Extracting Khmer Rouge Irrigation Networks from Pre-Landsat 4 Satellite Imagery Using Vegetation Indices

: Often discussed, the spatial extent and scope of the Khmer Rouge irrigation network has not been previously mapped on a national scale. Although low resolution, early Landsat images can identify water features accurately when using vegetation indices. We discuss the methods involved in mapping historic irrigation on a national scale, as well as comparing the performance of several vegetation indices at irrigation detection. Irrigation was a critical component of the Communist Part of Kampuchea (CPK)’s plan to transform Cambodia into an ideal communist society, aimed at providing surplus for the nation by tripling rice production. Of the three indices used, normalized di ﬀ erence, corrected transformed, and Thiam’s transformed vegetation indexes, (NDVI, CTVI, and TTVI respectively), the CTVI provided the clearest images of water storage and transport. This method for identifying anthropogenic water features proved highly accurate, despite low spatial resolution. We were successful in locating and identifying both water storage and irrigation canals from the time that the CPK regime was in power. In many areas these canals and reservoirs are no longer visible, even with high resolution modern satellites. Most of the structures built at this time experienced some collapse, either during the CPK regime or soon after, however many have been rehabilitated and are still in use, in at least a partial capacity.


Introduction
It is well established that Landsat and other remotely sensed imagery can be used to study land use change [1,2], agriculture [3], the impacts of climate change [4], and many other physical characteristics of the Earth's surface. Less emphasis has been placed on using older Landsat data to apply remote sensing analyses to historical and political questions. Much as current satellite information can be used to track destruction in Syria and genocide in Myanmar [5][6][7], remote sensing techniques can be applied to older Earth observation data to enhance our understanding of political situations of the past. This paper illustrates the benefits and drawbacks of using older, relatively low-resolution satellite imagery from Landsats 1-3 to fill in gaps in the historical record. The information extracted from these images will enhance our understanding of what was (literally) happening "on the ground" during times when very little information was available to those outside of Cambodia's borders.
The U.S. Geological Survey (USGS) and National Aeronautics and Space Administration (NASA) have been collecting satellite imagery since the first Landsat (originally named ERTS, for Earth Resources Technology Satellite) was launched in 1972 [8]. Landsats 1-3 differ from more recent satellites in spatial, spectral, and temporal resolution, as well as data quality and bit depth [9,10]. Although this presents challenges for remote sensing applications, the data from earlier Landsats are still useful for many analyses and should not be overlooked. This paper will focus on the nature of the early Landsat data used and the steps taken to prepare it.
The goal of this research was to determine the nature, extent, and relative success of all irrigation features built through forced labor during the Khmer Rouge regime, which existed in Cambodia from 1975 to 1979. We applied vegetation indices to Landsat scenes from Cambodia in 1973 and 1979 to highlight the presence of water over the entire country. By focusing on the low values in the index range, we were able to identify pixels containing canals and reservoirs. By comparing the early and later imagery we identified the water features that were built during the Khmer Rouge period. The goal first and foremost was to provide an estimate of the scale of the Khmer Rouge irrigation development that was as accurate as possible. The purpose of our survey was to answer the following questions: How many kilometers and dikes were built, and how many new reservoirs were created during the study period?
How large of an increase was this in comparison to pre-Khmer Rouge Cambodian irrigation networks? What was the relative success or sustainability of these irrigation features? Our research has uncovered thousands of kilometers of canals and reservoirs that were constructed between late 1975 to 1979 [11]. Many of the smaller canal networks cannot be detected in existing satellite imagery, indicating that this tally is certainly an undercount of the features that were built at the time.

Historical Background
After the capture of Phnom Penh in 1975, the Khmer Rouge, otherwise known as the Communist Party of Kampuchea (CPK) began forcibly relocating the population to rural agricultural areas with the goal of tripling rice production. This would have been accomplished by expanding irrigation systems and increasing to three harvests a year instead of two. Pol Pot, the leader of the CPK, envisioned a canal network arranged in 10-km grids, each with smaller 1 × 1 km canal grids nested inside them (Figure 1a,b) and each again with 100 m × 100 m grids nested inside them. In this way, the entire country could be supplied with water for rice [12][13][14]. This process of canal and dam building was implemented in multiple locations at the same time. Many structures were built quickly, by hand, without heavy equipment or expertise in the field [15].
Many claims about the extent and effectiveness of Khmer Rouge irrigation schemes reiterate the standard total view (STV) put forth by Michael Vickery. This view generalizes the performance of Khmer Rouge government control and of the Khmer Rouge irrigation development as connected symptoms of a poorly managed regime doomed to failure [15][16][17][18]. Repeated flooding of the landscape over the past 40 years has eroded the traces of many canals on the landscape, and continuous agriculture combined with current increases in development and deforestation all work towards erasing the sites that remain. No geospatial inventory of the structures built during that time period has been previously undertaken [11].
The success or failure of these structures and the role their failure played in the overall failure of the CPK regime has until now been a subject of much supposition and debate unsupported by field observation or physical evidence. While previous inventories of the irrigation infrastructure have been conducted by multiple sources [19][20][21], these have focused on repairing and expanding the existing irrigation structures and the nature of water governance [19,[22][23][24]. Conclusions about the effectiveness and success of the Khmer Rouge irrigation plan cannot be definitively made without a systematic, country-wide survey of the irrigation projects completed during the Khmer Rouge regime. Many claims about the extent and effectiveness of Khmer Rouge irrigation schemes reiterate the standard total view (STV) put forth by Michael Vickery. This view generalizes the performance of Khmer Rouge government control and of the Khmer Rouge irrigation development as connected symptoms of a poorly managed regime doomed to failure [15][16][17][18]. Repeated flooding of the landscape over the past 40 years has eroded the traces of many canals on the landscape, and continuous agriculture combined with current increases in development and deforestation all work towards erasing the sites that remain. No geospatial inventory of the structures built during that time period has been previously undertaken [11].
The success or failure of these structures and the role their failure played in the overall failure of the CPK regime has until now been a subject of much supposition and debate unsupported by field

