GEOG 6110 Environmental Analysis through Remote Sensing

Utilize Remote Sensing Techniques to Detect Areas of High Soil Salinity

Introduction

Soil degradation is very seriously affecting agricultural production across the world. (Gorji, Qadir, Scudiero 2015, 2017)  There are many different causes of soil degradation, both physical and chemical. Among the leading chemical contributors is salinization, or increasing the level of soluble salt in the root layer of soil. (Rattan, 2004) Soil salinity, when it reaches a certain threshold, can be very detrimental to agricultural production.  Within the US, the National Resource Conservation Service (NRCS) and US Department of Agriculture (USDA) offer much literature on the subject of testing and monitoring salinity levels (USDA-NRCS).  Most of these methods, however, require labor intensive ground tests over repeated periods of time. This is not feasible to conduct an analysis on the impact on of soil salinity of a national or regional level, agricultural production tends to cover large areas of land. (Gorji, Scudiero, Allbed) Determining if salinity levels can be sensed remotely would be very beneficial and highly useful.  

There are many contributing factors that lead to increased levels of salinity in soil.  These factors can be grouped in two categories, natural and human or agricultural factors.  Natural contributors include, eroding parent material that contains salt minerals, temperature, rainfall, and flooding.  Agricultural or human contributors are related to poor agricultural practices. (Rattan, Gorji, Scudiero 2015, Scudiero 2015, Al-Khaier) Over fertilizing soil to increase crop yields and poor irrigation practices can have a powerful impact on salinity levels.  Fertilizer, which can be beneficial to production in appropriate amounts, can also lead to increased deposits of salt when over utilized. Irrigating crops with poor quality water, especially in times of little rainfall, can also cause increased levels of salinity in soil. (Singh, Shrivastava) Without the rain to dilute the soluble salts, soil can reach a harmful level of salinity.  Potential for crop yield reduction varies with the type of crop, but on average begins around 4 decisiemens per meter (ds/m). (Scudiero et al. 2015, San Joaquin River Restoration Program) High salinity concentrations impact crops at every level, from germination to growth. Increased amounts of soluble salt require the plants to exert more energy in absorbing water, which detracts from the overall health of the plant. (Al-Khaier) 

The area chosen to focus on for this study is in the north east corner of Cache County, Utah, see Figure 1.  A study conducted by Utah State University in 2008 has provided a large number of EC tests throughout a densely agricultural region of the area. These test locations are also shown in Figure 1, along with the underlying band of soil type such as Trenton-Jordan-Cache. (Cardon, Hawkes 2008) Utah is within the region of the US affected by naturally saline soils, by which the parent material for the soil contains higher than normal levels of sodium salts. (Gorjiet al.)  The climate in Utah is usually classified as arid or semi-arid. Areas of higher elevation tend to receive more rainfall than the valleys. (Fisher) The study area is located in a region of lower elevation between mountain ranges.

Data

Electric conductivity tests were chosen as ground truth data primarily for the widespread acceptance as a measure of soil salinity as well as for the availability. (Rhoades et al.)  A study conducted by the Utah State University in 2008 provided the necessary EC results of 75 different test locations. Several of the test locations are at various spots in the same field, the distribution of which is over a large area in Cache County.  The method of EC testing used was the Bureau Cup method, which is frequently used for field testing as it is both practical and inexpensive. (Cardon, Rhoades et al.)  The “cup” in this method is a 50 cm cylinder which is constructed from hard rubber. Two large electrodes of nickel-plated brass measure the electrical currents in the soil paste that is put in the cell. This method has been used for well over 100 years, and while criticisms have been made against the method, it is still quite effective at producing reliable results for EC levels in soil. (Rhoades et al.)

It was important to determine if the any of the results, either from remote sensing analysis or ground truth EC tests, were a product of natural occurrence or potentially caused by agricultural practices.  Soil data from the NRCS was utilized for this purpose. Vector data for the study area with soil type attributes were used as a comparison on all levels of analysis. The data shows several different bands of soil in the study area, see Figure 1.   Most of the tests were conducted in the Trenton-Jordan-Cache and Quinney-Lewiston-Layton-Kidman soil bands, see Figure 2. These bands, while geographically similar, are different in makeup and natural salinity concentration. Quinney-Lewiston-Layton-Kidman is primary a silty-sandy loam, which is characterized by low levels of salinity. (NRCS soil survey 1974)  Trenton-Jordan-Cache is a silty clay loam with high to severe levels of salinity occurring naturally. The other bands of soil in the study area are not prone to high levels of salinity.(NRCS Soil Survey 1974)

Multispectral imagery from the Landsat 8 sensor was used in the testing and analysis of the soil indices.  The imagery consists of six bands and includes the visible spectrum, NIR and two bands of SWIR. Imagery was initially used from August 30th of 2008 as this was most representative of the dates for the EC testing.  The imagery had already gone through atmospheric and geometric correction. Once the file format was converted, all of the image analysis procedures were conducted in ENVI.

Methods

The Cache Valley study area follows similar patterns in depth of soil concentration as mentioned in Scudiero et al. 2015, only three of the 450 EC tests have a result of 4ds/m or greater in the top level of soil (1-4”).  The forty-six EC tests which returned great than 4 ds/m are all a foot or lower below the surface. This is a result of irrigation in the agricultural plots, the water causes the salt to seep lower in the soil but remain within the root level.  Due to this phenomenon a focus was placed on vegetation indices for potentially identifying salinity levels. Without high concentration of salt on the soil surface it is not likely to register as strongly in the spectral data, however photosynthetic response of plants can still be sensed.

