SENTINEL-1&2 FOR NEAR REAL TIME CROPPING PATTERN MONITORING IN DROUGHT PRONE AREAS. APPLICATION TO IRRIGATION WATER NEEDS IN TELANGANA, SOUTH-INDIA
Keywords: machine learning, rice detection, Sentinel data fusion, water detection, irrigated water needs estimates
Abstract. Indian agriculture relies on monsoon rainfall and irrigation from surface and groundwater. The inter-annual variability of monsoon rainfalls is high, which forces South Indian farmers to adapt their irrigated area extents to local water availability. We are developing and testing an automatic methodology for monitoring spatio-temporal variations of irrigated crops in near real time based on Sentinel-1 and -2 data feed over the Telangana State, South India. These freely available radar and optical data are systematically acquired worldwide, over India since 2016, on a weekly basis. Their high spatial resolution (10–20 m) are well adapted to the small size field crops that is common in India. We have focused first on drought prone areas, North of Hyderabad. Crop fraction remains low and varies widely (from 10 to 60%, ISRO-NRSC, Bhuvan). Those upstream areas, mainly irrigated with groundwater, are composed by less than 20% of irrigated areas during the dry season (Rabi, December to March) and up to 60% of the surface is used for crop production during the Kharif (June to November), which includes rainfed cotton and drip irrigated maize crops and inundated rice. A machine learning algorithm, the Random Forest (RF) method, was automatically used over 6 growing seasons (January to March and July to November, from 2016 to 2018) from the Sentinel-1&2 data stacked for each season, to create crop mapping at 10 m resolution over a study area located in the north of Hyderabad (210 by 110 km). Six seasonal land cover field surveys were used to train and validate the classifier, with a specific effort on rice and maize field sampling. The lowest irrigated area extents were found for driest conditions in Rabi 2016 and Kharif 2016, accounting for 3.5 and 5% with moderate classification confusion. This confusion decreases with the increase of irrigated crops areas during Rabi 2017. For this season, 22% of rice and 9% of irrigated crops were detected after heavy rainfall events in September 2017, which have filled surface water tanks (3.4% of the surface area) and groundwater (Central Groundwater Board observations). From this abundance situation, the surface water detected for each season decreased regularly to less than 0.3% together with the rice and irrigated area extents respectively from 22 to 11% and 10 to 3%, despite a good monsoon 2017. Groundwater level show similar trends, with a drop from 20 meters depth in October 2016 and 2017 to more than 30 m in June 2018 (more recent available data). The deficit of the monsoon 2018 will certainly bring this situation to a hydrological drought at the beginning of 2019, probably similar to the Rabi 2016 situation. The estimated Irrigated Water Demand (IWD) varies from 51 to 310 mm/season, depending on water availability. This methodology shows the potential of automatically monitoring, in near real time, with standard computers, irrigated area extents presenting fast high resolution variability. As it is based on standard global satellite acquisitions, it is foreseen to be used for other regions, for any studies on farmer’s adaptation to climate and hydrological variability, as a proxy to estimate irrigation water needs and water resources availability. In Telangana for instance, it provides an inventory of crop production and irrigation practices before the implementation of mega project infrastructures funded by this new state: - the Kâkâtiya tank restoration program to enhance monsoon runoff capture or the Kaleshwaram project to divert Godavari river water toward upstream Telangana region through tunnels and canals in 20 giant reservoirs.