Study Area
Cambodia has a tropical climate dominated by two main monsoon seasons, with 88% of the precipitation for the region occurring between May and October. The majority of Cambodia is alluvial plain, with a slope of less than 1%, and sandy well drained soils. As part of the lower Mekong Watershed, the Tonle Sap River and the Tonle Sap Lake are part of a unique hydrological system that changes flow direction with the arrival of the wet monsoon season. During the flood season, the Mekong flow increases such that it forces a reversal in the flow direction of the Tonle Sap River away from the Mekong Delta and back into the Tonle Sap Lake which serves as a collection reservoir for the greater Mekong watershed. Over 50% of the sediment collected in the Upper Mekong Basin finds its way into the Tonle Sap Lake, making it a nutrient rich source for the region's fisheries. The Tonle Sap's Remote Sens. 2019, 11, 2397 4 of 21 annual floodwaters also provide an ideal environment for the cultivation of floating rice. Forty percent of Cambodia's population relies on subsistence farming, mainly of rice. Only 7% of Cambodia's agriculture relies on irrigation [25][26][27][28].

Satellite Data Acquisition
Satellite imagery from Landsats 1-3 was downloaded from the USGS Earth Explorer website. Landsat multispectral scanner (MSS) data was chosen as there are more datasets available than other contemporaneous forms of imagery. Landsat has moderate temporal resolution (18-day revisits for Landsat 1-3), and frequent revisits over the same area give more opportunities to gain cloud-free images.
Scenes were collected from the year 1973 to serve as "before Khmer Rouge" reference images, and scenes from the year 1979 served as the "after".
Landsat scenes were identified by path and row as opposed to specific latitude and longitude coordinates. Fifteen individual path/row locations were required to cover the entire country of Cambodia ( Figure 2). Seven of the 1973 path/row locations and five of the 1979 locations had only one usable scene.

Differences in Satellite Image Data Collection and Format
Early Landsat imagery differs from current imagery in several ways. Primarily, lower spatial resolution and the inability to collect data in the 0.45 to 0.52 micrometer bandwidth (the blue band) of the electromagnetic spectrum makes data from Landsats 1-3 incompatible with modern Landsat imagery [29]. Spectrally, the lack of a blue visible light band in the 1970s imagery prohibited true color visualization, so all images discussed were first imported into Clark Lab's TerrSet and made into false color infrared composites. The spatial extent of Landsat 1-3 pixels also differs from later

Differences in Satellite Image Data Collection and Format
Early Landsat imagery differs from current imagery in several ways. Primarily, lower spatial resolution and the inability to collect data in the 0.45 to 0.52 micrometer bandwidth (the blue band) of the electromagnetic spectrum makes data from Landsats 1-3 incompatible with modern Landsat Remote Sens. 2019, 11, 2397 5 of 21 imagery [29]. Spectrally, the lack of a blue visible light band in the 1970s imagery prohibited true color visualization, so all images discussed were first imported into Clark Lab's TerrSet and made into false color infrared composites. The spatial extent of Landsat 1-3 pixels also differs from later satellite imagery. Earlier Landsats collected spatial information in a 79 × 57 m pixel format, post-processed to a 60 m pixel, whereas current Landsat imagery is a standard 30 m pixel for most spectral bands. In terms of information, early Landsat scenes suffered from inconsistent data quality: band transmission errors and other causes of striping resulted in poor image quality and data dropout. Earlier imagery was collected in a four-band, eight-bit format [9,10].

Reference Data
Reference data included Google Earth imagery, ESRI basemap imagery, and primary data collected in the field. During the 2015 and 2016 field seasons a preliminary set of known irrigation locations was collected via ground survey. Points were taken at both surviving canal and dam locations and at the locations of failed, or failing and eroding structures. Interviews with local residents provided rough construction methods and dates, as well as location data for areas where irrigation structures were no longer visible on the surface. Point data and attributes were then imported into ArcMap to aid in visual recognition of irrigation features in the Landsat scenes. Once potential canals and reservoirs were identified, they were then verified by comparison with known structures in both the Landsat data sets and modern satellite imagery. Irrigation development has been a part of Cambodia's history since before 500 A.D., and large-scale ancient irrigation systems still dominate large parts of the landscape, especially in the Siem Reap Province, home to Angkor Wat. Sources used to eliminate ancient canals from the database included maps derived from French aerial reconnaissance in the 1930s as well as recent work by Evans, Fletcher, Pottier, Kummu, and Penny [26,30,31]. Evans and Moylan's [32] study comparing historical imagery of Angkor hydraulic networks with their modern remnants and Khmer Rouge alterations was particularly helpful.

