Spatio-Temporal Determination of Irrigated Paddy Rice Pixels Using Evapotranspiration and Vegetation Indices: A Case Study for Ca River Basin in Vietnam

Improving irrigation monitoring and efficiency is a current priority of the Government of Vietnam, focusing primarily on the agricultural sector which consumes most of the available surface and groundwater. This paper presents how remote sensing can be used in an integrated manner to achieve better understanding of key eco-hydrological processes including precipitation, evapotranspiration, irrigation and crop growth. The results indicated that Normalized Difference Vegetation Index derived from Moderate Resolution Imaging Spectroradiometer (MODIS) can be applied to determine irrigated pixels on a spatial and temporal basis. The validation using measured water level showed a Pearson correlation of 0.7 proving the high accuracy of this method. The inclusion of these technologies is deemed necessary to improve water resources monitoring and management and hence, ensure long-term drought resilience and water and food security. Ca River Basin in the central Vietnam was selected as a case study to test this approach.

waters related processes are interlinked (Rebelo et al., 2014). Basin-wide sustainable management approaches are required to maintain the key functions of a watershed, including agriculture, biodiversity, energy, health and leisure (Falkenmark & Rockström, 2006). As such, it is imperative to have insight in water management strategies, consumption and demand patterns for all users and functions within a river basin (Hoekstra, 2014;Karimi et al., 2013;Shilpakar et al., 2012). Conventional methods to calculate basin wide water balances often use interpolated weather data including land-use to estimate the total amount of rainfall and evaporation. Since most basins in Vietnam are poorly gauged and data is often difficult to obtain, this is not a viable option (Bastiaanssen & Steduto, 2015).
Several papers indicated that remote sensing technology is suitable for monitoring of hydrology (Neale & Cosh, 2010;Pietroniro & Leconte, 2000;Bastiaanssen & Harshadeep, 2007;Karimi et al., 2013) and water resources (Melesse et al., 2016;Dembele & Zwart, 2016). Remotely sensed precipitation (Serrat-Capdevila et al., 2014), evaporation, water storage has been used worldwide for water management purposes (van Eekelen et al., 2015;Jia et al., 2009;Mirales et al., 2013). Evaporation derived from energy balance approach was used to monitor water consumption and use from local field level to global scale (Templeton et al., 2014;Anderson et al., 2011). Vegetation indices calculated from spectral reflectance provided information on crop growth, water deficit and drought (Hu et al., 2020;Nanzad et al., 2019). These data are especially valuable for policy makers in data scarce areas, as they can be linked to land-use, hydrology and all other water related issues within a river basin. This information is vital to achieve sustainable land management practices.
In addition to climatic and flow data, mapping of irrigated area also required information on crop growth. Traditional assessment involves locally collected data at field level and from census sources. The fieldwork became extensive and often impossible for large or transboundary basin (Sruthi & Aslam, 2015). Determination of vegetation growth from satellite imagery for basin-wide cultivation is a modern and appropriate method of mapping and assessing irrigation. The Moderate Resolution Imaging Spectroradiometer (MODIS) presents a considerably accurate and operational dataset to classify irrigated area at reasonable resolution (Wardlow & Egbert, 2010;Shahriar Pervez et al., 2014). Usman et al. (2015) applied MODIS NDVI to detect temporal changes for irrigated cropland in Pakistan. Biggs et al. (2006) utilized MODIS to map land cover and irrigated pixels for small scale farming in India using MODIS and Landsat time series. Ambika et al. (2016) published irrigated maps for India from 2000-2015 using similar approaches.
In Vietnam, the quality and quantity of the available water resources is already affected by over exploitation and unsustainable land management practices. In the highlands, water induced soil erosion is considered as a threat as it reduces crop yield and pollutes water resources. Heavy rain showers cause floods in the lowlands during the tropical monsoon season, whereas the lowlands are plagued by drought in the dry season . Low river levels in the dry season also cause saline water to enter the river mouth, causing contamination of the ground and surface water. Mismanagement of the natural resources has declined the forest carbon stock, degraded the forest and led to dangerous levels of toxins in the groundwater.
In this paper, we apply the remote sensing approach for mapping irrigated paddy rice in the Ca river basin in Central Vietnam. The Ca river basin is a poorly gauged basin which experiences all of the above mentioned problems of floods, droughts, salinity, water quality deterioration and overlapping mandates.

