I am an absolute novice when it comes to SAR data. So please excuse me if this is a poor question.
I work in the sugarcane industry and would find a way to monitor the area harvested to date as well as crop still on the ground in our cane supply footprint using SAR. Our crushing season runs from April to December. So any fields that can be harvested in this window need to be at full canopy by April. Up until now I have been using Sentinel 2 to extract the harvestable area as at 1 April with NDVI. Incrementally with each new Sentinel 2 data set, I remove the harvested area from the original harvestable area. This gives us a good indication of yield/ha as well as how much area is still remaining. Up until now this has gone well, but we cloud cover is becoming a challenge.
What would be the easiest approach/technique to monitor crop removal during periods of prolonged cloud cover? Leaf Area Index sounds like it might be the right way to go, but I cannot find any tutorial on producing an LAI image, are there any available? I read a term differential interferometry, would this be suitable considering that sugarcane is about 2m+ tall before harvesting there would be bare earth after harvest?
I think you should start with S-1 GRD-products and look at backscatter time-series of VV and especially VH-polarisation to see whether harvesting is detectable or not.
InSAR-coherence time-series would be the next step to try since bare fields should have a much higher coherence than unharvested fields. But for this to have a real change of working your test site should be in the area that are covered by 6-day repeats.
I have tried with S-1 GRD and you are correct, the backscatter on the VH-polarization does detect the crop removal. However, there seems to be a lot of noise and I would not be able to use this alone on the smaller fields. I will read about InSAR-coherence time-series and see if I can understand it and if I am able to produce something usable.
You could try to reduce the noise in the VH time-series by multitemporal speckle filtering.
Regarding coherence, is your test area in the 6-day repeat zone? If not I think it’s less likely to work. Also, since coherence is computed over a sliding window it is of inherently lower resolution than intensity -> unlikely to work on small fields.
Thanks for the additional info. I will first try the multi-temporal speckle filtering and see how I go. As for the 6 day repeat zone, I don’t even know what that is unfortunately .
Sentinel-1 takes images at 6 day intervals (in the best case). So if you want to get best coherence the shorter the time between both images, the less temporal decorrelation can be expected (coherence goes down).
I am looking at an area in Southern Mozambique which is unfortunately not in the 6-day repeat zone. I tried the Multi-temporal speckle filter without much luck. It works on the larger fields (± 8 Hectares), but there is still a lot of speckle. Where fields are less than 1 hectare, I cannot see detect if they have been harvested.
how much images did you use for the multi-temporal speckle filter? I would at least suggest 12 images (1 per month) so you have robust statistics on surface characteristics: Single or Multi-temporal speckle filter?
If you use GRD data, I see no reason to apply Multi-looking. As well - if your area is comparably flat, you can skip the Terrain Flattening, which surely takes most of the time, and directly calibrate to Sigma0.
I have another question. I am wanting to classify an agricultural area where they grow rice or sugarcane. The rice is dry-land and harvested about 3 times a year and the sugarcane is harvested once. What would be the best way of classifying these 2 crops?
First I would use GRD data. Coherence derived from SLC can help you but since you have different crops with different growing cycles I would say amplitude can give you already good results (plus take into account that you are new to SAR, so start first by understating GRD products and then move to complex data)
The methodology you have proposed before (plus the corrections done by Abraun) seems good enough.
I would suggest you to select S1 data taking into account the growing cycle of both crops. Try to find a time where the structural difference between both is as large as possible (S1 is sensitive to that).
Then you have to think about a classifier. Any experience on that? You can start with some basic approach and then move to SVM, RF etc. (I guess you are doing a supervised classification so you have training/validation data)
Another thing, you may want to derive texture information from your data to increase the number of layers for the classifier.
I also used Sentinel 1 GRD data in crop removal monitoring in my recent study. The processing steps are followed.
Read
Apply Orbit File
Radiometric Calibration
Terrain Correction
Write
Subset
Multi-temporal Speckle Filter
However, one of the reviewer did not agree with the process. Following is his question.
Why did you perform terrain corrections before multi-temporal filtering? Performing the multi-temporal speckle filtering after the terrain geocoding means that the speckle statistic -first and second order - are changed, hence affecting filtering itself.
.
