Stg - 28 july 2016
In order to develop the Sci and Tech group program, please consider how you would respond to the following questions related to goals, functions, and operations of the SciTech group. Your (collective) responses will be collated and formatted as input for a description of our goals, principles and activities. Please take time to comment the one-page project.
- 1 Do you agree with the statement that “Our main purpose as a scientific and technical group is to lay out (and implement) a detailed plan for the production of a 100 meter resolution soil map for Canada.” (?). If not, or if you are only in partial agreement, please provide some thoughts or alternative language to describe what you think might be a better description of our goal.
- 2 Please describe a general approach that you think would be effective for producing a Canada-wide soil map at 100m resolution.
- 3 What do you see as the main obstacle for achieving the goal using the approach you described above?
- 4 What do you think you can contribute to the achievement of the goals of this group?
- 5 What should this group do to ensure that your participation in this initiative is a rewarding activity for you and your employer?
- 6 Proposal for objectives of the CDSM-STG
Do you agree with the statement that “Our main purpose as a scientific and technical group is to lay out (and implement) a detailed plan for the production of a 100 meter resolution soil map for Canada.” (?). If not, or if you are only in partial agreement, please provide some thoughts or alternative language to describe what you think might be a better description of our goal.
- If we want to be precise, the goal could include explicitly the soil properties that we wish to predict (e.g. by order of relevance, feasibility, data availability, etc...). Some objectives can be accomplished over the short (< 1 year), medium (>1 year), or long terms (>2-3 years).
(Bob Mac comments)
- I am not so supportive of making a 100 m map of soil properties the main goal of this group. A map is a one-of product and once the map is completed, then what? I prefer to set as a goal establishing a platform, data sets and procedures, for creating and continuously updating predictive maps of soils for Canada at successively finer resolutions, beginning at 250 m or 100 m and ending at perhaps 5-10 m.
- Rather than make a plan the focus, I would prefer to actually make rapid production of facilities and data sets the initial focus so that we "learn by doing" and not by talking.
- I see the first version of 100m as a visit card and I think it is a realistic goal with thae available ressources. I agree with Bob that we should go for the creation of a platform, data sets and proceduree, but I think that it should be a long term goal. I think the 100 map goal will leave us sufficient liberty to see what working and what's not. If we are planning well and coordonnates our development, we will be able to reuse all the code and data develop in a second phase that could be oriented among the developement of a geographical information and inference system of soil properties.
Please describe a general approach that you think would be effective for producing a Canada-wide soil map at 100m resolution.
1. a good assessment is needed about the availability and quality of soil data in Canada, both in forest and agricultural ecosystems: pedons, soil maps, covariates (DEM, climate, vegetation, etc...).
2. define a number of soil properties that can (easily) be modeled over the short term in order to have quickly some products that we can show to the community (including funding agencies). I think that we need to be strategic here. As an example, I think about modeling soil texture (percent of sand, sill, clay) because it 1) is often measured in the same quantitative way among data sets (no big issues among data bases), 2) soil texture usually has good predictive performances, 3) soil texture is available in most pedon data bases, and 4) it is relevant for land uses and resource management projects.
3. collect and assemble all the covariates available at national scale.
4. identify a number of statistical methods that are relevant to compare: machine learning methods (random forests, boosted regression trees, KNN, neural networks, vector machines, etc...), parametric approaches (linear, additive models), (approximate) Bayesian approaches, kriging, etc...
5. define the type of statistical models that we want to compare: non-spatial vs spatial explicit models, with linear vs non-linear effects, with simple vs interaction effects terms, with hierarchical structures (mixed models) or not.
6. Use model averaging techniques or not.
7. agree on metrics and methods to evaluate the performance (validation) of the model/method combinations: bootstrap, leave-one-out cross-validation, k-fold cross-validation, group-one-out cross-validation, etc... metrics: pseudo-R2, MSE, RMSE, bias, Dawid-Sebastiani Scores, etc...
8. mapping mean and variance estimates. Identify mapping validation techniques as well (not only statistical validation).
9. stressing advantages and limits of the final products
10. steps 1-9 can be split (or not) according to the type of ecosystem (agriculture vs forest), administrative boundaries (e.g. provinces), etc... see what is the most optimal. -
What do you see as the main obstacle for achieving the goal using the approach you described above?
- time and data availability/accessibility.
- IT structure
- Work organization
- Data availability among Provinces
- Covariates availability among Provinces
What do you think you can contribute to the achievement of the goals of this group?
- modeling processes (calibration, comparisons, evaluation and validation processes)
(Bob Mac comments)
- design, calculation of some terrain derivatives at coarser resolutions (250 m, maybe 100 m), review and harmonization of soil profile data.
(JDS suggestions) We need to
- Define a clear structure of the data in and data out
- Organize of script tools around this structure.
- Develop a clear documentation on the wiki.
- Organize code in bitucket structure
What should this group do to ensure that your participation in this initiative is a rewarding activity for you and your employer?
- define deliverables of interest over the short, medium and long terms. These include scientific studies published in peer-reviewed journals, final products freely available for external communities (predictive maps, scripts, methods, etc...) and that are updated regularly.
(JDS suggestions) Julien .... and that are updated regularly. ==> this go in agreement with the developement of a geographical information and inference system of soil properties.
