Small Water Bodies Identification

Examination of on-farm small water bodies in Roza Irrigation District

Here I describe the method to identify on-farm small water bodies for agricultural areas using satellite and airborne imagery for Roza Irrigation District, Washington.

Water Indices

There are multiple algorithms which have been proposed for identifying water bodies with remote sensing imageries such as supervised and Unsupervised classification (Olmanson et al. 2008; Verpoorter et al. 2012), single and multiple band thresholding (Jain et al. 2005; Sun et al. 2012), and spectral water index method (Xu 2006; Feyisa et al. 2014; Mishra and Prasad 2015). Among these methods, spectral water index method is widely used because it is easier to apply and has shown to have good accuracy for open water bodies (Jiang et al. 2014).

Normalized Difference Water Index (NDWI) (McFEETERS 1996) is calculated as

NDWI = (Green - NIR) / (Green + NIR)

It maximizes the reflectance properties of water. This index has produced good results compared to single or multiple band thresholding methods (Jain et al. 2005). The major limitation of this index is that it cannot provide good results for water bodies in urban areas. But since this work is focused on agricultural areas, NDWI was used as the first index for analysis.

Modified Normalized Difference Water Index (MNDWI) (Xu 2006) is calculated as

MNDWI = (Blue - NIR) / (Blue + NIR)

It has been developed to work better in urban areas. While NDWI overestimates open water pixels, MNDWI has shown to suppress land, vegetation and soil noise to enhance water information. This index was used for this work.

Figure 4 Reflectance of water at different wavelengths (Top). The reflectance values for a point inside a water body using band values from Landsat 8 (Bottom).

Landsat Imagery

The initial analysis was conducted using Landsat 5 and Landsat 8 Satellite imageries for finding the water bodies from 2009 to 2015. Landsat has three visible bands (Blue – 0.45 to 0.51 µm, Green – 0.53 to 0.59 µm and Red – 0.64 to 0.67 µm) and a near infrared (NIR – 0.85 to 0.88 µm) which can be used for computing MNDWI. These bands have a resolution of 30 m. Although the lowest temporal resolution of Landsat imagery is 15 days, there are some areas where the path of satellite overlaps and these areas are more frequently imaged. Images for analysis were selected from the month of June and July. To create a composite image for the year and to remove effects of clouds and its shadows, median reflectance values were selected from the stack of imageries from the selected periods. This assumes that the extent and signal from water bodies during the selected period remains constant.

National Agricultural Imagery Program (NAIP) imagery

After initial analysis with Landsat, NAIP imagery was used to identify small water bodies. The fine resolution of NAIP was used to ascertain the location and extent of water bodies computed using Landsat. NAIP has four total bands - Red, Green, Blue and Near Infrared. This is a dataset provided by USDA Farm Service Agency. These imagery is acquired and distributed at least every 3-year cycle for the entire continental US free of cost. The imageries used for our analysis were 2009, 2011, 2013 and 2015. NAIP imagery does not provide reflectance values but provides scaled values between 0 and 255 for all of the bands. Since no intensities are provided in the dataset, care must be taken to compare, mosaic and stack these data. Since spectral water indices described above use ratio of differences in band values, these imageries can be used without further processing.

Analysis Platform - Google Earth Engine

Google Earth Engine is a cloud-based geo-processing platform where the required satellite and airborne imagery are stored at the cloud and processing is done at the cloud server. This platform was chosen for this work because: (1) No satellite or air-borne imagery needs to be downloaded (2) Processing of these large datasets is very fast using cloud-based server (3) Scripts written for a specific region can be applied to a different region without much hassle (4) Application Program Interface (API) is available in both JavaScript and Python, which are commonly used programming languages which makes it easier to update the scripts as better algorithms are added or updated.

Process

The process used for the two imageries varied slightly due to their spatial resolutions.

Spectral indices for medium resolution imageries (e.g. Landsat) work better to identify water bodies but the accuracy decreases with size of water body identified. Based on visual inspection of water bodies from google earth imagery (at ~0.5m resolution), we observed that identifying smaller water bodies, usually lower than 4 times the resolution of the image was challenging (for example - it is challenging to identify water bodies with smallest dimension smaller than 120 m using a 30m resolution Landsat imagery). Nevertheless, we did not filter water bodies based on their size for this analysis.