Image Processing Workflow
Since the goal of the Khmer Rouge was to create a nationwide irrigation network, the scope of our analysis needed to include the entire geographic area of Cambodia in order to calculate an estimate of the canals and water storage features constructed [14,33]. Landsat data filled that requirement, with the added advantage of having been collected during the time in question. Once acquired, the data was processed as follows in Figure 3:

1.
Import all bands into Terrset. Create false color near infrared rasters to load into ArcMap.

2.
Geo-reference composite images and individual bands if necessary.

3.
Import near infrared (NIR) and red bands into QGIS to run multiple vegetation index algorithms.

4.
Import vegetation indices into ArcMap and manually digitize water features. See figures for results.

5.
Calculate total kilometers of canals and dikes built during the study period, and total number of water storage features created.
To create three-band raster composites, Terrset was chosen due to its streamlined geotiff file creation process. ArcMap's integration of a world imagery basemap made rectifying imagery and ground truthing irrigation features more efficient. QGIS's System for Automated GIS Analysis (SAGA) plugin, which calculated multiple vegetation indices simultaneously, sped up the image processing step and allowed more time for comparative analysis.
2. Geo-reference composite images and individual bands if necessary. 3. Import near infrared (NIR) and red bands into QGIS to run multiple vegetation index algorithms.
4. Import vegetation indices into ArcMap and manually digitize water features. See figures for results. 5. Calculate total kilometers of canals and dikes built during the study period, and total number of water storage features created.

False Color Composites
The first step in building the image database involved filtering the images by cloud cover and image quality in TerrSet. Due to the lack of a blue band in Landsats 1-3 data, 654/RGB (false color infrared) images were created for each dataset. In this format, healthy vegetated areas appear as bright red and clear water bodies as black or dark blue due to the ability of water to absorb infrared rays [8]. At this stage, defects common to early Landsat scenes, such as data banding, were obvious. In order to maximize the cloud-free pixel area between the multiple images, all scenes were processed [34][35][36]. Cloud-free sections of an imperfect image, even if severely striped, filled in many missing data areas. In Figure 4 the left side displays a severely banded image, yet the large reservoir in the southwest is still easily visible. To create three-band raster composites, Terrset was chosen due to its streamlined geotiff file creation process. ArcMap's integration of a world imagery basemap made rectifying imagery and ground truthing irrigation features more efficient. QGIS's System for Automated GIS Analysis (SAGA) plugin, which calculated multiple vegetation indices simultaneously, sped up the image processing step and allowed more time for comparative analysis.

False Color Composites
The first step in building the image database involved filtering the images by cloud cover and image quality in TerrSet. Due to the lack of a blue band in Landsats 1-3 data, 654/RGB (false color infrared) images were created for each dataset. In this format, healthy vegetated areas appear as bright red and clear water bodies as black or dark blue due to the ability of water to absorb infrared rays [8]. At this stage, defects common to early Landsat scenes, such as data banding, were obvious. In order to maximize the cloud-free pixel area between the multiple images, all scenes were processed [34][35][36]. Cloud-free sections of an imperfect image, even if severely striped, filled in many missing data areas. In Figure 4 the left side displays a severely banded image, yet the large reservoir in the southwest is still easily visible.

Georeferencing Images
While Landsat scenes are processed by USGS and NASA to align with a modern coordinate system, early images do not always line up with the modern datums within a GIS. For this study, all composite images were corrected to WGS 1984 UTM zone 48 North. Reference points were created using identifiable landmarks that could be located in both the original Landsat scene and the modern ArcMap satellite imagery data and rectifying the two images [8]. Roughly 100 reference points were required to rectify each image. Cambodia's lack of development and large forested areas made identification of common landmarks difficult, and changes to the landscape made over almost 50 years made some of the early landmarks indistinguishable. Once this process was completed for the composite image, it was repeated on all the original data bands to create a new set of four geo-referenced bands and one geo-referenced composite image. In addition to correcting spatial accuracy, this process removed extraneous "no data" cells from the raster extent. The corrected images were then used as input for vegetation difference indices.

Vegetation Index Analyses
Although false color infrared gave a quick visual baseline for each data scene, the low-resolution imagery did not always give a clear indication of water containing areas. Due to seasonal variations in flow and sediment load, water visibly varied in color from black to light greenish blue, or reddish tan. Cloud shadows could also be mistaken for bodies of clear water when observed in NIR/R/G format. In order to increase the visibility of water on the landscape, vegetation indices were run on each Landsat scene. The normalized difference vegetation index, or NDVI, was used to extract water features from the surrounding landscape. The NDVI was developed by Rouse [37] using MSS data to minimize variations in reflectance measurements due to topographic effects and to enhance the difference between vegetation and non-vegetated land areas [38]. NDVI is a ratio of the red and near infrared (bands 5 and 6, respectively, in Landsats 1-3) [39]. It is one of the most commonly used vegetation indices and is efficient at detecting irrigated land. It has the advantage of providing a small range in values, −1.0 to 1.0, and a clear cutoff value, indicating that all values below 0 identify non-vegetated land or water bodies [34,40]. Both the corrected transformed vegetation index (CTVI) and the Thiam's transformed vegetation index (TTVI) expand upon the NDVI, and are part of the same family of slope-based vegetation indexes. The TVI was created to correct index values so that the histogram reflected a more normal distribution. The CTVI altered the TVI in order to eliminate most negative values and return a higher quality vegetation index (VI) image. Thiam found the results from the CTVI to be too noisy, and overestimate greenness, so altered the TVI once again by dropping the first part of the equation, resulting in the TTVI [39].
Vegetation indices were calculated in QGIS 2. 18, an open source GIS program. QGIS provides access to third party geoprocessing tools, offering the ease of a "black box" tool with supporting documentation available to the user. By partnering with independent software providers in this way, QGIS provides access to image filters, remote sensing workflows, and other raster analysis methods that might not be available in any one GIS or remote sensing package. One of the third-party providers is SAGA. Within QGIS, SAGA provides a module that computes seven slope-based vegetation indices simultaneously. This gave us the opportunity to produce NDVI, CTVI (corrected transformed vegetation index), and TTVI (Thiam's transformed vegetation index) outputs for each of the downloaded scenes quickly and efficiently. The vegetation index equations appear below.

