ML supervised learning generating tiff distort

[tosen1990]

When using(fully filled with data value) tiff toMultibandRaster throws null pointer exception,which I still yet cannot understand why there are some null tile in the RasterFrameLayer. Input geojson Output classified result My rough code

So I think out changing the rf by

    val rr: RDD[(SpatialKey, MultibandTile)] = rf.select(rf.spatialKeyColumn, array($"bd_6")).as[(SpatialKey, Array[Tile])]
      .rdd
      .filter {
        case (_: SpatialKey, b) if b(0) == null ⇒ false  // remove any null Tiles
        case _ ⇒ true
      }
      .map { case (sk, tiles) ⇒
        (sk, MultibandTile(tiles))
      }

And the classified result is distorted, the whole raster is squeezed in upper left direction compared with original one. It seems the tlm meta doesn't behave correctly and I really cannot explain what causes this weird behavor.

[tosen1990]

And I even create a new tlm to generate the tiff,it's better than the one above with only tiles in the right-most and bottom cannot output normally. Code is here For now,I still cannot figuire out what the point is.

[tosen1990]

This can easily reproduce by using demo data L8_LB2-Elkton-VA with tile size 16. The result is: