Vegetation Effect on Soil Moisture Retrieval from Active Microwave Data
Tarendra Lakhankar (Ph.D. Student) tarendra@ce.ccny.cuny.edu, Hosni Ghedira ghedira@ce.ccny.cuny.edu, Reza Khanbilvardi khanbilvardi@ccny.cuny.edu
Dept. of Civil Engineering, The City College and the Graduate Center of CUNY, Convent Avenue at 138th Street, New York, NY 10031 Phone: 212 650 8598
1. Soil Moisture and Microwave Data
3. Soil Moisture, SAR, Optical Depth and NDVI Data (July 12, 1997)
R e g re s s io n A n a lys is o f S A R B a c k s c a tte rin g a n d S o il M o is tu re
• The brightness temperature and backscatter coefficient from
microwave remote sensing data is related to soil moisture
based on dielectric properties of soil and water.
SAR Backscattering (DN)
120
• The relationship between backscatter coefficient and soil
moisture becomes non-linear and complex by the presence of
vegetation cover present on ground surface.
100
80
60
U pper 95% C onfidenc e Lim it
40
Low er 95% C onfidenc e Lim it
U pper 95% P redic tion Lim it
20
Low er 95% P redic tion Lim it
0
0
5
10
15
20
25
30
35
40
S o il M o is ture (% )
• The vegetation cover is the function of normalized vegetation
difference index and vegetation optical depth.
R a n g e o f B a c k s c a tt e r in g v a lu e s (D N ) v s S o il M o is t u re C la s s e s
Mean SAR Backscattering (DN)
90
• The microwave energy backscattered from vegetated terrain is
highly dependent on the dielectric constant of the soil and
vegetation, which is in turn directly affected by water content.
• Radar energy is able to operate independently of cloud cover,
smoke and solar illumination which hardly limit the use of
optical sensors such as SPOT and LANDSAT.
80
70
60
Vertical line shows ± one
standard deviation of mean
50
40
0 -1 0
11 -2 0
2 1 -3 0
Soil Moisture
Soil Moisture from
(Unsupervised Classification)
ESTAR
31 <
S o il M o is tur e C la s s (% )
2. Data Acquisition and Study Area
SAR Image
Std Deviation
Mean
Homogenity
Neural Network
Output (supervised)
NDVI
Optical Depth
5. Neural Network Optimization
4. Neural Network Classification Methodology
A best neural network is described as a combination of better accuracy, training
stability and processing time. Those three parameters depend on the network size and
the network complexity, which are the function of the number of network layers and the
number of nodes in each layer.
SAR Data (25 x 25 m resolution)
ESTAR Data (800 x 800 m resolution)
To optimize the internal configuration of the neural network, the same network was run
25 times for each architectural configuration. The results showed that, by using two
hidden layers with an equal number of nodes, the standard deviation of the 25 runs is
very low. Further, the increase of the number of hidden nodes increases the training
and the classification time without any improvement of the overall accuracy. The
results indicate that better classification accuracy was reached when the number of
hidden nodes is the same in each hidden layer.
SSM/I Data (25 km x 25 km)
38°30’ N
Soil Moisture Data
165 km x 495 km
(Res. 800 m)
37°30’ N
When using a single layer, the number of nodes should be greater than the number of
input data to get reliable results. Further, the variance of overall accuracy in 25 runs
becomes more stable when the number of hidden nodes is less than 22.
Study Area
140 km x 300 km
~ 70 SSM/I pixels
36°30’ N
SAR: July 12, 1997
35°30’ N
SSM/I Data
350 km x 350 km
(Res. 25 m)
98°00’ W
96°30’ W
95°00’ W
93°30’ W
8. Preliminary Results and Discussion
A subtractive clustering method is also used to estimate the soil moisture from a
combination of input data. The subtractive clustering method assumes each data point is
a potential cluster center and calculates a measure of the likelihood that each data point
would define the cluster center, based on the density of surrounding data points.
(NAD 27)
0.54
0.50
0.40
0.45
0.50
0.55
0.60
0.65
0.70
0.75
0.80
0.85
Radii
Methodology applied in soil moisture estimation
< 0.15
Class 2
Class 3
35
4
Class 2
26
43
187
29
242
Class 3
6
34
125
165
Nil Pixel
8
15
27
50
75
240
185
0.69
Total Pixel
60%
Class 3
Class 1
36
1
3
40
Class 2
21
173
19
213
Class 3
4
29
116
149
Nil Pixel
14
37
47
98
Total Pixel
75
240
185
0.65
30%
Nil P ixel
20%
Correct pixel
Incorrect pixel
0.000-0.010
0.011-0.020
0.021-0.030
0.031 <
Fig: Effect of optical depth class on classification accuracy
Effect of optical depth on classification accuracy
0.40
9. Future Approaches
0.30
0.20
Nil Pixels
Correct Pixels
0.40
Total Pixel
40%
0%
0.00
Class 2
50%
10%
Incorrect Pixels
Class 1
0 .35 <
70%
0.50
0.10
Confusion Matrix (Threshold = 0.6)
0 .25 - 0 .35
Optical Depth Class
Total Pixel
4
0.15 - 0.25
NDVI Values
0.60
Percentage
Class 1
0.40
Effect of NDVI on classification accuracy
Effect of the selected threshold on the
overall classification
Class 1
0.50
0.30
¾ The prediction made by neural network is higher than fuzzy logic in several runs of the model, but
we found that the prediction made by fuzzy logic is more stable in nature.
