, ESA, and GLAD datasets show a similar total discrepancy from the ARMSTAT data, DW shows a substantial overestimation of cropland area.
Figure 4. Total discrepancy between cropland areas in datasets and ARMSTAT