1) I would tend to agree with Bob to some extent and that the 100 m might not be the MAIN goal, but I still think it's an important one and to keep ourselves focused, the 100 m map might be a good place to start that we can do right now. In terms of the rapid production of 'facilities & data sets' as a focus, I don't know what the resource requirements for something like that would be - what is a 'facility'? I would also agree with Bob that there could be less (detailed) planning and more doing and the ideas that I suggested during my visit would get us 'doing' something at least with minimal (formalized) planning.
2) Start with small scale test sites before moving onto national scale mapping where we then have to address big-data issues. In a way, this would be part of the 'planning' but it's more like 'planning-by-doing' before committing to the national scale work.Ideally, by testing a bunch of smaller sites using a variety of modeling types, we might be able to come to a better consensus to a good national-scale approach rather than each of us arguing in favour of our own approaches.
3) I wouldn't say there are main obstacles doing test sites, but really, the challenge is the big data issue and whether or not certain modeling approaches that work effectively for small sites would still work when concerning big data (maybe starting at 250 m resolution might be a less daunting task instead of 100 m)
Another goal, as I said previously, would be the production of a review article - especially for the people that are more tied into academia. A review article talking about the history and future of DSM within Canada would be very valuable as a document that summarizes some of the group's objectives as well as a means of communication to national and international scientists of the Canadian context in DSM. I think that we should try and link the usefulness of DSM in address various environmental and resource management issues, regionally, because those are the things that potential funders of this project will care about. For instance, why do we need DSM for forest & agriculture management, what are some of the biggest issues where DSM can be used as a tool to provide better insight into some of these issues or challenges, etc. Again, I think an inventory of DSM people and work that has been carried out at various scales would be an excellent place to start. I don't think the technical group should be relied completely on for the writing of this document, but rather, it would be beneficial as an exercise by the entire working group. Honestly, I think the Communication team for this entire project is going to be really important as they should be providing the context as to why any of this work is useful and why people should care about what we do or have an interest. Effective communication is sort of the linchpin that connects the technical group (and the people that figure out HOW we do stuff) to the general public and potential funding groups.
Proposal for objectives of the CDSM-STG
My take on objectives (RAM):
1. Set up a platform for communication and sharing so that we (and anyone else who is interested) can collaborate to organise ourselves and jointly agree on key objectives and further activities. Get us all using this platform consistently and recruiting others to join and use it too.
2. Jointly create and agree on a document that sets out the agreed short and long terms goals of the group and establishes norms and rules for interaction and contribution.
3. Find someone to lead a process to identify and contact holders of soil field and lab data (through some combination of targeted enquiry and self reporting) and then itemise their holdings. If possible, acquire original holdings in their raw state, if owners are willing to make them available. Store these as is on a DropBox style repository.
4. Find someone or some organisation(s) that has (have) the technical capabilities, infrastructure and desire to build, host and maintain web infrastructure needed to:
a) collate, host and make available soil profile data for Canada (centralised or distributed) (For an example see: http://www.isric.org/data/wosis )
b) produce, collate, host and distribute environmental covariate data (grid maps) for Canada (or by province). Let's try to share and avoid duplication as much as possible. (For an example of hosting and distributing global gridded covariates see http://www.worldgrids.org/doku.php This will soon be replaced with an improved platform)
c) create, maintain and host a web mapping platform to provide shared functionality to produce maps of any desired property or class, on demand, at any resolution.(ideally). See http://gsif.isric.org/doku.php for an example of how to create and share reusable code and documentation of methods.
d) build and maintain a web platform for accessing, visualising and distributing maps of predicted soil properties or soil classes by province or Canada wide (see SoilInfo as an example, http://soilgrids.org/ http://soilinfo-app.org/#/
5. Jointly create and agree on what measures we feel are needed to assess the quality of any contribution (profile data, covariate data, prediction map, platform) and agree on metrics and standards that contributions need to report so that they can be evaluated in terms of quality and usability. Don't necessarily reject any contribution but assign all offered contributions some objective evaluation of quality or reliability. (e.g. RMSE error relative to independent calibration data for predictions) Publish these as standards.
6. If possible, create a quality control data set of profile observations that can be used to evaluate the accuracy and quality of predictions made for anywhere in Canada (by province, region or country-wide). Ideally, this data set should be restricted and not available for use in developing any models used to make predictions. Use data subsetted from data identified and collected in 3 above to create this validation data set. Use the validation data set to evaluate SoilGrids250m predictions in the first instance and then apply to any other maps produced for any part of Canada, as they become available. Rank all maps produced in terms of relative RMSE error. Merge maps using weighted averages based on relative RMSE error.
7. Agree on and then set targets for producing maps of predictions of soil properties and soil classes, by province, for all of Canada at specified resolutions by specified delivery dates. For example:
a) Up to 280 layers of predicted values at 250 m (based on SoilGrids250m - quality controlled and evaluated for Canada by 2017)
b) Up to 280 layers of predicted values at 100 m (based on Canadian led improvements to SoilGrids250m at 100 m, with additional point data and additional 100 m covariates by 2018. (Probably produced, quality checked and distributed by province)
c) More than 280 layers of predicted values at 25-30 m (based on SRTM or new DEM for all Canada with additional point data and covariates added to b) above. by 2020