Confusion Matrix (Threshold = 0.5)
0.60
¾ The fuzzy logic model was also used to predict the soil moisture using the similar input data and
gave a low variation in soil moisture estimation accuracy.
¾ These results gave us a thought to couple these two techniques to come out with better and
reliable method such as “neuro-fuzzy” to improve the soil moisture estimation accuracy.
Accuracy assessment was carried out using confusion
matrices generated from the comparison between real
values (truth data) and predicted values (estimated data).
The two following matrices show that the increase of
threshold limit from 0.5 to 0.6 leads to more Nil pixels in the
final soil moisture map. However, it makes more sure that
the classified pixels have been correctly classified.
0.90
Output:
Soil Moisture Data
0.70
¾ The neural network model shows higher potential to estimate soil moisture. However, high
variation in soil moisture estimation accuracy has been observed at different runs of NN model.
0.58
Neural Network Model
Simulation
Confusion
Matrix
¾ Adding the vegetation data is important to improve the accuracy at dry soil condition, where the
dominance of vegetation water content is higher.
0.62
Optical Depth
and NDVI
Textural data
SAR image
Training
Validation
¾ The correlation between SAR backscattering and soil moisture is better with higher soil moisture
content, which validates the study carried out by Wang et al. (2004) stipulating that the
dominance of soil moisture on backscattering is higher at wet soil than dry soil.
7. Threshold Limit and Confusion Matrix
To avoid that the network forces the classification of all the
pixels, we have introduced a threshold (value between 0
and 1) to decide if a class will be assigned to the input pixel
or if this pixel will be considered as unclassified. Thus, a
pixel is considered unclassified if all output values are
lower than this threshold; otherwise the pixel is assigned
the class corresponding to the neuron with the highest
value. In this project, the threshold value has been varied
from 0.4 to 0.7.
ESTAR Soil
Moisture Data
¾ High optical depth and NDVI values generate more confusion in the NN prediction.
7 input variable
5 input variable
0.66
¾ The influence of various parameters (NDVI, Vegetation optical depth, textural data) can be better
understood by classifiers like neural network and fuzzy logic.
¾ The areas with lower NDVI values showed better classification accuracy due to less contribution
of vegetation to the backscatter.
0.70
Accuracy
The study area is located in Oklahoma, USA (97d35’W, 36d15’N). One Radarsat-1 image
acquired on July 12th, 1997 by ScanSAR Narrow Mode at an incidence angle range of 20°-39°
with a resolution of 25 m was used in this study. The soil moisture and other vegetation data
(NDVI, vegetation water content, vegetation b parameter, etc) were collected during SGP97
mission have been used in this research. The SGP97 experiment was a large, interdisciplinary
experiment carried out in 1997 with the objective to test formerly established soil-moisture
retrieval algorithms for the ESTAR Instrument (Electronically Scanned Thinned Array
Radiometer) L-band passive microwave radiometer at 800-meter resolution.
Calibration and Geocoding
¾ The additions of optical depth and NDVI information to NN model have significant effect (increase
the accuracy by ~6-10%) on the final soil moisture accuracy.
The algorithm:
o Selects the data point with the highest potential to be the first cluster center.
o Removes all data points in the vicinity of the first cluster center (as determined by
radii), in order to determine the next data cluster and its center location.
o Iterates on this process until all of the data is within radii of a cluster center.
The variable radii is a vector of entries between 0
and 1 that specifies a cluster center's range of
influence in each of the data dimensions, assuming
the data falls within a unit hyperbox. Small radii
values generally result in finding a few large clusters.
The Good values for radii is estimated as 0.55 based
on various run of model for a set of input data.
SAR Data Acquisition
¾ This study demonstrates a promising capability of neural network to retrieve soil moisture maps
from active microwave data.
Percent Accuracy
34°30’ N
99°30’ W
6. Fuzzy Logic Method
SAR Data Characteristic:
Product: ScanSAR Narrow
Beam Mode: (B) W2 S5 S6
Projection: UTM Zone 14S
Earth Ellipsoid: Clarke 1866
Percent of Pixel
(Res. 25 km)
0.45
0.50
0.55
Threshold Limit
Soil Moisture Classes:
Class 1 : < 10 %
Class 2: 11-10 %
Class 3: 21 % <
Nil Pixel
0.60
0.65
0.70
¾ Development of soil moisture retrieval algorithm using a
combination of parametric and non-parametric tools such
as maximum likelihood, neural networks, fuzzy logic etc…
¾
Assess the effect of normalized difference vegetation
index (NDVI) and vegetation optical depth on the retrieval
of soil moisture from microwave data.
¾ Production of qualitative and quantitative soil moisture
maps with different levels of accuracy.