Visual Evaluation and Irrigation Mapping
All Landsat images and vegetation indices were resampled with bilinear interpolation and stretched using standard deviation symbology in ArcMap, with a 30% increase in contrast. A diverging color ramp was applied to highlight the contrasts between high and low index values. The images that follow were produced from Landsat scene LM3137051979280AAA10 (Landsat scene names have been reformatted since this data was originally acquired, the current Landsat product identifier for this scene is LM03_L1TP_137051_19791007_20180420_01_T2).
Canals were mapped by comparing known locations from field data with both the before and after data sets of satellite images, and using both the satellite basemap available in ArcMap and Google Earth for reference. Canals were often confused with roads in the false color images, however the vegetation index images provided clear differences in values between roads and water. When mapping, binary color ramps were used to contrast the green vegetation along road edges (high NDVI values) from water filled canals (low NDVI values). Uncertainty came in when viewing older canals full of vegetation or eroded former canals that had been plowed over, which usually called for comparing the three vegetation indices with each other and with modern Google Earth imagery. Figures 5 and 6 show the results of the canal and water storage mapping.
Remote Sens. 2019, 11, x FOR PEER REVIEW 9 of 24 Canals were mapped by comparing known locations from field data with both the before and after data sets of satellite images, and using both the satellite basemap available in ArcMap and Google Earth for reference. Canals were often confused with roads in the false color images, however the vegetation index images provided clear differences in values between roads and water. When mapping, binary color ramps were used to contrast the green vegetation along road edges (high NDVI values) from water filled canals (low NDVI values). Uncertainty came in when viewing older canals full of vegetation or eroded former canals that had been plowed over, which usually called for comparing the three vegetation indices with each other and with modern Google Earth imagery. Figures 5 and 6 show the results of the canal and water storage mapping.

Comparison of Index Performance
While all three of the vegetation indices listed provided an acceptable advantage over the false color infrared in water detection, the index that most reliably detected water storage in most cases was the CTVI. The CTVI output reliably identified even very small water bodies while still maintaining distinct boundaries at the water's edges. NDVI outputs often over emphasized water pixels, which made it difficult to distinguish between healthy vegetation and water bodies. The TTVI results tended to falsely identify clouds as surface water, which necessitated referring back to the false color infrared composites for confirmation. This algorithm worked well for identifying water bodies, but overrepresented wet soils and failed to clearly delineate between reservoirs and wet fields or flood plains. Figure 5 shows the 1979 false color infrared, CTVI, NDVI, and TTVI outputs for the same location in Banteay Meanchey province, Cambodia. In false color, the land surface appears indistinct and flooded, with blue coloring suggesting water across most of the image. The NDVI increases contrast between water (blue) and land (yellow to brown), but the overall image is still hazy. Distinction between water and farmland is clearer in the TTVI than the NDVI, but the TTVI tended to overemphasize water. This can be seen in the lower left corner of the image, where the large blue (water) area has the same index value on both sides of the dividing dike, a uniform blue shade across the entire section. Contrast this to the same area in Figure 7a,b, where there is a stronger (darker) water reading on the left side. This was a drawback when mapping water storage features, as it failed to pick up differences in signal between wet land and deeper, stored water. However, the ability of the TTVI to clearly define a contrast between water and land made it the most useful for mapping canals, as it successfully separated the thin linear features from the surrounding landscape. The spectral signal of water is such that it supersedes all other signals in a pixel, making even the fivemeter wide canals detectable [11]. In the CTVI imagery (Figure 7d), containment dikes and reservoirs

Comparison of Index Performance
While all three of the vegetation indices listed provided an acceptable advantage over the false color infrared in water detection, the index that most reliably detected water storage in most cases was the CTVI. The CTVI output reliably identified even very small water bodies while still maintaining distinct boundaries at the water's edges. NDVI outputs often over emphasized water pixels, which made it difficult to distinguish between healthy vegetation and water bodies. The TTVI results tended to falsely identify clouds as surface water, which necessitated referring back to the false color infrared composites for confirmation. This algorithm worked well for identifying water bodies, but overrepresented wet soils and failed to clearly delineate between reservoirs and wet fields or flood plains. Figure 5 shows the 1979 false color infrared, CTVI, NDVI, and TTVI outputs for the same location in Banteay Meanchey province, Cambodia. In false color, the land surface appears indistinct and flooded, with blue coloring suggesting water across most of the image. The NDVI increases contrast between water (blue) and land (yellow to brown), but the overall image is still hazy. Distinction between water and farmland is clearer in the TTVI than the NDVI, but the TTVI tended to overemphasize water. This can be seen in the lower left corner of the image, where the large blue (water) area has the same index value on both sides of the dividing dike, a uniform blue shade across the entire section. Contrast this to the same area in Figure 7a,b, where there is a stronger (darker) water reading on the left side. This was a drawback when mapping water storage features, as it failed to pick up differences in signal between wet land and deeper, stored water. However, the ability of the TTVI to clearly define a contrast between water and land made it the most useful for mapping canals, as it successfully separated the thin linear features from the surrounding landscape. The spectral signal of water is such that it supersedes all other signals in a pixel, making even the five-meter wide canals detectable [11]. In the CTVI imagery (Figure 7d), containment dikes and reservoirs could be clearly distinguished from the surrounding flooded landscape, while the large blue area in the lower left corner clearly separated the water storage section (left) with the inundated area on the right.  In the false color image (a), haze obscured flooded areas and boundaries between water and land were hard to define. In the NDVI (b), the contrast between wet areas (blue) and farmland (yellow to brown) was much clearer, but boundaries still remained blurred. The TTVI (c) isolated the water signal but overemphasized it. This could be seen in the lower left corner of the image, where the large blue (water) area has the same index value on both sides of the dividing dike, showing a uniform blue shade across the entire section. Contrast this to the same area in (a,b), where there was a stronger (darker) water reading on the left side. CTVI (d) results showed the clearest distinction between water storage and flooded fields, and proved the most successful in defining borders between fields (pale blue) and water storage structures (darker solid blue).