Study Area
Ca River Basin is a basin cover Laos and Vietnam. The total area is approximately 27,000 km 3 with an estimated population of 5 million inhabitants  The increasing need for electricity by the growing population and water resources in the dry season has led to the development of a series of mainstream and tributary dams. The total effective storage capacity of the medium and larger size dams is 2.2 km 3 /yr. An additional volume of 126 mm 3 /yr is under construction in Ban Mong. With the annual rainfall being 53.8 km 3 /yr, only 4% of the rainfall can be stored (or 8% of the exploitable water resources). This is low for a mountainous catchment, and the potential for small dams for rainwater collection and sustainable hydropower production is great.
There are three large irrigation systems in place, which supply water to the agricultural areas but also supply water for industrial and domestic purposes.

Elevation and Landuse
A Shuttle Radar Topography Mission (SRTM) Digital Elevation Model (DEM) map was retrieved from USGS. The resolution of 1 arc-second or 30 m ( Figure   1) is deemed sufficient to determine irrigated pixels.
The landuse data was collected from SERVIR Mekong Regional Land Cover Monitoring System. This landuse for Ca river basin was studied including Poor-  Figure 2 shows that the lowlands along the coast are mostly used for human settlements and irrigated rice cultivation, but also forest plantations, fruits, and annual or perennial crops. Paddy rice can be grown up to three seasons a year in summer, autumn and winter. Rice cultivation in Summer-Autumn (May -September) is most common, but roughly 70% of the farmers grow also rice in Winter-Spring (January -May), about 30% grows rice year round. Farmers with three rotations are more common in Nghe An (45%) as compared to Ha Tinh (13%). Uplands are dominated by secondary, broadleaved medium forest and rich forest. Most parts of the uplands actually consist of a mosaic of forests and farmlands, where primary forests have been degraded to support goods and services for the local livelihoods. Dry or upland rice cultivation can be found throughout the highlands, whereas the plains are used for more intensive forms rice cultivation. Slash and burn techniques are applied in the uplands for rice or annual crop cultivation, dependent on the market demand. Logging can be found along the main roads, supplying wood for the national and international market. Broadleaved rich forests are associated with national protected areas. Climate Research Unit (CRU) for climatic conditions in Vietnamese river basins and concluded that CHIRPS performance yielded good correlation with station measurements. CHIRPS came with a better resolution of 0.05˚ × 0.05˚ as compared with other products. In this study we used the precipitation data from CHIRPS (Funk et al., 2014).
Rainfall provides 53.8 km 3 /yr (or 1985 mm/yr). The spatial distribution of rainfall is shown in Figure 3. Many areas in the downstream part receive more (2300 mm/yr) than the upstream end in Lao PDR (1250 mm/yr). Flood risk is thus not only caused by runoff from the upstream mountains, but to a large extent influenced by local rainfall. The difference between precipitation and evapotranspiration (water yield) was calculated as precipitation minus evapotranspiration. Figure 5 presents the water yield calculated for 2018. As presented P-ET peaks at approximately 300 -500 mm during July and August when precipitation surplus evapotranspiration. This P-ET can be negative during dry season (between November to June).
The flow measurement for 11 years at Yen Thuong station (Figure 1) was collected to validate the accuracy of irrigated area mapping. The location of Yen Thuong station determines the amount of water flowing into the large irrigation schemes downstream.

Vegetation Index
The Normalized Difference Vegetation Index (NDVI) can be used to monitor C. N. Luong et al. the growth of vegetation cover. NDVI value reflects the concentration of green biome in vegetation leaf and hence, indicates vegetation health and dynamics.
It can give indication on the impact of water shortage and water scarcity to vegetation cover, including crop. NDVI is calculated using the following formula: In which: NIR ϕ is reflectance from near thermal infrared; red ϕ is reflectance from red band. Regional and global spatially determination of irrigated lands conducted earlier using MODIS dataset (Reed et al., 1994). In this study, NDVI dataset was downloaded from MODIS Global MOD13C2. MOD13C2 is cloud-free spatial composites of the gridded monthly 1-km geographic (lat/lon) Climate Modeling Grid (CMG). Cloud-free global coverage is achieved by replacing clouds with the historical MODIS time series climatology record.
NDVI at resolution of 1 km is sufficient to map paddy rice for the 27,000 km 2 basin. Popular rice plants cultivated in Nghe An and Thanh Hoa need approximately 3 -5 months to complete its grow cycle from seeds to matured rice. Hence, a monthly time step of NDVI maps can capture this growth. There are two rice seasons per year in the basin, namely Winter-Spring and Summer-Autumn. The total annual artificial water withdrawals are 4.6 km 3 /yr (2.0 and 2.6 in dry and wet season respectively). The largest portion of the withdrawals is from three large scale irrigation schemes. These irrigation schemes have an intake capacity of about 35 m 3 /s and supply water for irrigation, industry and domestic use. Most of the water is allocated to for irrigation (2.7 km 3 ). Figure 6 illustrated the NDVI dynamics from January to December in 2018. The irrigated mapping was conducted for the period from 2008-2018. NDVI time series can be used to capture discrepancy during growing period and hence, helps separate irrigated and non-irrigated pixels. The information on NDVI was Figure 6. NDVI in Ca River basin calculated for 2018. C. N. Luong et al. used in combination with data on slope derived from SRTM DEM map. Since paddy rice in Ca River Basin was cultivated in rather flat land (<5% slope), this filter was applied to classify irrigated and non-irrigated pixels.

