Data normalization
Normalized seed data and the methodology behind it
We start with public tech sheets, extract factual product details, preserve the source layer, and build a separate normalized layer so product information is easier to compare across crops, brands, and formats.
Source
Publicly available tech sheets are the source
The records are built from publicly available tech sheet PDFs and the factual agronomic details contained in those documents.
Independent normalization work
The comparison value comes from extracting, cleaning, and normalizing those facts into a consistent schema that can be used across brands.
What normalization does
It aligns inconsistent source formats into one cleaner structure.
Rating scales become comparable
Different source systems use different score ranges and directions. We normalize rating-style fields into a common scale so values can be understood more consistently.
Text gets standardized
Abbreviations, shorthand, and inconsistent wording are cleaned into readable display values where the meaning is clear and stable.
Missing values become consistent
Empty, unavailable, or not-rated values are normalized into a shared missing-value format so gaps are easy to recognize.
Source-only details stay source-only
Not every field should be forced into a numeric score. Some fields remain descriptive because collapsing them would remove useful meaning.
Core rules
Simple rules make the normalized layer predictable.
All normalized values stay stored as strings
Even when a field becomes a normalized rating, it remains stored in a consistent display-friendly form rather than being mixed across multiple data types.
Missing or unavailable values become NA
Values such as empty fields, not rated, or unavailable markers are normalized into NA so missing information is easy to spot and handle.
Score direction stays consistent
When a field is normalized into a rating scale, the target meaning stays the same: 1 = weakest and 9 = strongest.
Some source scales must be flipped
If a source system uses low numbers to indicate better performance, the value is reversed during normalization so the shared scale remains consistent.
Real normalization examples
These are the kinds of conversions happening behind the scenes.
1 to 9 rating systems
If a source already uses a 1 to 9 system in the correct direction, it can stay as-is. If the direction is reversed, the score is flipped to match the common comparison scale.
1 to 7 rating systems
Some source systems use 1 to 7 bands. Those are mapped into the shared 1 to 9 scale using a fixed conversion table so the normalized layer remains consistent.
Percentage ratings
Percentage ratings such as 40, 50, 60, 70, 80, and 90 are aligned to the shared comparison scale when appropriate.
Text-based rating bands
Some source systems describe performance with phrases instead of numbers. Those phrases can be mapped into the normalized rating scale when the source meaning is clear enough to support it.
Readable field cleanup
Not everything is a score. A lot of the work is making descriptive fields usable.
Abbreviations become readable text
Short codes and inconsistent shorthand are expanded into clearer display values when the meaning is known, stable, and useful for comparison.
Plant characteristics stay descriptive
Fields like plant height, canopy type, growth habit, flower color, pod color, ear flex, ear placement, and related traits remain readable descriptors instead of being forced into generic scores.
Continuous measurements stay direct
Some numeric values are already meaningful as captured and are not converted into score bands. In those cases, the normalized layer preserves the direct value.
Source descriptors remain visible
Some disease or resistance fields carry specific source context that should not be flattened into one simplified rating. Those values stay as descriptive source-backed fields.
Where interpretation is applied
Normalization is explicit, limited, and documented.
What we interpret
Normalization may require deciding whether a scale is reversed, mapping text bands into shared rating bands, expanding abbreviations, or deciding when a direct measurement should stay unchanged.
What we do not do
We do not treat normalization as rewriting the source data or hiding the original format. The normalized layer is an independent comparison tool, not a replacement for the raw capture.
Quality control
Unresolved values are surfaced instead of silently guessed.
Normalization reporting
The process generates reporting to catch unknown score values and unmapped vocabulary so incomplete rules can be fixed directly.
Why that matters
It is better to flag an unresolved value than to force a weak conversion. That keeps the normalized layer more trustworthy and easier to improve over time.
Comparison with traceability
The end result is a cleaner comparison layer backed by source-aware rules, preserved raw data, and a process that can be reviewed and extended as new formats appear.