Linear Feature Detection
Even given modern imagery, canals varied in width and often had empty or missing sections that made them hard to trace long distances. This was complicated by sun-glint or high sediment load. While clear water absorbs most of the infrared that hits it, sediment filled water reflects more light back to the sensor and often registers higher index values similar to bare ground. A particularly effective example of this occurs in Banteay Meanchey Province, where two intersections presented similarly in the false color imagery, however one intersection displays canals and the second did not.  In the false color image (a), haze obscured flooded areas and boundaries between water and land were hard to define. In the NDVI (b), the contrast between wet areas (blue) and farmland (yellow to brown) was much clearer, but boundaries still remained blurred. The TTVI (c) isolated the water signal but overemphasized it. This could be seen in the lower left corner of the image, where the large blue (water) area has the same index value on both sides of the dividing dike, showing a uniform blue shade across the entire section. Contrast this to the same area in (a,b), where there was a stronger (darker) water reading on the left side. CTVI (d) results showed the clearest distinction between water storage and flooded fields, and proved the most successful in defining borders between fields (pale blue) and water storage structures (darker solid blue).

Linear Feature Detection
Even given modern imagery, canals varied in width and often had empty or missing sections that made them hard to trace long distances. This was complicated by sun-glint or high sediment load. While clear water absorbs most of the infrared that hits it, sediment filled water reflects more light back to the sensor and often registers higher index values similar to bare ground. A particularly effective example of this occurs in Banteay Meanchey Province, where two intersections presented similarly in the false color imagery, however one intersection displays canals and the second did not.  (Figure 8b), a similar line is evident. On the left it starts as a pale line edged in red running through a mixed light blue and brown background. As you follow the line towards the right, it becomes lighter and more pronounced against the bright red vegetation on the landscape. It, too, has a somewhat offset intersection running through the center of the image, however these lines are not as precisely defined and have a rougher outline than the lines at intersection a. The paths of the two intersections are shown enlarged in blue below. At both locations, a diagonal line cuts across the image from the east to the west. At intersection a (Figure 8a) a, this line is lighter and almost white compared to the bright red vegetation surrounding it. Multiple perpendicular lines meet or cross it, although these are faint and almost fade into the brown/red background. At intersection b (Figure 8b), a similar line is evident. On the left it starts as a pale line edged in red running through a mixed light blue and brown background. As you follow the line towards the right, it becomes lighter and more pronounced against the bright red vegetation on the landscape. It, too, has a somewhat offset intersection running through the center of the image, however these lines are not as precisely defined and have a rougher outline than the lines at intersection a. In Figure 9, Sites a and b display similar diagonal intersecting lines. In the top row, CTVI images are displayed using a standard-deviation gamma stretch. The lower index values of Site a remain lighter than the background values. Their light-yellow appearance indicates values in the middle of the index value range. This contrasts with Site b, which displays the dark brown (high CTVI value) lines of healthy vegetation, often found lining roadways. With the standard-deviation stretch, linear features are clear and contiguous, with good definition while still retaining vegetation (b) and nonvegetation (a) qualities. When the stretch was changed to a histogram equalized view (bottom row), linear definition was compromised, but Site a's canals appeared distinctly in blue, low index values, confirming the presence of water. What can be seen of Site b's linear features were much darker brown. This color contrast in the CTVI highlighted the difference in infrared reflection between vegetation and water, providing the information necessary to distinguish canals from roads.   Figure 10 shows Site a in 2019 Planet Labs imagery, with a 3 m spatial resolution, much higher than the 1979 Landsat 3 imagery (60 m resolution). The major intersection that was at the center of Figure 8a appears in the top right corner. This imagery illustrated just how difficult it was to find relict canals in true color imagery, especially after 40 years had passed. There was still a need for CTVI analysis even with high spatial resolution, as roads, canals, and vegetation can appear many different ways in a single true-color image.
The red inset rectangle ( Figure 10a) highlights a wide, dark green canal to the south of the gray asphalt road on the left of the image. Close to the light-brown dirt road running north-south, canals begin running on both the south and north banks. Once the asphalt road crosses the dirt road, the canal south of the main road ends. The dirt road has a small tertiary canal in dark green running along its west side. Dark green colors indicated that the canal was older and had been overgrown by vegetation.
The blue inset rectangle (Figure 10b) is less that 400 m away from the red inset, but the dirt road running north-south at this location is a pale white line. The canal here also runs on the west side of the road, but is fragmented, and has various levels of reflectance in each section. The northmost canal section follows the dark green appearance of the previous wide canals. The section immediately to the south is a light blue, due to sun-glint reflecting off the surface. The last section is as wide as the light blue section above it, but the same color as the dirt road it runs beside. Contextually, the width, location, and the knowledge that the soil was very silty can help to identify this section as canal and not road. The light color can be explained by the fact that water very high in sediment reflects similarly to soil in the visible spectrum. It requires CTVI analysis to verify this description, however.  Figure 10 shows Site a in 2019 Planet Labs imagery, with a 3 m spatial resolution, much higher than the 1979 Landsat 3 imagery (60 m resolution). The major intersection that was at the center of Figure 8a appears in the top right corner. This imagery illustrated just how difficult it was to find relict canals in true color imagery, especially after 40 years had passed. There was still a need for CTVI analysis even with high spatial resolution, as roads, canals, and vegetation can appear many different ways in a single true-color image.
The red inset rectangle ( Figure 10a) highlights a wide, dark green canal to the south of the gray asphalt road on the left of the image. Close to the light-brown dirt road running north-south, canals begin running on both the south and north banks. Once the asphalt road crosses the dirt road, the canal south of the main road ends. The dirt road has a small tertiary canal in dark green running along its west side. Dark green colors indicated that the canal was older and had been overgrown by vegetation.
The blue inset rectangle (Figure 10b) is less that 400 m away from the red inset, but the dirt road running north-south at this location is a pale white line. The canal here also runs on the west side of the road, but is fragmented, and has various levels of reflectance in each section. The northmost canal section follows the dark green appearance of the previous wide canals. The section immediately to the south is a light blue, due to sun-glint reflecting off the surface. The last section is as wide as the light blue section above it, but the same color as the dirt road it runs beside. Contextually, the width, location, and the knowledge that the soil was very silty can help to identify this section as canal and not road. The light color can be explained by the fact that water very high in sediment reflects similarly to soil in the visible spectrum. It requires CTVI analysis to verify this description, however.  Figure 11 shows the intersection at Site b. The small red inset (Figure 11a) focuses on the road that crosses the image from west to east. This is a light beige dirt road flanked by houses on either side. Running vertically through the image is a natural waterway, in dark greens. This channel has a much more organic shape when contrasted to the manufactured canal to the east. In the CTVI and false color imagery, it was this channel to the west that appeared as a water on our Landsat imagery, and the more engineered canal to the east was not detected. The eastern canal may be more modern in construction. As to why the natural waterway appeared as vegetation in the CTVI image, comparing the color of the vegetation in the channel to the vegetation in the town to the east (top right corner of blue inset image), the natural waterway appears overgrown to the point where it would have a high enough infrared reflectance to register as vegetation as opposed to water. The more modern canal, highlighted in the blue inset (Figure 11b), is more to type, with a dark green channel and two light dirt roads on the east and west banks. As this channel proceeds north toward the town, however, we picked up sun-glint and sedimentation that made this wider canal section appear lighter and more similar to the dirt roads.  Figure 11 shows the intersection at Site b. The small red inset (Figure 11a) focuses on the road that crosses the image from west to east. This is a light beige dirt road flanked by houses on either side. Running vertically through the image is a natural waterway, in dark greens. This channel has a much more organic shape when contrasted to the manufactured canal to the east. In the CTVI and false color imagery, it was this channel to the west that appeared as a water on our Landsat imagery, and the more engineered canal to the east was not detected. The eastern canal may be more modern in construction. As to why the natural waterway appeared as vegetation in the CTVI image, comparing the color of the vegetation in the channel to the vegetation in the town to the east (top right corner of blue inset image), the natural waterway appears overgrown to the point where it would have a high enough infrared reflectance to register as vegetation as opposed to water. The more modern canal, highlighted in the blue inset (Figure 11b), is more to type, with a dark green channel and two light dirt roads on the east and west banks. As this channel proceeds north toward the town, however, we picked up sun-glint and sedimentation that made this wider canal section appear lighter and more similar to the dirt roads.

