Memory (RAM) issues using intersect from raster package

Question

I have trouble getting the intersection between two large SpatialPolygonsDataFrame on R. My polygons data represent buildings and administrative boundaries, and I am trying to get the intersection polygons between them.

I understand that the intersect function from the raster package and gIntersection from the rgeos package can do this job (with a few differences) but they cannot handle all my polygons at once (about 50.000 polygons/entity).

For this reason, I have to split my calculation within a loop, saving the result for each step. The problem is: these functions keep filling my physical memory, and I cannot clean it. I tried using rm() and gc(), but it does not change a thing. The memory issue crashes my R session, and I cannot do my calculation.

Is there a way to free the RAM during simulation, within loops ? Or to avoid this memory issue ?

Here comes a reproducible example, for random polygons.

library(raster)
library(sp)
library(rgeos)

#Generating 50000 points (for smaller polygons) and 150000 (for larger polygons) in a square of side 100000
size=100000

Nb_points1=50000
Nb_points2=150000
start_point=matrix(c(sample(x = 1:size,size = Nb_points1,replace = T),sample(x = 1:size,size = Nb_points1,replace = T)),ncol=2)
start_point2=matrix(c(sample(x = 1:size,size = Nb_points2,replace = T),sample(x = 1:size,size = Nb_points2,replace = T)),ncol=2)

#Defining different sides length
radius=sample(x = 1:50,size = Nb_points1,replace = T)
radius2=sample(x = 1:150,size = Nb_points2,replace = T)

#Generating list of polygons coordinates
coords=list()
for(y in 1:Nb_points1){
  xmin=max(0,start_point[y,1]-radius[y])
  xmax=min(size,start_point[y,1]+radius[y])
  ymin=max(0,start_point[y,2]-radius[y])
  ymax=min(size,start_point[y,2]+radius[y])
  coords[[y]]=matrix(c(xmin,xmin,xmax,xmax,ymin,ymax,ymax,ymin),ncol=2)
}

coords2=list()
for(y in 1:Nb_points2){
  xmin=max(0,start_point2[y,1]-radius2[y])
  xmax=min(size,start_point2[y,1]+radius2[y])
  ymin=max(0,start_point2[y,2]-radius2[y])
  ymax=min(size,start_point2[y,2]+radius2[y])
  coords2[[y]]=matrix(c(xmin,xmin,xmax,xmax,ymin,ymax,ymax,ymin),ncol=2)
}

#Generating 75000 polygons
Poly=SpatialPolygons(Srl = lapply(1:Nb_points1,function(y) Polygons(srl = list(Polygon(coords=coords[y],hole = F)),ID = y)),proj4string = CRS('+init=epsg:2154'))
Poly2=SpatialPolygons(Srl = lapply(1:Nb_points2,function(y)Polygons(srl =  list(Polygon(coords=coords2[y],hole = F)),ID = y)),proj4string = CRS('+init=epsg:2154'))

#Union of overlapping polygons
aaa=gUnionCascaded(Poly)
bbb=gUnionCascaded(Poly2)

aaa=disaggregate(aaa)
bbb=disaggregate(bbb)

intersection=gIntersects(spgeom1 = aaa,bbb,byid = T,returnDense = F)

#Loop on the intersect function
pb <- txtProgressBar(min = 0, max = ceiling(length(aaa)/1000), style = 3)

for(j in 1:ceiling(length(aaa)/1000)){
  tmp_aaa=aaa[((j-1)*1000+1):(j*1000),]
  tmp_bbb=bbb[unique(unlist(intersection[((j-1)*1000+1):(j*1000)])),]
  List_inter=intersect(tmp_aaa,tmp_bbb)
  gc()
  gc()
  gc()
  setTxtProgressBar(pb, j)
}

Thank you !

Answer 1

You can consider using the st_intersects and st_intersection functions of package sf. For example:

aaa2 <- sf::st_as_sf(aaa)
bbb2 <- sf::st_as_sf(bbb)
intersections_mat <- sf::st_intersects(aaa2, bbb2)
intersections <- list()
for (int in seq_along(intersections_mat)){
  if (length(intersections_mat[[int]]) != 0){
    intersections[[int]] <- sf::st_intersection(aaa2[int,], 
    bbb2[intersections_mat[[int]],])
  }
}

will give you an intersection_mat of length equal to aaa, and containing , for each feature of aaa, the "indexes" of bbb elements with which it intersects ("empty" if no intersection found):

> intersections_mat
Sparse geometry binary predicate list of length 48503, where the predicate was `intersects'
first 10 elements:
 1: 562
 2: (empty)
 3: 571
 4: 731
 5: (empty)
 6: (empty)
 7: (empty)
 8: 589
 9: 715
 10: (empty)

, and an intersection list containing the list of intersecting polygons:

>head(intersections)
[[1]]
Simple feature collection with 1 feature and 0 fields
geometry type:  POLYGON
dimension:      XY
bbox:           xmin: 98873 ymin: 33 xmax: 98946 ymax: 98
epsg (SRID):    2154
proj4string:    +proj=lcc +lat_1=49 +lat_2=44 +lat_0=46.5 +lon_0=3 +x_0=700000 +y_0=6600000 +ellps=GRS80 +towgs84=0,0,0,0,0,0,0 +units=m +no_defs
                        geometry
1 POLYGON ((98873 33, 98873 9...

[[2]]
NULL

[[3]]
Simple feature collection with 1 feature and 0 fields
geometry type:  POLYGON
dimension:      XY
bbox:           xmin: 11792 ymin: 3 xmax: 11806 ymax: 17
epsg (SRID):    2154
proj4string:    +proj=lcc +lat_1=49 +lat_2=44 +lat_0=46.5 +lon_0=3 +x_0=700000 +y_0=6600000 +ellps=GRS80 +towgs84=0,0,0,0,0,0,0 +units=m +no_defs
                        geometry
1 POLYGON ((11792 3, 11792 17...

(i.e., intersections[[1]] is the intersection between polygon 1 of aaa and polygon 571 of bbb)

HTH.

Answer 2

The example works fine for me (8 GB RAM), after a few changes to the loop. See below. Tese changes are not related to memory use --- you were not storing the results.

List_inter <- list()

for(j in 1:ceiling(length(aaa)/1000)){
    begin <- (j-1) * 1000 + 1
    end <- min((j*1000), length(aaa))
    tmp_aaa <- aaa[begin:end,]
    tmp_bbb <- bbb[unique(unlist(intersection[begin:end])),]
    List_inter[[j]] <- intersect(tmp_aaa,tmp_bbb)
    cat(j, "\n"); flush.console()
}

x <- do.call(bind, List_inter)

Alternatively, you could write the intermediate results to disk, and deal with them later:

inters <- intersect(tmp_aaa,tmp_bbb)
saveRDS(inters, paste0(j, '.rds'))

Or

shapefile(inters, paste0(j, '.shp'))