TITLE:
Prediction of Soil Salinity Using Multivariate Statistical Techniques and Remote Sensing Tools
AUTHORS:
Moncef Bouaziz, Mahmoud Yassine Chtourou, Ibtissem Triki, Sascha Mezner, Samir Bouaziz
KEYWORDS:
Remote Sensing, Spectral Indices, Soil Salinity, Principal Component Analysis, Cluster Analysis
JOURNAL NAME:
Advances in Remote Sensing,
Vol.7 No.4,
December
26,
2018
ABSTRACT:
Soil salinity limits
plant growth, reduces crop productivity and degrades soil. Multispectral data
from Landsat TM are used to study saline soils in southern Tunisia. This study
will explore the potential multivariate statistical
analysis, such as principal component analysis (PCA) and cluster analysis to identify the most correlated spectral indices and rapidly predict salt
affected soils. Sixty six soil samples were collected for ground truth data in
the investigated region. A high correlation was found between electrical
conductivity and the spectral indices from near infrared and short-wave
infrared spectrum. Different spectral indices
were used from spectral bands of Landsat data. Statistical correlation between
ground measurements of Electrical Conductivity (EC), spectral indices and
Landsat original bands showed that the near and short-wave infrared bands (band
4, band 5 and 7) and the salinity indices (SI 5 and SI 9) have the highest
correlation with EC. The use of CA revealed a strong correlation between
electrical conductivity EC and spectral indices such abs4, abs5, abs7 and si5.
The principal components analysis is conducted by incorporating the reflectance
bands and spectral salinity indices from the remote
sensing data. The first principal component has large positive associations with
bands from the visible domain and salinity indices derived from these bands,
while second principal component is strongly correlated with spectral indices
from NIR and SWIR. Overall, it was found that the electrical conductivity EC is
highly correlated (R2 = -0.72) to the second principal component (PC2), but no correlation is observed between EC and the first principal
component (PC1). This suggests that the second component can be used as an
explanatory variable for predicting EC. Based on these results and combining the spectral indices (PC2 and abs
B4) into a regression analysis, model yielded a
relatively high coefficient of determination R2 = 0.62 and a low
RMSE = 1.86 dS/m.