Canal and Water Storage Inventory
Excluding ancient Angkorian systems, we identified a total of 2540 km of canals in the "before" dataset, and 9772 km of canals in the post-Khmer Rouge satellite data. This is an increase of over 7000 km between 1973 and 1979. Although canals as small as five meters across could be detected, this is certainly an underestimate of the scope of the construction completed during this time. Field surveys in areas where the grid system has been maintained indicate that these canals were spaced one km apart, and smaller canals were spaced at 100 m intervals between them (see Figure 1).
Reservoirs increased from roughly 32 detectable in the pre-Khmer Rouge era to over 350 separate reservoirs by 1979. A rough typology of storage types was created based on the method of water containment and relationship of the water storage feature to both topography and hydrology of its location. Six types were identified: external storage, internal storage, segregated storage, linked storage, topographically controlled storage, and large-scale canal systems. For a more in-depth discussion of the typology characteristics, see Tyner et al. [11].

Canal and Water Storage Inventory
Excluding ancient Angkorian systems, we identified a total of 2540 km of canals in the "before" dataset, and 9772 km of canals in the post-Khmer Rouge satellite data. This is an increase of over 7000 km between 1973 and 1979. Although canals as small as five meters across could be detected, this is certainly an underestimate of the scope of the construction completed during this time. Field surveys in areas where the grid system has been maintained indicate that these canals were spaced one km apart, and smaller canals were spaced at 100 m intervals between them (see Figure 1).
Reservoirs increased from roughly 32 detectable in the pre-Khmer Rouge era to over 350 separate reservoirs by 1979. A rough typology of storage types was created based on the method of water containment and relationship of the water storage feature to both topography and hydrology of its location. Six types were identified: external storage, internal storage, segregated storage, linked storage, topographically controlled storage, and large-scale canal systems. For a more in-depth discussion of the typology characteristics, see Tyner et al. [11].

