Seismic Line Ecological Recovery Analysis Closed
RFP51223004
Closed at: 2:00p.m., September 27, 2022
Awarded
This bid has been awarded to Caslys Consulting Ltd..
The Regulator is seeking a vendor to assist with the attribution of linear spatial data for 2D seismic lines in northeast BC.
Download Seismic Line Ecological Recovery Analysis DescriptionDescription
The goal of this project is to attribute linear spatial data for 2D seismic lines in northeast BC with information about vegetation coverage and recovery metrics to support tactical and operational restoration planning for future restoration projects.
Questions and Answers
If you have any questions about this posting, please email procurement@bc-er.ca.
Thank you for your question regarding manual visual interpretation and machine learning. We specifically included the term ‘manual visual interpretation’ in the RFP as we are not confident in the suitability of machine learning for this project. This project is not a pilot or exploratory initiative, so we are seeking a high-level of confidence in the data attributes that are collected. Machine learning is challenging for this project as the seismic line dataset is not very precise. The linear geometry is approximately -/+2m in some places, so alignment with remotely sensed data, such as high-resolution satellite imagery is not perfect. In most places the geometry is not sufficiently precise to classify the line based on the spectral response at the line location. It is possible that machine learning could be used for some of the attributes (like is the line located in a treed/non-treed area), and possibly canopy coverage, and alternate usage, but the shrub information may require visual (human) interpretation. In addition to the alignment issues, there are significant shadows along seismic lines, from adjacent trees, which complicates the success of machine learning algorithms. We are open to considering different innovative proposals but have a strong focus on accurate results. We recognize that ‘manual visual interpretation’ is resource intense, so have assigned a significant budget to this project.
You are welcome to propose different approaches to gathering the data. You can put forward two alternate ‘proposals’ within the proposal if you mean two completely different approaches to the project. Please be clear if there are any cost differences for the two different approaches.
You are welcome to suggest different approaches within the
workflow to analyze and interpret the data. For example, you may use a
different method for considering canopy cover to evaluating alternate land use.
The manual interpretation is important for most of the workflow. We specifically included the term ‘manual visual interpretation’ in the RFP as we are not confident in the suitability of machine learning for this project. This project is not a pilot or exploratory initiative, so we are seeking a high-level of confidence in the data attributes that are collected. Machine learning is challenging for this project as the seismic line dataset is not very precise. The linear geometry is approximately -/+2m in some places, so alignment with remotely sensed data, such as high-resolution satellite imagery is not perfect. In most places the geometry is not sufficiently precise to classify the line based on the spectral response at the line location. It is possible that machine learning could be used for some of the attributes (like is the line located in a treed/non-treed area), and possibly canopy coverage, and alternate usage, but the shrub information may require visual (human) interpretation. In addition to the alignment issues, there are significant shadows along seismic lines, from adjacent trees, which complicates the success of machine learning algorithms. We are open to considering different innovative proposals but have a strong focus on accurate results. We recognise that ‘manual visual interpretation’ is resource intense, so have assigned a significant budget to this project.
The VRI composite layer BC Land Classification is a potential source for if an area adjacent to a seismic line is treed or not treed. The treed non-treed value in the VRI/BC Land Classification System comes from the crown closure attribute in VRI of greater or equal to 10%. The challenge is that sometimes the data is incorrect and there are no trees in the area, due to fire or clearing. We do not expect this project to include a full crown closure assessment of northeast BC, but for there to be a basic validation to ensure that seismic lines are clearly tagged as being in tree or non-treed areas. The intent is to confirm the presence of trees at a high-level validation to confirm if there are trees in treed areas, and no trees in non-treed areas, probably using satellite imagery. This could be from visual interpretation or using a classification approach (you may even be able to use our existing height model for this). It is important to identify the lines that are in non-treed areas as these generally do not have a restoration requirement. A restoration program would not be launched to plant trees on a seismic line if the adjacent area was a grassland, an agricultural field, an open grazing area, a forest fire or a clearcut. We are not specifically interested in the crown closure – just if this is generally an area where a line is in a treed area or not. I think fixing or manipulating the VRI data and its polygons would be a huge task and very time consuming.