Use of water specific spectral indices for high resolution imagery (e.g. NAIP) to identify water bodies is challenging because the indices also highlight shadows from built up areas of larger trees. Due to the high resolution, there are multiple occasions where small features in the map cast shadows due to which water bodies are confused with shadows. So, a general threshold cannot be used for identifying water bodies. This problem can occur with medium sized satellite imagery (e.g. Landsat) as well, but at such a resolution in the region selected, where the land features are relatively plane, the degree of misidentification is lower.

The first step in identifying water bodies with these imageries was to use thresholds of water index (MNDWI). The thresholds were either based on literature (e.g. Landsat derived MNDWI thresholds water at 0) or using a value from a distribution of MNDWI for manually delineated water bodies (Figure 6) (e.g. NAIP imagery thresholds most of the water bodies in our study area at MNDWI > 0.4). Using these thresholds, there was a positive bias in the water pixel identification i.e. there were more pixels identified as water bodies than there were water bodies in the image. Some built-up areas and shadows were also identified as water bodies using the threshold method.

The second step was to remove urban areas and shadows from this threshold image. It was done using a K means classification scheme. Red, Green and Blue (RGB) bands of the image were first masked using the threshold image and these bands were then used for classification. The final result was a classified image of the threshold image as water and non-water pixels. These water pixels were then aggregated as water bodies.

Results

· Using Landsat Imagery: Landsat 5 and 8 were used to identify water bodies. The difference in spectral resolution of the two satellites causes for some differences in water bodies identified by each. So, for comparison purposes, it is necessary to be aware of the years for each satellite was used. Here, we resent only results from 2013 – 2016 which were obtained using Landsat 8. Figure 7 shows the location of water bodies for 2016.

Figure 8 shows the changes in total area of small water bodies in Roza Irrigation District. It can be observed that in 2015, there was a reduction of about 40 acres of small water bodies while in 2016, the area of water bodies increased. This decrease in small water bodies in the year 2015 coincides with a drought in the region. This is also evident in the reduction of mean flow in 2015 at the outlet of Lower Yakima Basin.

The potential of water storage in these small water bodies for Roza irrigation district was estimated using these results. For this, we assume that any pixel which has been identified as a small water body on any year between 2013 and 2016 remains a water body and does not change into any other land use type. Using this assumption, we estimated the area for total water storage potential until 2016 to be about 155 acres. Various proportion of this potential has been utilized in different years in Roza. About 60% of the total potential were used to store water in 2013 and 2014, 45% of the potential was used in 2015 and 80% of the total potential was being utilized in 2016.

Figure 7 Location of small water bodies in the year 2016. The sizes of the dots represent the size of water bodies (exaggerated)

Figure 8 Changes in total area of water bodies in Roza Irrigation District (left). Mean annual flow for an outlet at Lower Yakima Basin.

· Using NAIP Imagery: NAIP imagery was available only for 2009, 2011, 2013 and 2015. These imagery had Red, Green, Blue and NIR bands but not reflectance values. So, the results were dependent on the quality of imagery available. We used imageries for 2013 and 2015 to compare the area of water bodies with those obtained using Landsat imagery.

Comparison of results from Landsat and NAIP imagery:

A visual comparison of the water bodies identified using Landsat and NAIP imagery shows that high resolution imagery is capable of identifying smaller water bodies accurately (Figure 9). This inspection also helped visually assess the accuracy of the identified water bodies. The accuracy was higher in rural (agricultural) areas than in urban (built-up) areas.

Figure 9 Comparison of water bodies identified using Landsat and NAIP imagery.

Comparison of total area of water bodies identified using NAIP imagery shows that the total area identified was about four times greater than that identified using Landsat imagery (Figure 10). Accuracy assessment was not performed only visually for this analysis.

This analysis helps recognize the importance of high resolution imagery to identify changes in water bodies at a farm-scale. NAIP imagery is not recorded annually but once every three years. So, this analysis provides a loose basis to use Landsat imagery to get an estimate of total acreage of small water bodies in Roza. A major assumption is that the growth/decline of small water bodies is proportional to growth/decline of larger water bodies (identifiable using Landsat).

Figure 10 Comparison of total area of water bodies identified using NAIP imagery and Landsat.