Accuracy Assessment
To evaluate the success or sustainability of the irrigation features, first we looked at the larger structures as a whole-were the water storage features that we detected via the satellite imagery verified by ground-truthing? To what degree were the features detected-were structures identifiable "in toto" or were large components of irrigation systems missed by the CTVI index? The satellite mapping process revealed over 350 water storage features, all but one of which were Khmer Rouge constructed or improved. All of these were at least partially traceable in modern imagery, although not necessarily visible at ground level. Even for modern and rehabilitated dams, the amount of water stored at any given time varies, and a reservoir may be drained for irrigation use and sit

Accuracy Assessment
To evaluate the success or sustainability of the irrigation features, first we looked at the larger structures as a whole-were the water storage features that we detected via the satellite imagery verified by ground-truthing? To what degree were the features detected-were structures identifiable "in toto" or were large components of irrigation systems missed by the CTVI index? The satellite mapping process revealed over 350 water storage features, all but one of which were Khmer Rouge constructed or improved. All of these were at least partially traceable in modern imagery, although not necessarily visible at ground level. Even for modern and rehabilitated dams, the amount of water stored at any given time varies, and a reservoir may be drained for irrigation use and sit empty until the next rainy season (an example of this was the Sla reservoir that was mapped with very low water levels during the 2018 field season). Because of this and the tendency of structures to erode in monsoon season, the evidence of dams, dikes or gates at a site all served to verify previously existing water storage. Of the reservoirs identified in the satellite imagery, 13 were verified as Khmer Rouge by ground survey, and five were determined to have been canal structures. One was verbally confirmed by a local resident, but unable to be reached by road. In addition, 14 new Khmer Rouge dams were identified in a field survey, for a total of 27 dams ground-truthed. All but one of these were visible in the 1979 imagery once we knew where to look based on ground data. When ground truthing canals, 276 satellite-identified points were surveyed as possible canals or canal fragments, and 272 of these proved to be accurate. Table 1 shows the accuracy matrix comparing the satellite feature detection and the ground survey. The category "other" describes canals or dams that were not mapped previous to the ground survey. Several factors contribute to the relatively low number and low accuracy of water storage data points as compared to canal features. One, canal features were usually mapped at multiple points along their length, at gates or culverts or failure points. This means that canal points were nearer to each other and more numerous than water storage points. Often, during monsoon season canals over topped or breached, leading to large flooded areas in the imagery that were seen as storage. Areas where canals had broken or breached were sometimes mistakenly identified as individual reservoirs, when in fact they were two parts of the same canal. These large flooded areas resulted from the failure of the canal system transporting the water, rather than the storage features themselves. Finally, the size and scope of some of the water storage features, especially those that were topographically constrained, combined with the distance and limited time in the field prevented us from taking as many points as would have been desirable for a larger reference data set. Most of the water storage features (373 in total) were verified through satellite imagery available on Google Earth.
Separating individual canals even in modern imagery is often impossible, especially if those images were taken during flood season. For reference we used a combination of government and NGO (non-governmental organization) reports on scheme conditions and improvement plans. We then compared the canal networks detected in the 1979 CTVI imagery with modern networks known to have been modified from original Khmer Rouge construction. Following Treffner [43], we focused on identifying irrigation schemes and sub-projects rather than individual canals. Much of the data available consisted of lists of irrigation projects by district without location information, with details given only on planned rehabilitation project sites. Given these limitations, of the over 9000 km of canals traceable in the 1979 imagery, 897 km were linked to 32 irrigation schemes, in 11 provinces.

Success or Failure?
In terms of judging the success or failure of the Khmer Rouge structures, we looked at factors that would apply to any irrigation structure. Did the structure perform its most basic function? If it was a canal, did it transport water to the intended area? If it was a water storage feature, how effective were the structures (dam, gates, or culverts) at retaining and distributing water as needed?
At three water storage sites, containment dams were not completed before the fall of the Khmer Rouge regime. These structures were all in the CPK Northwest zone, which remained under the influence of the Khmer Rouge until 1998 [24]. Two of these were large topographically constrained structures, Kampong Puoy and Trapaeng Thmar, with dikes stretching over 6 km and 13 km, respectively. The third was the Prek Chik irrigation diversion dam, secondary intake structure for the large-scale Damnak Ampil/April 17th canal system, which stretched over 48 km. The Damnak Ampil was both unable to cope with heavy monsoon rains, which led to frequent canal breaching due to overland flow, and too large and permeable to distribute enough water to farmers along its length during the dry season. Both this dam and a nearby dam to the northeast, Thleam Ma'oam, collapsed repeatedly during construction. We were unable to procure details for each of the water storage sites visited, but of the 11 dams for which we have failure information, five lasted fewer than three years, two failed after roughly six years, and only two lasted more than a decade. Seven were left unused or abandoned at some point, five of those permanently.
Functionally, at nineteen of the water storage sites the dams are at least partially operable at the current time, and most of these have been rehabilitated at least once. Erosion around gate structures and inoperable gates were the primary complaints. At Trapaeng Thmar, the largest water storage site, water distribution, gate repair, and canal maintenance has been fractious and plans for rehabilitation poorly understood since construction [43]. This area of Cambodia suffered from conflict between Khmer Rouge splinter groups and the government until 1998. As of 2010 only upstream farmers could benefit from the reservoir during the dry season [43]. The projected goal for the Trapaeng Thmar project by the Khmer Rouge was to irrigate 3000 ha. At present the upstream irrigated land has increased from 50 hectares to 300 hectares, but downstream fields remain unirrigated and with no more access to water during the dry season than in the 1970s. In southwest Cambodia, Kep province, four Khmer Rouge reservoirs were linked along a single river. Upstream of this cascade system a large, post-Khmer Rouge dam (Chamka Bey) was built that funneled more water than the next dam downstream could hold, resulting in breaching of the dam at Kraal Kau Ta Pom and downstream flooding [44]. In the case of the Damnak Ampil irrigation scheme, as of 2009 water did not reach the secondary canals due to their deterioration [19]. The permeability of the canals causes high water loss and neither upstream or downstream residents are able to access water to farm dry season rice [45]. The Jan 6th/Traing Krasaing system has repeatedly breached since its construction due to monsoon-related runoff, and despite being in a water-rich area only supplies water to 3% of its potential service area [46]. Table 2 lists the conditions of water storage sites that we were able to visit.
Even more fundamental than whether a particular structure still functions is whether or not it still exists. Figure 12 shows Kep Province, Cambodia, where many reservoirs were visible in the 1979 imagery that were barely traceable in modern imagery. Field survey and local interviews confirmed that multiple reservoirs, seen as dark blue or black in the 1979 near infrared image (Figure 12a) were built by the Khmer Rouge; these are visible as blue in the 1979 CTVI images (Figure 12b). In the high resolution images from Planet Labs (Figure 12e,f), the two reservoirs in the lower left corner are no longer visible. In the northeast, a large dike south of the river can be traced, but does not hold water. Even more fundamental than whether a particular structure still functions is whether or not it still exists. Figure 12 shows Kep Province, Cambodia, where many reservoirs were visible in the 1979 imagery that were barely traceable in modern imagery. Field survey and local interviews confirmed that multiple reservoirs, seen as dark blue or black in the 1979 near infrared image (Figure 12a      Applying vegetation indices to pre-Landsat 4 imagery gave us the unique ability to view these irrigation systems immediately after their construction. This allowed us to view the structures in their original configurations, before floods, failures, and multiple rehabilitations altered their function or layout. However, challenges to using this data exist and we are still experimenting with ways to make this process more efficient and productive.
The ability to automatically detect irrigation features would greatly speed up the process. Multispectral classification was attempted, but large pixel size combined with poor quality data rendered it unsuccessful. In the future we hope to use Hexagon KH-9 imagery from the same time period to provide a higher resolution view of the area. Extracting raster attributes in ArcMap in order to create water-only rasters has been promising. By removing all but the values below the average value in the CTVI rasters, we are left with a raster of only water storage and flooded features. This works well to isolate even very small reservoirs. Unfortunately, the canal values are too close to the surrounding wet soil values to be able to separate the two. An image convolution filter might improve results. Several cloud removal processes have been attempted but many require multiple images of the same area or the use of the blue spectral band, which is unavailable in early Landsat imagery [47]. We have tried a partial homomorphic filter process with limited success [48]. Braaten et al.'s [49] cloud and cloud shadow identification algorithm is a promising alternative.

