we have been doing a lot of testing with the L4B but noticed two problems. Even if I combine L4B results based on different buffers and tree masking, I only get results for 36% of the parcels, based on S1/S2. The majority of the mowing is based on S2 only but these mowing events are heavily influenced by clouds. We manually checked the S2 pictures and half of the S2 activities are thin clouds, so we get a 50% error in activity detection in this group. We got very good results when we have an activity based on S1/S2. It feels to me that there are some difficulties in picking up a decent S1 signal for the true activities we detect in S2. Is there an easy way of making the S1 detection more sensitive to change? It feels strange to have so many S2 activities were I cannot get a S1 activity as well. Is there maybe something in the configuration file I can adjust? Thanks for all the help!
to make the S1 processor more sensitive to the SAR temporal trends variations, you can modify the probability of false alarms parameter in the configuration file (section: “S1 Processing”, parameter: pfa = 3.0e-7).
It is set as 3.0e-7, by default. You could try to change it adopting 3.0e-6, at maximum 3.0e-5, but this increases also the possibility to have false detections. I hope it helps.
Thank you Laura, I will give it a try! I was just testing to change some of the other parameters and a lowering of the min_cohe_var also resulted in more S2/S1 results. I haven’t chekced the quality yet but is it reasonable to change this parameter? I guess it relates to the coherence between two S1 pictures?
that parameter (related to coherence window and impact on its RMS) belongs to a set of parameters that have to not be changed (S1 constants). For S1 you can try to change the pfa, as already discussed.
For S2 you could act on the decreasing_abs_th (S2 processing parameters) that represents the threshold for the NDVI decreasing between two consecutive acquisitions. Obviously, concerning S2 performances, it can not solve limitations due to temporal trend gaps caused by persistence of cloud or by clouds not well masked.
My previous post was wrong, I needed to correct something else aswell and now I got a few more S2/S1 results but its still only around 25% of all parcels which get a combined result.
indeed it was strange. The adoption of pfa=3.0e-5 for sure increases S1 detections, but this does not imply that there is an intersection with S2 detection (combined S2\S1 results).
Could you please give more details on the configuration that you used when generating the L4B product? Did you used also L8 products or only S2? Also, the S2 products where pre-processed with MAJA or you used the L2A products downloaded from SciHub (and created with ESA Sen2Cor i.e. you used the downloader.use.esa.l2a = true flag in the config table of the system)?
I used the dashboard to create L4B products, so i am not entirely sure if L8 was used. I included it when I created the site but nore sure if it is part of the default settings for the dashboards jobs. The pre-processing was done with MAJA. In the config file, I changed the settings to:
pfa = 3.0e-5
prod_type_list = SNDVI
sc_fact = 1000
corrupted_th = 0.1
invalid_data = -10000
decreasing_abs_th = 0.10
decreasing_rate_th = -0.000001
increasing_rate_th = 0.9
high_abs_th = 0.75
low_abs_th = 0.4