Ecology Live with Nathalie Pettorelli- Satellite imagery, time series, fusion and land cover mapping

The British Ecological Society is broadcasting free online talks on the latest ecological research during the coronavirus lockdown period. In this talk, Nathalie Pettorelli of the Zoological Society of London talks about Satellite imagery, time series, fusion and land cover mapping: why is this at all relevant to ecology?

Information on land cover and land cover change is key for ecological research and environmental management. In most cases, this information is derived from satellite information. How this is done shapes the accuracy and comprehensiveness of the information derived. In this talk, Nathalie explores how the recent availability of high spatial and temporal resolution optical and radar satellite imagery has dramatically increased opportunities for mapping land cover at fine scales using case studies in Africa and Asia.

REFERENCES
Lopes, M., Frison, P.‐L., Durant, S.M., Schulte to Bühne, H., Ipavec, A., Lapeyre, V. and Pettorelli, N. (2020), Combining optical and radar satellite image time series to map natural vegetation: savannas as an example. Remote Sens Ecol Conserv. doi:10.1002/rse2.139
Pettorelli, N., Laurance, W.F., O'Brien, T.G., Wegmann, M., Nagendra, H. and Turner, W. (2014), Satellite remote sensing for applied ecologists: opportunities and challenges. J Appl Ecol, 51: 839-848. doi:10.1111/1365-2664.12261
Pettorelli, Nathalie & Schulte to Buehne, Henrike & Shapiro, Aurelie & Glover-Kapfer, Paul. (2018). Conservation Technology Series Issue 4: SATELLITE REMOTE SENSING FOR CONSERVATION. 10.13140/RG.2.2.25962.41926.

[BritishEcolSoc]

There were lots of questions during the seminar that we didn't have time to answer. Nathalie Pettorelli has kindly answered a number of those we received below.

[BritishEcolSoc]

Romain Courault: I read that Kappa coefficient is in discussion to assess land cover assessment accuracy, what do you think?


Nathalie: There certainly is an ongoing discussion as to what is the best way to assess land cover map accuracy; we generally try to use multiple approaches and adapt to the changing recommendations. In this particular situation, however, the point was to look for changes in accuracies between different approaches (fusion, consideration of temporal information, etc) – the choice of the validation metric should therefore not affect our conclusions.

[susmitaaown]

Very relevant to me as my potential PhD requires use of radar and LiDAR

[BritishEcolSoc]

Vanessa Bastos: There are many advanced remote sensing techniques that could be helpful for movement ecology studies, such as multisensor and multiscale approaches. How could we facilitate the absorption of this knowledge by the ecology community and how long do you think it will take for ecology to "catch up" with the up-to-date development in remote sensing ?


Nathalie: I believe the ecological community has become more equipped over the past years to deal with technological advances in remote sensing, with many ecologists now working with remote sensing specialists and vice and versa, and remote sensing having become part of the accepted toolkit. Increased training in remote sensing within ecological degrees is certainly helping this integration.

[BritishEcolSoc]

Jasper Slingsby: This obviously comes with massive computational requirements. Is it feasible to apply at scales larger than the reserve?


Nathalie: Yes – we did it for the whole of Jambi 😊 But you are right; there are computational challenges associated with the use of these approaches at very large scales

[BritishEcolSoc]

Joan Bauza: In relation to long time series, how do you see the possibilities and accuracy of Google Earth Engine and also including a GLCM matrix on the classification process?


Nathalie: Google Earth Engine is being increasingly used and could therefore become an important platform for ecologists. That said, it highly depends on its sustainability and accessibility. Classification processes are always evolving, particularly in periods of rapid progresses in computational power – what will be used as standard in 5 years probably doesn’t exist right now
.