Allbed  and Gorji provided several indices to test in the study area.  From this list several were identified as providing little valuable feedback.  Among the vegetation indices utilized are the Normalized Differential Vegetation Index, Enhanced Vegetation Index, and Ratio Vegetation Index.  A range of eight different soil indices were utilized as well, see Figure 3.  

Several classification methods were applied to the Landsat Imagery.  Test data was derived from the EC results. A pixel to pixel comparison was conducted for the test locations, and surrounding pixels were chosen to make up a region of interest (ROI) around the test location based on pixel similarity.  Three training ROIs were determined to match the EC levels illustrated in Figure 4, tests with greater than 4 ds/m, between 3 and 4 ds/m, and less than 3 ds/m. Methods of classification used were Mahalanobis, Maximum Likelihood, Binary Encoding and Spectral Angle Mapper.  No exclusion was made for test sites based on geographic location or land cover. A confusion matrix was derived for each method of classification.

Initial analysis was visual inspection of the indices compared to the NRCS soil bands and 2008 EC test results.  EC test results were divided into three categories, 0-3 ds/m, 3.01- 4 ds/m, 4.01 and greater. Figure 4 shows the distribution of the test results over the study area.   As would be expected there are more higher threshold EC test results in the band prone to salinity, Trenton-Jordan-Cache. The ratio is 11:7 compared to Quinney-Lewiston-Layton-Kidman.  Middle range threshold results primarily occur in Quinney-Lewiston-Layton-Kidman. Very few of these results occur in the first four inches of soil, as such the most likely indicator of salinity levels would be reduced vegetative output.  See figures 4-9 for a comparison of vegetation and soil indices to EC test results.  (Figures 4-9 were created for use in a project paper and due to a data loss, cannot be updated for a web format.)

Results/Discussion

Distribution of the soil tests was concentrated in two bands with different physical makeup and relatively opposite salinity levels. Lewiston-Layton-Kidman is a more agriculturally suited soil with low natural saline levels.  As would be expected, the Trenton soil prone to saline levels contained more of the higher EC tests. However, the ratio was not as disproportionate as expected, and the suitable EC test results were well dispersed among the two bands.  This could be an indication of restorative agricultural practices in place, in the case of the Trenton soil, and perhaps poor practices in the Quinney plots.

Review of the vegetation focused indices provided some insight on the study area.  There is some small variation in brightness among the active plots, indicating an uneven level of productivity.  This may be unrelated to the salinity level however, as there is not a pattern emerging with either the soil bands or the EC test results.  While the Trenton-Jordan-Cache has more EC tests indicating elevated salinity levels, there does not appear to be any plots with significant difference in brightness.  With the level of variation among the farm plots, it is difficult to determine if there is an actual correlation between the brightness of the vegetation and the soil salinity. 

The soil indices provide similar uncorrelary insight.  NDSI provides a very clear contrast between vegetated land cover and less vegetated land cover, and highlights an area of plots along the river bed that have consistent variance from the surrounding plots.  This could be related to the level of moisture in the soil along the river. In subsequent analysis, it would be beneficial to consider water saturation and moisture levels and the impact on that would have on the spectral characteristics of the soil.

Classification of the Landsat Imagery with the designated ROIs proved to be very difficult.  Gorji and Scudiero et al. 2015 discuss the pixels for classification selection as choosing the pixels with similar values as the ones the test data falls on.  In this particular image the pixels are 30 m x 30 m and the similarity within the agricultural plots is less than expected. A minimum of six pixels that had similar values as the test were selected around each test location.  The middle EC range had only seven test locations, as such the training ROI was very small, less than fifty pixels. For the upper and lower range there were enough test locations to create ROIs with at least seventy pixels. Such small testing and validating ROIs could have impacted the classification.   Mahalanobis classification yielded an overall accuracy of 31.28% and kappa coefficient of -0.06. The result of maximum likelihood classification was an improved 39.96% overall accuracy and kappa coefficient of -0.05. Binary Encoding classification resulted in 24.02% accuracy and a kappa coefficient of -0.10, this method was done without any threshold values applied to the classes.  Spectral Angle Mapping provided negligible results, with and overall accuracy of 3.91% and kappa coefficient of -0.02.  

Conclusions

As exploration in analyzing indices to improve the reach of remote sensing in the agricultural community, this undertaking has been a success.  A certain level of success was achieved in application of the vegetation and soil indices. Upon first examination however of the lack of distinguishing visual characteristics from the soil indices, corresponding to the EC tests, it was apparent that the study area and imagery would not provide the level of detailed feedback that was being sought.  

The products of the four classification methods indicates there is little to no correlation between the datasets.  The small pixel count of the testing data and relative coarseness of the imagery may be the culprits of these results.  Many similar studies have met with much higher levels of accuracy and potential application, and it is perhaps methodology that needs to be altered to create a useful practice for detecting soil salinity in Utah.  Additional EC tests would be desirable for further analysis to create the ground-truthed base of the analysis. Control over the location of the EC tests would be critical for future analysis, as the coarseness of the Landsat imagery almost requires wider dispersion of test locations.  Higher resolution imagery, such as NAIP or products provided by private companies may be able to provide the nuance within an agricultural plot that the EC test intimate exists. 

Remote Sensing has proven in other studies and settings to be more cost effective and with time more accurate than manual data collection in detecting and monitoring soil salinity.  While initial results with this study did not prove fruitful in the way of producing an accurate salinity map, it has tested and discarded several methods, as well as point the way for future analysis.  To achieve the desired result more testing will be required as well as a more layered approach to imagery and ground truth data comparisons.

For more information, read the full report here.

MGGIS Program Skills

• GIS Analysis

• Spatial Data and Algorithms

• GIS Workflow

• Cartography and Graphic Design

• Spatial Analysis

• Communication Skills