Conclusions
While the STV maintains that these structures failed due to lack of planning, improper design, or unskilled workmanship, it is apparent that, while faulty, many of the structures are still being used for their intended purpose. In some cases, even when completely rebuilt, as in the case of the Damnak Ampil intake dam or the dam at Thleam Ma'om, the locations and general layout rely on the systems initially put in place during the Khmer Rouge period. In cases like Trapaeng Thmar, functionality has been partially improved by the completion of control gates to release water downstream, but the access to irrigation water remains problematic. This is as often as much a failure of current planning or lack of cooperation between villages as it is hydraulic in nature [24]. Our field work has observed significant erosion in canals that have been widened and improved within the last few years, due to the inherent instability of the alluvial soil. Without entirely lining canals with stone or concrete, it is hard to see how any canal, modern or Khmer Rouge, could withstand the fluctuations in flow during the monsoon season. Applying vegetation indices to pre-Landsat 4 imagery gave us the unique ability to view these irrigation systems immediately after their construction. This allowed us to view the structures in their original configurations, before floods, failures, and multiple rehabilitations altered their function or layout. However, challenges to using this data exist and we are still experimenting with ways to make this process more efficient and productive.
The ability to automatically detect irrigation features would greatly speed up the process. Multispectral classification was attempted, but large pixel size combined with poor quality data rendered it unsuccessful. In the future we hope to use Hexagon KH-9 imagery from the same time period to provide a higher resolution view of the area. Extracting raster attributes in ArcMap in order to create water-only rasters has been promising. By removing all but the values below the average value in the CTVI rasters, we are left with a raster of only water storage and flooded features. This works well to isolate even very small reservoirs. Unfortunately, the canal values are too close to the surrounding wet soil values to be able to separate the two. An image convolution filter might improve results. Several cloud removal processes have been attempted but many require multiple images of the same area or the use of the blue spectral band, which is unavailable in early Landsat imagery [47]. We have tried a partial homomorphic filter process with limited success [48]. Braaten et al.'s [49] cloud and cloud shadow identification algorithm is a promising alternative.

Conclusions
While the STV maintains that these structures failed due to lack of planning, improper design, or unskilled workmanship, it is apparent that, while faulty, many of the structures are still being used for their intended purpose. In some cases, even when completely rebuilt, as in the case of the Damnak Ampil intake dam or the dam at Thleam Ma'om, the locations and general layout rely on the systems initially put in place during the Khmer Rouge period. In cases like Trapaeng Thmar, functionality has been partially improved by the completion of control gates to release water downstream, but the access to irrigation water remains problematic. This is as often as much a failure of current planning or lack of cooperation between villages as it is hydraulic in nature [24]. Our field work has observed significant erosion in canals that have been widened and improved within the last few years, due to the inherent instability of the alluvial soil. Without entirely lining canals with stone or concrete, it is hard to see how any canal, modern or Khmer Rouge, could withstand the fluctuations in flow during the monsoon season.
Germann and Epp [50] emphasize that one of the singular qualities of an archaeological site is that the physical context of a site contributes as much information as the objects that are found there. Watershed development can irrevocably destroy that context as well as bury or wash away a site's material artifacts. While they focus on the damage that reservoir and watershed development cause to the archaeological material contained within the site, in the case of the Cambodian dams and canals, the sites themselves are the material culture. Our study has shown that often these structures have been built, washed away, and rebuilt multiple times. The history of their construction and reconstruction is inseparable from the history of the people killed during their production [51,52]. The best physical record that we have of their production is recorded in the Landsats 1-3 imagery. It is important that we continue to salvage what we can to fill this gap in our historical knowledge by retrieving as much information as possible from these images.