LynkerAnalyticsGordonMorris
commented
2 years ago When talking about data quality, I use a helpful metadata element in the ANZLIC Metadata - “lineage”. This allows a description of how the source data was found, created, and developed. While its an optional obligation, I use it in the production of my imagery and encourage it for use in the Artificial Intelligence / Machine Learning environment.
As noted in the ANZLIC Metadata user guide: “This field should be used to indicate whether the data are observations, analyses (re-analyses), forecast (based on initial states including observations), simulations or other sources of data. It could also be used to include the platform/mission in the source of data (e.g. Ship, aircraft, satellite, satellite id). There may be a need to use pairs [source, processing step] to provide additional information. May contain references (e.g. URL) to external information on the processing and source”
I like how it's easy to read, and folk can put as much or as little jargon as they like, as long as it reads well. This would imply it could be machine-read in the future.
I can see "lineage" (or indeed the "abstract") being used at a plot level, as an example "soil condition inferred from fertilizer placement data supplied by SpreadTheGood Ltd on 4 May 2022 under ideal placement conditions with a 1 second time interval (approximately 5m) then rasterized to a 5 metre grid to match existing data holdings").
For example in some of our current work, we have the following lineage statement: "Lineage. Lynker Analytics support to Toitū Te Whenua Land Information New Zealand includes the provision of datasets derived from Artificial Intelligence. This includes layers derived from Machine Learning (ML), which can be described as mathematical models that generate predictions of features from imagery. The pixel segmentation maps generated from the ML models are vectorised autonomously using a rules-based process. The polygons are validated using a human captured test data set with semi-automated quality control while human review and editing are used for a small percentage that fail automated acceptance criteria."
We then have a statement on quality assurance for the ANZLIC Metadata profile: Quality Assurance. Lynker Analytics use a rigorous quality assurance process, including imagery and geometry review. Polygons in several training areas were captured to the LINZ specification by a human operator. This “truth” data set was used to finetune pre-trained models held by Lynker Analytics. A proportion of the objects are withheld from training to be used as quality assurance for the derived building data layer. Separate neural network models are then used in a QA network. Image classification neural networks automate the assessment of captured building polygons against the source imagery. Captured polygons that fall outside accuracy thresholds relative to prior building capture or that are identified by our QA network will trigger final human review. Building placement is within a 4 pixel tolerance of the absolute position of the building with respect to the imagery provided.
Users may refer to the [ Data Dictionary where users can link to] for detailed information about the criteria used to define this dataset.
In the above instance we've not gone into great detail about the type of tools and Lynker Analytics procedures use, but it gives enough for an average end user to assess that this data itself has been through a quality assurance process.
cookeac
When interchanging a map layer, it can be helpful for the recipient to understand the:
Key data quality attributes are:
Many of these require a formal QA review (using ISO 19158) to populate, which I suspect won't happen with most farm-level data, except perhaps for data sets compiled formally at national or regional level.
It seems that the most relevant to farm scale data might be:
Feedback needed: Could those who are interested in data quality metrics please comment with their needs/thoughts?