Thank you for noting this discrepancy. This is an error and as such, we will honor the larger 500m maximum segment line (for segments in the existing data). Approximately eighty-nine percent of line segments in the existing attributed seismic line dataset are less than 250m – so there are not a significant number of longer segments. It is best if the line length is defined by parameters like land use, vegetation, fire and cutblock so that the line segment is uniform in character (as opposed to having standard breaks in the data). You are welcome to retain the longer segments (truncated at less than 500m) and if you are dissolving the data to combine segments try not to generate super long segments. Segment length should be defined primarily by environmental parameters – line breaks where the line transitions to a cutblock, a forest fire, a grassland, agricultural land so that line character is uniform. The VRI attributes in the data already do this, but you can modify, adjust, or dissolve as you see fit.
We do not expect the line to be segmented wherever there is canopy cover as this would result instead in many small segments. Instead. We are looking for a general sense if there is canopy cover over the line or not. Therefore, we have this index card concept for shrubs and canopy cover. Is the canopy absent, occasional, patchy, or uniform?
You are welcome to ‘clean’ the data to remove small segments. I would suggest simply combining very short segments with adjacent segments. You may want to consider dissolving on certain attributes to remove some of the breaks. This would make more sense for attributes that are less likely to influence line character – like land ownership. You will want to retain the breaks in the data for significant land changes like cutblocks, treed-non treed areas, possibly fires etc. Wetlands are important, but the Freshwater Atlas used to define the breaks is not extremely accurate for edges. When amalgamating please try to keep line segment length less than 250m where possible, except for those already long segments (ref #1).
That is an interesting question. Shadow is challenging. In some instances, you can still see line character even where there is shadow – it is often evident if there is shadow falling on an open clear line (the bowling alley conventional old seismic line), as opposed to shadow on a line that has some texture and growth (lumpy/bumpy with regeneration). I think in some instances it is fair to use shadow but we would not expect this for all lines where there is shadow. If the line is not visible because there is no clear cut in the bush on high resolution data, we would like the data to capture this in the line width attribute. Generally, if the line width class is not visible then canopy cover is “full coverage” and there is no need to capture woody shrub, or alternative use information. Note not-visible needs to be captured in high-resolution data, as nothing is very visible on low-resolution data!
The current crown closure and treed/non-treed attributes from VRI are not 100% accurate. It is a good first pass. We suggest using it to group and then verify. For example, you can filter for non-treed and then check visually or using classified image data to see if there are any obvious treed areas. We do not expect you do a full crown closure estimate across the whole study area. Some remote sensing classification techniques should allow you to separate where there are trees and no trees. You can also try using the existing height model for this. Identifying treed/non-treed is important- it is unlikely we will restore lines in open un-treed areas (like agricultural land, pasture, grassland, high alpine, cutblock etc).
We have access to a spot mosaic that is streamed into our GIS systems. It is pan-sharpened. We can also acquire the raw spot pan and color/IR data for an additional contract charge. The data is usually 1-2 years old. Note that the seismic lines are not always clear on the spot imagery. You may need to consider subscription access to higher resolution data – even if it is slightly older (see MAXAR, UP42, Bing etc)
It’s a 5m data product. It is not perfect but a good general dataset. I think it is 2018.
We subscribe to the Planet lab SPOT mosaic data. Under our subscription we can get access to raw SPOT data Pan and RGBN for a small fee. We already have some older data but could acquire the 2021 data. Note that the SPOT data still doesn’t have sufficient resolution to see shrubs.
Higher resolution data is available from MAXAR and UP42 on a subscription basis. Some of this is pay per use which I find complicated. In conversations with data vendors my understanding is that access for northeast BC would cost about $300k for six months, I think this can be negotiated with data vendors. The coverage is variable by year. It looks like there is good coverage in 2018 – although there may be clouds. I do not have any experience with these subscription services. Bing and Google Earth have good data in most of northeast BC but we would need to check licensing. I would imagine we can use this to derive attribute data if we are not reselling the data. I don’t think the budget is sufficient to fly data or order data from a new satellite constellation.
Yes – we are open to considering multiple proposals from vendors. Please keep the different approaches and associated costs separate.
Amendments
There are no Amendments at this time.
Competition ID: RFP51223004