SAM (Spectral Angle Mapper) Classification Question

[agraham1]

Hi Everybody- this is my first post! 

 

I have a specific question regarding SAM and its usefulness in classifying multispectral imagery (not hyperspectral). 

I am thinking of perhaps getting endmember data with a field spectrometer for different plant types, and then using the endmember data to classify a multispec. image. 

However, how would this work? As the multispec sensor is only 5 narrow channels, (RGB, red edge and NIR) and the spectrometer would likely cover all of these channels. If I use spectrometer endmember data, it would only classify in relation to whats available in the bandwidth of the 5 channels.....is what I am assuming. Do you guys think this could be a robust way to distinguish between plant types? 

 

The multispec. sensor is high spatial res. (5cm/pixel). Let me know what you all think . Thanks!

 

Alex

[mamadouba]

Spectral Angle Mapper can be a robust material detection algorithm for hyperspectral sensor data, but you lose the spectral fidelity needed when you reduce the dimensionality to something like 5 bands. It would not do you much good to collect field spectra with an ASD if you are simply going to resample to the sensor band widths, which is what you would have to do. For SAM to be effective, it needs enough n vector angle measurements to determine the spectral similarity between the field collected and pixel based spectra; the highest unit value in gradiens will determine a possible "hit" of the target material. In your case, the spectral profile of different vegetation types for a five band system will look spectrally identical (I.e. High absorption in the blue and green wavelengths, slight reflection in the red wavelength, possible angular difference in the red edge, and very high reflection in the NIR). In summary, you may be able to distinguish plant types using your proposed system but only in general classes such as conifer vs deciduous vs wetland vs cultivated. The red edge may also help determine possible plant stress. However, you will not be able to distinguish plant types using 5 bands if you are referring to plant species, regardless of using field spectra and SAM.

[agraham1]

Spectral Angle Mapper can be a robust material detection algorithm for hyperspectral sensor data, but you lose the spectral fidelity needed when you reduce the dimensionality to something like 5 bands. It would not do you much good to collect field spectra with an ASD if you are simply going to resample to the sensor band widths, which is what you would have to do. For SAM to be effective, it needs enough n vector angle measurements to determine the spectral similarity between the field collected and pixel based spectra; the highest unit value in gradiens will determine a possible "hit" of the target material. In your case, the spectral profile of different vegetation types for a five band system will look spectrally identical (I.e. High absorption in the blue and green wavelengths, slight reflection in the red wavelength, possible angular difference in the red edge, and very high reflection in the NIR). In summary, you may be able to distinguish plant types using your proposed system but only in general classes such as conifer vs deciduous vs wetland vs cultivated. The red edge may also help determine possible plant stress. However, you will not be able to distinguish plant types using 5 bands if you are referring to plant species, regardless of using field spectra and SAM.

 

Hello mamadouba- 

 

Thanks for your reply- I coincidentally checked this thread for the first time in a while and was happy to see that you replied yesterday. 

 

Things have changed slightly since my original post--I still have the same sensor and am now more focused towards identifying locations of wheat stress within a single crop over the course of a season. This of course entails many frequent flights and ground truthing work. I am considering the identification of stress with red edge or some combination of the bands- not really sure yet how I will make this work-lots of reading to do. Mainly I am curious as to which types of stress (fusarium, leaf rust, etc.) I can detect- and if I can detect them early- and if they will be distinguishable. My sensor is high resolution (sub 5cm/pixel)- just wondering if you had any ideas or recommendations for me. Thanks  

[mamadouba]

agraham1,

 

Your research topic is definitely interesting. I think you are on the right track with applying the red edge for detecting plant stress/health, especially deriving a broad-band spectral disease index (SDI) from red edge transformations or ratios.  You may not be able to distinguish what type of plant disease is causing the stress since this is reserved mainly for hyperspectral and even fluorescence imaging sensors, but detecting stress that changes the chlorophyll content of the plant is very likely.  Red edge detection is all about the shape, phase, and amplitude of the spectral signature when comparing healthy and stressed vegetation.  Can you share what sensor and platform you are using (aerial vs. hand held) and the actual band wavelengths?  Many of the red edge indices that I am familiar with have a specific wavelength range to work.  I can also also point you to a variety of related literature if you wish?  Any work by Alfredo Huete will be well worth reviewing.  Not only did he develop the Soil Adjusted Vegetation Index (SAVI), but he has a wealth of experience in researching agricultural systems.

[agraham1]

Yes, exactly. Hyperspectral sensors are still quite far away from being an economically valid solution for farmers and even crop consultants; the sensors are just simply too expensive at the moment. Currently, the best way in detecting specific stress is perhaps having knowledge in the times that certain stressors or diseases typically appear- and to try to detect them in respect of chlorophyll or nitrogen flux at certain windows of the crop's development- is what I am thinking: this is sort of a work around solution. The question is, how early can these be detected with just reflectance data alone. If its not detectable before or at the time a farmer may typically apply blanket treatment of, say a fungicide, then its impractical.  

 

I am using micasense - (first number is center wavelength, second number is bandwidth): 

 

 

475 - 20

560- 20

670-10

720-10

840-40

 

spatial resolution ~5cm/pixel @60 metres. Im mounting this on a UAV and capturing a mosaik of the field crops. 

 

 

I'll check out Huete's work- thanks. If you'd like, you can private message me on this forum to talk more in depth. These bands are fairly narrow- my main concern is the narrow red edge channel. Anyways, thanks for your help! 

Alex