[fernandotroyaandrade]

Very interesting presentation, sadly I could saw it just by youtube. I have 2 questions:
1) How could you combined the resolution of Sentinel-1 and Sentinel-2 for the fusion? Using Sentinel-1 classification as input in Sentinel-2 classification, or another more complex technique?
2) In the case you have few information as ground true data in a study area (1% of the pixels or less), is still possible to use part of that as training set and part for validation (something like 70 and 30% respectively)?

[BritishEcolSoc]

Fredrick Guya: We would like to understand fisheries phenology in Lake Victoria through the study of temporal changes in Lake Victoria's primary productivity. Is it possible to use satellite imagery to study temporal changes in lake productivity?


Nathalie: Yes it is; this is how scientists monitor eutrophication of lakes for example, so the methodology is trialled and tested.

[BritishEcolSoc]

Christina Kallimani: So combining Optical & Radar only resulted in visually better maps but didn't increase accuracy of classification?


Nathalie: For Benin that’s correct yes
.

[BritishEcolSoc]

Chris Thomas: Does low classification success provide evidence of environmental gradients, such as grass-shrub-tree-woodland savana comntinuums, rather than defined categories?


Nathalie: Possibly; one problem with land cover mapping is related to trying to fit a continuous world into defined categories. But we also had other issues that could explain the poor accuracies, which will be worth exploring.

[BritishEcolSoc]

KM Florentine Weber: What product would you recommend to look at land cover before 2000, saying 70ies - 80ies?


Nathalie: Sadly there wouldn’t be much choice; when looking for free and easily accessible imagery, Landsat is pretty much the only option.

[BritishEcolSoc]

Kristina Graves: Why did you pick the land classes you did? Why not classify to lower level, plantation compared to forest, different types of plantations etc.


Nathalie: The distinction of classes is generally informed by local knowledge, the research question to be explored and the expected likelihood that these classes will be mappable from space given the training information at hand. Our choices reflected the broad categories that were deemed appropriate to map the main land cover categories defining the landscape under study, given our research objectives. These classes could have been different if chosen by other people, in other situations.

[BritishEcolSoc]

Katja Philippart: Do you have any experience for combining different satellite information for coastal waters?


Nathalie: Yes; we did combine optical and radar data to map the distribution of mangroves.

[BritishEcolSoc]

Anonymous Attendee: Why did you choose Random Forest for your classification algorithm? Did you consider any other algorithms?


Nathalie: We did, but random forest was deemed the most robust in this case, which is why we went for this one
.

[BritishEcolSoc]

Sarah Chisholm:Is it possible to use these time series and fusion methods to create higher quality landcover maps that cover a global extent?


Nathalie: Theoretically yes, although this would require significant computational power.

[BritishEcolSoc]

Romain Courault: How do you combine optical and radar datasets? Is that a kind of "clustered raster" fusion?

[BritishEcolSoc]

Mayanglambam Singh: Taking an image of a forest whose canopy cover is good but the floor of the forest is dried or exploited and the ones whose canopy is removed but the floor of the forest is intake ...in such a situation the will the images be similar or if so how do we filter



Nathalie: Radar is vey sensitive to the 3D structure of ecosystems, and so this type of information should be able to distinguish forests with dense undercover from forest with a very sparse undercover. I cannot guarantee that this would work, but would recommend considering the combination of both optical and radar information to help differentiate these types of forests.

[BritishEcolSoc]

Andy Cleaves: What ideas do you have to improve the classification of grass and scrub savanna?


Nathalie: One limitation was related to ground truth data access; we would need to improve access to these data to improve on our ability to train our classification I think

[BritishEcolSoc]

David Atkinson: How frequent are images produced for a location from the Copernicus programme?


Nathalie: Sentinels have been designed to give a high revisit frequency of 5 days at the Equator; however, that frequency depends on where your site is located and whether or not you are using optical imagery (which are sensitive to clouds, with clouded images being unusable). In Benin, we were able to use >30 images covering one year.

[BritishEcolSoc]

Tyanita Puti Marindah Wardhani: Do you use any particular software or programming language to proceed the satellite image and do classification?