Results and Discussions
A comparison of monthly NDVI time series for different landuse classes and precipitation are shown in Figure 7. For each landuse class, a monthly average time series was plotted against precipitation. Since higher photosynthesis level yielded higher biome and thus, NDVI values, it can be seen that forest classes shown higher NDVI values throughout the year.
The NDVI time series show two distinguished peaks in April-May and August-September respectively. These two peaks are in coincident with the crop calendars when paddy rice matures during these two months before being harvested in June and September (depending on the lunar calendar). In contrast, NDVI for other forest and fruit tree classes show a consistent development from January (middle of dry season) until September when precipitation reaches its peak during monsoonal season. By distinguishing this feature, irrigated paddy rice can be identified from other land use classes or from rain-fed paddies.
More precisely, Figure 8 shows NDVI for paddy rice for an 11-year period from 2008 until 2018. The peaks of NDVI during Winter-Spring crop (February-June) are mostly in April while in several years it was in May, e.g. in 2010 when there is a drought event. Paddy rice is normally harvested in June. The next season Summer-Autumn normally starts during June-July.
The monthly NDVI for different landuse classes so that NDVI for paddy rice peaks twice a year during March-April and August-September. It agrees with a fact that in Ca River basin, most of the downstream paddy rice farm grow two seasons a year. The highest NDVI for paddy rice is 0.65 -0.7 when the rice reaches its highest biome. The harvesting period for Winter-Spring crop season is during end of May to beginning of June and for Summer-Autumn is during September.
These values also match with previous studies, for example FAO recommended NDVI between 0.6 and 0.7. Other landuse classes such as cropland (e.g. maize), forest, grassland or orchard only have one peak per year during September, agreeable with the rainy season. Figure 9 shows that irrigated area and water level in Yen Thuong are highly correlated. Basin evapotranspiration also adds to this consistency. For instance, the year 2010 observes both irrigated area and ET in March drops to its lowest in the observation timeline, at 89,660 ha and 59.74 mm/month. Meanwhile, water level at Yen Thuong station reduces to only 98 cm. Similar results were observed in 2015 when total irrigated area dropped to 102,000 ha and water level was at 86 cm. ET in that year reached 77.68 mm/month.    While in 2017, a wet year, both irrigated area and ET in March peak at 126,944 ha and 103.6 mm/month. Water level at Yen Thuong in March maintained at 142.8 cm showing its capability to secure enough water for irrigation. Table 1 described the correlation calculated using Pearson formula for three datasets: Irrigated area (ha), evapotranspiration-ET (mm) and water level at Yen Thuong (cm). The result indicated that irrigated area and water level has high correlation at 0.6 while irrigated area and ET only yielded at 0.3. On the other hand, ET and water level has a correlation result of 0.7, reflecting that high water level in the river basin resulted in higher plant water intake and hence, higher evapotranspiration. This analysis proved that the methodology applied in this study can improve the accuracy of irrigated area mapping using remote sensing, as compared to field measurement.

Conclusion
This study presents a comprehensive overview of water resources and irrigated area mapping in the Ca river basin using remote sensing. It was demonstrated that remote sensing-derived irrigated area using NDVI has a good agreement with theory and field observation.
While this study provides very specific numbers on irrigated paddy rice pixels, they should be considered as a best estimate rather than absolute figures. More elaborate in-situ information on river flows and water extractions can further refine the numbers presented. Furthermore, it should be noted that all analyses strongly rely on the quality of the land use map. Nevertheless, the data offers exciting insights for policy makers to promote more effective water resource management strategies in the Ca river basin. This mapping of irrigated land provides more insights that allow policy makers and planners to identify alternative approaches such as improving water use efficiency, land management, cultivation of more water efficient crops, crop-rotations and inter-cropping etc.