You can always use D3DXSPRITE, but if you want to go a little deeper than that...

I'd create a list of texture ids on coordinates (ie. int world_array[16][16]).

There are several ways you can sort the ids, but I'd start with just creating the quads in the vertex/index buffer when it comes across the respective id.
Don't forget to count the amount of hits on the id because you don't want to render a quad with the incorrect id.
EX:
// lock vertex buffer
for(int z=0; z<max_ids; z++)
for(int x=0; x<16; x++)
for(int y=0; y<16; y++)
{
if (world_array[x][y] == z)
{
count_id[z]++;

// create quad in vertex buffer
}
}
// unlock vertex buffer

// lock index buffer
for(int x=0; x<16; x++)
for(int y=0; y<16; y++)
// create indices

// unlock index buffer

Use DrawIndexedPrimitive w/ D3DPT_TRIANGLELIST for each texture using count_id[tex_id] as the primitive counter times two (quads are two tris).

Use the list because a fan may create strange looking primitives on the screen, and because not all of those primitives are going to be in a row/column. It may also create the primitives with the incorrect id.

The most important thing you can do is put your small textures together into a big graphics file (.PNG, .TGA, .JPG, .BMP, etc.) so that you don't have to change to a new texture anywhere near as often. Just use texture coordinates to pluck out the sprite you need.

The suggestion of using indexed triangle lists is excellent. Then you can render multiple sprites in one batch, even if they share no vertices and are completely disjoint, so long as they DO share a texture (hence the importance of combining them).

For optimal speed look into using dynamic vertex and index buffers. They are FAR more efficient for sprite-type 2D engines because you can change the vertices every frame (for particles position, animation, color, size, etc.) and still get excellent performance. Research the proper usage of the locking flags as they can be tricky and you have to get it right to get the speed.