In an otherwise successful L4A-analysis, roughly 1% of parcels have NULL attributes in the predictions and CT_decl field too.
Most of these parcels are very small or narrow, however there are also some very large parcels with empty crop type predictions. The issue doesn’t seem to be related to any specific crops either.
Is this an indication of failed classification and if so, is there a way to find out the reason? Of interest are especially the large parcels with NULL predictions.
I did not find an explanation to this and was hoping that the great folks here would be able to enlighten me.
This was my first guess as well and indeed most of the parcels with NULL predictions are very small or narrow, but usually they still have S1PIX > 0 or S2PIX >= 3.
Furthermore, there are also parcels with very large and “neat” geometries that also have NULL predictions.
Thus there must be something more to it than parcel geometry.
I would want to reply on this topic, since it seems that this problem has been happening with our crop classification as well.
I have gone over the documentation to verify all things concerning the classification process, so I began filtering out data in the following manners:
Exclude all parcels that hold a class for which there were no accuracy measures in the output accuracy matrix (there were not enough parcels for that class to be classified)
Exclude all parcels that do not satisfy the requirement of a minimum 1 S1 Pixel AND the requirement of a minimum of 3 S2 Pixels
As I was trying to get the system to classify 15 500 parcels over a region of one tile, the system gave me satisfactory results, although 4.5% (circa 700) of the parcels after going through the aforementioned filters remained with NULL predictions even though some of them were very large in their area (and consequentially in number of S1 and S2 pixels)
I tried to visualize where these parcels lie on the map, and even though I could find them randomly in more places, I noticed an obvious pattern of islands in what can somewhat be considered almost a corner of the tile:
Since we saw on some other topic that if the accuracy of a certain class is good, we can try to reclassify the parcels using S2 products only, which we did, but we would definitely want to know whether this behavior happens in more users and could be somewhat of a bug in the system, or it’s a configuration/parameter we might have overlooked.
After looking back to other classification results they could be various reason to be not classified :
too small parcels (less than 1S1Pix and 3 S2Pix)
small crop type with less than 30 parcels
parcels that has problem in their geometries (those parcels are not presents in the results of the classification) :
GeomValid = 0 (invalid geometry)
Duplic = 1 (duplicated parcel)
Overlap = 1 (parcel overlapping another parcel)
I found this last category when merging the “VECTOR_DATA/Parcels_all_with_predictions_0923_1422.csv” file with the /mnt/archive/lpis/site/year/decl_site_year.gpkg file by NewID.