The Truth about Vegetation Species Identification from Imagery
Anonym
Lately the topic of vegetation analysis, in particular species identification, has been coming up like gang busters. I spent a lot of time during my years at the Carnegie Institution department of Global Ecology looking at invasive species in tropical environments—think dense tropical forests. Even with hyperspectral data it was a huge challenge. Vegetation isn’t like minerals. It’s carbon, Hydrogen, nitrogen and oxygen. Absorption features are more driven by leaf composition and health than by a species having unique feature, whereas a mineral has absorption feature unique to it or a family of minerals caused by electron bonds or presence/absence of other compounds. Vegetation of the same species can look vastly different. Even leaves on the same plant can look completely different from one another. In other words in the intra species variability is greater than the inter species variability. Or as I say too often, vegetation is hard.
Take my chlorotic plant here. According to Wikipedia “ chlorosis is a condition in which leaves produce insufficient chlorophyll. As chlorophyll is responsible for the green color of leaves, chlorotic leaves are pale, yellow, or yellow-white. ”One leaf looks vastly different from the rest of the plant, but it’s the same plant. Species ID with spectral data is challenging because there are so many manifestations of how a single plant can appear. To create a spectral library of one species of plant is extremely difficult—there are literally endless permutations of how that species can appear. You really can’t use one spectrum to go find others. If you did, a classifier would go find others like that spectrum. But whatever state that spectrum is in (dry, wet, healthy) does not truly giving you a species ID. That’s not to say you can’t do some discrimination. Conifers and deciduous trees can be separated same for grasses and certain kinds of shrubs with certain data. But sub alpine fir from Douglas fir with only spectra is really challenging.
RayKokaly, USGS SpecLab, has a nice paper on vegetation biochemistry observable from imaging spectroscopy. My key take away from this is you can extract a ton of biophysical information from imagery and use that to study ecosystem health and productivity, but again plants have similar biochemistries, so it’s not like that can be applied to species identification. These chemistries are not unique to a species.
The four things that can be most helpful toward getting to species ID are the following:
- Temporal data: Certain species emerge before others. Plants bloom at different times, have different responses to water stress, die at different times, etc. Even just two images from different times can help and more can provide a higher degree of accuracy.
- Contextual information such as, “you aren’t going to find palm trees growing in the mountain west, so don’t look for them there, look for lodge pole, Douglas fir, and blue spruce. ”Narrow your options and you’re more likely to have some success with a library and target detection. Again I would recommend HSI data. MSI would probably still be confounded depending on the spatial resolution.
- Object based analysis: This method works well with high resolution data, so maybe the spectral signature isn’t enough information, but crown shape, clustering, and texture can all indicate certain species. ENVI’s Feature Extraction Tool can provide this type of analysis.
- Deep Learning: We’re just starting to scratch the surface on what deep learning can do in vegetation. We have a working tool in house and if you have a hard vegetation problem you need solved, give me a shout. This yelp article on identifying food pictures gives a good overview of how deep learning works.
As I said in the beginning, vegetation is hard. Not impossible, but hard. But there are a number of things you can try to improve your results. Invasive species detection, weed infestation, and finding endangered plants are all areas of research in vital need of solutions. If you have a hard vegetation problem and want to talk it through, please reach out to me. It’s one of my favorite topics, in case you couldn't tell.