Figure 3: Socioeconomic and land cover variables in the first two canonical dimensions. Canonical loadings for each original variable in the first two canonical dimensions. Variables definitions are listed in Table 2.

Mentions: The first dimension of the canonical correlation captured the strongest association between the socioeconomic and land cover variables. The ordination of census tracts along the first canonical dimension captured a gradient from slum areas (positive loadings) to higher income areas (negative loadings) (Figure 3). This gradient can be inferred from the canonical loadings associated with this dimension: positive for crowding and corrugated roofs, negative for pavement, white-painted roofs and income (Figure 3, Table 5). Based on the canonical coefficients of the land cover variables, corrugated roofs and pavement variables most strongly influenced the first dimension (Table 4). Of the socioeconomic variables, income and bathroom variables had the most influence in the first dimension (Table 4). Based on the canonical loadings, the second canonical dimension was mostly driven by socioeconomic variables such as bathroom, garbage and crowding (positive) and income and water (negative) (Figure 3, Table 5). The majority of households in Salvador have access to sanitation and piped water, which may account for why these variables explain less of the overall variance. Based on this lack of infrastructure, communities that are highly loaded in this dimension are likely newly-squatted areas. The land cover variables did not weigh strongly in this dimension, however dimension 2 still identifies important deprivation characteristics (Table 5). The third dimension explained only a 0.11 of the variance and differentiated high-income areas with high proportion of white-painted (typical commercial areas in Salvador) or tiled roofs from lower-income areas.