you can import the measurements as points (by CSV or Shapefile) and use the correlative plot view to build a regression.
The results of soil moisture depends on the size of your AOI and the number of soil moisture sensors.
If you have to study an area, let’s say 100km x 100km using satellite data, you are gonna need a network of sensors scattered around your AOI (might be 100, 200, e.t.c). These sensors will provide you with accurate measurements of water content and their geographic location using GPS equipment.
In order though to estimate the the soil moisture content for the whole satellite scene, you will probably need to apply machine learning algorithms (Neural Network, SVM, e.t.c) in order to make predictions of the soil moisture content for the whole satellite scene.
I agree. There is no simple relationship. But applying ML would require several spatial input parameters. One single SAR image is not sufficient. I therefore like the idea discussed here to calculate an index value (based on multi-temporal SAR data) which can be directly related to the measured moisture variations.
yes agree, on SAR image is not sufficient. We would need other auxiliary data (optical data potentially as well)
the simplest way to do it is the one you described.
Means “wet weather conditions“ that it must rain while the acquisition is ongoing or short before the acuisition? I thought precipitation while radar meassurements leads to unwanted noise?
I’d say shortly before the image acquisition would be the ideal case, but certainly not during the measurement.
Not sure you are aware of the Copernicus product from the copernicus services on Soil moisture from Sentinel-1 that will be soon coming.
@johngan Thanks for your detailed replies. I’m implementing the change detection technique, as you described in your post from Oct 18 however on Sentinel 1 multi temporal imagery, not ASAR WS imagery as the paper describes.
A few questions I have are,
- is the VH polarization of S1 more appropriate than VV for change detection?
- is it necessary to compute angular correction for S1? since each pixel has a characteristic local angle that is unchanged throughout the year over my study area.
- lastly, is it possible to produce a high resolution (10 m) soil moisture index map using Sentinel 1 imagery? the paper you cited only mentions spatial resolution once, i.e. 75 m ASAR pixels are averaged to a 1 km grid to allow low sensitivity pixels to be processed. In my case multi-temporal imagery (3 yrs) gives me a high sensitivity > 3 dB over the arable and barren land of my study area, hence is it necessary to average the resolution of S1 images or can I retain the 10m resolution?
The problem with VV polarization is that the vertically polarized microwaves are affected by vegetation (HH pol is less sensitive to vegetation) and the penetration through soil is minimized. VV polarization can detect the differenceswithin the vertical vegetation structure from various growth stages. As M.R.Saradjian and M.Hosseini, 2011 mentioned, after experiments, HV cross-polarization proved to be more accurate than HH or VV polarization for areas with some vegetation.
The backscatter coefficient as a function of incidence angle declines. The larger the incidence angle the lower the sensitivity of the microwave. Hence, we can identify larger errors in our measurements at large incidence angles due to the low sensitivity. It is recommended to perform an incidence angle normalization at approximately 30 degrees.
You can perform soil moisture analysis using Sentinel-1 data. there is no need to down-sample your 20m resolution to a coarser one
Thanks @johngan for that reply
to your second point, incidence angle normalization is necessary to correctly analyse spatial distribution of soil moisture in areas where topography varies? Is it still important if I am only concerned with per pixel changes in soil moisture over time.
Further I looked into the forums regarding incidence angle normalization, specifically these posts,  and  and this paper my interpretation is that the normalization process gives us gamma-naught which is sigma-naught/cos(local incidence angle). is this correct?
The papers you are looking at refer to
terrain flattening which is different from
incidence angle normalization.
Terrain flattening is applied to an area with local tomography. Due to the terrain variations, your back-scatter intensity is affected accordingly leading to misleading results. Hence, by applying terrain flattening we can remove some of the ambiguities introduced due to topography.
In terms of incidence angle normalization, it is required when we do measurements over flat areas. The backscatter intensity drops as the incidence angle increase (figure below)
If you look at the graph, you see that the larger the incidence angle the lower the amplitude. Looking at the images, we see how eh brightness varies from left to right. So, the brightness variation does not have to do with the properties of the materials (as someone would expect), but instead, with the incidence angle range. For accurate results, we need to correct for this.
Have a look at this paper here
So, if your area of interest is flat, then you need to do this step.
In case your area of interest has a lot of tomography, then you can apply terrain flattening
Not all materials scatter in an isotropic manner, therefore the incidence-angle can affect backscatter. In other words, part of the intensity variation can be signal.
thanks johngan, that reference helps.
so the squared cosine correction should work over land.
and if I have to perform both terrain flattening and local incidence angle correction does the sequence matter?
first terrain flattening and then incidence angle normalization or vice versa?
Yes I agree, part of the intensity variation in the SAR image is signal.
The point i am trying to make is that Incidence angle variation can affect our results when working, for instance, on SAR image classification over flat areas (e.g sea ice types, land cover, e.t.c).
When materials of the same type are distributed across the SAR image , the backscatter intensity of this material located in near range will vary with the one located in the far range of the SAR image. Hence, due to the differences in intensity, this will end up having high misclassification errors.
You have to choose one or the other and not do both.
Hi，Jojngan.Do I have to add a step （Apply Orbit File）？
You just need to terrain correct your time-series data
Hello！You said：“ In my case multi-temporal imagery (3 yrs) gives me a high sensitivity > 3 dB over the arable and barren land of my study area”.How Can I check the sensitivity by SNAP. Thank you！
Dear johngan, I’m currently working on my diploma thesis in which I compare different change detection approaches for soil moisture retrieval. Two of the compared methods are those proposed in the article you reference in this thread ( Synergetic Use of Sentinel-1 and Sentinel-2 Data for Soil Moisture Mapping at 100 m Resolution). I understand the logic of both the methods and I was able to reproduce all of the steps. However, in the case of the first of these methods I got confused in the last step - applying the following formula:
Could you please explain what does the last element of the formula ( Mvmin (i,j,d) ) represent?
Mvmin(i,j,d) refers to the driest value (Mvmin) acquired by the Sentinel-1 over a date
In the equation, in order to derive
Mvmin, it requires a time series of S1 images from which you can derive the min and max values. Hence,
Mvmin(i,j,d) refers to Mvmin of a specific day instead of deriving Mvmin from a time series image acquisitions.