Aiding Irrigation Census in Developing Countries by Detecting Minor Irrigation Structures from Satellite Imagery Chintan Tundia, Pooja Tank and Om Damani Indian Institute of Technology - Bombay, Mumbai, India Keywords: Minor Irrigation Census, Object Detection, Deep Learning, Remote Sensing, Computer Vision. Abstract: Minor irrigation structures such as well and farm ponds play very important roles in agriculture growth in developing countries. Typically, a minor irrigation census is conducted every five years to take inventory of these structures. It is essential that an up to date database of these structures be maintained for planning and policy formulation purposes. In this work, we present the design and implementation of an online system for the automatic detection of irrigation structures from satellite images. Our system is built using three popular object detection architectures - YOLO, FasterRCNN and RetinaNet. Our system takes input at multiple res- olutions and fragments and reassembles the input region to perform object detection. Since currently there exists no dataset for farm pond and the only publicly available well dataset covers a small geographical region, we have prepared object detection datasets for farm ponds and wells using Google Maps satellite images. We compare the performance of a number of state of the art object detection models and find that a clear trade-off exists between the detection accuracy and inference time with the RetinaNet providing a golden mean. 1 INTRODUCTION In developing countries, irrigation plays an impor- tant role in farming and agricultural growth (Kirpich et al., 1999). Minor Irrigation structures account for a huge part of irrigation infrastructure due to short con- struction period and low investment required. These structures, such as wells, check dams, and farm ponds (NIC , MeitY., 2014), have cultivable command area up to 2000 hectare, . In India 65% of the agricul- ture depends on minor irrigation (Frenken, 2012). It is essential that an up to date database of these struc- tures be maintained for planning and policy formu- lation purposes for which minor irrigation census is conducted every 5 years. The census data is collected from village level workers/administrators, revenue or land records and a survey of different government and private scheme owners. After collection of field data, the data entry is done on an online portal. State gov- ernments monitor the progress of field work, data en- try and validation work. The validated data is again examined by the Central Government before the final report generation. Moreover, prior to conducting cen- sus, training workshops are conducted at central as well as regional levels. Thus the process of conduct- ing minor irrigation census involves lot of cost and effort. As per (V. K. Bhatia, et. al, 2010) there are sev- eral obstacles in conducting minor irrigation census. Census officials face difficulties like unavailability of village records, villages being located at remote areas, and difficulty in explaining villagers technical terms. Sometimes there is a delay in data collection due to lack of sufficient staff, non-cooperation by farmers, elections, and floods. These problems make census lengthy, error prone and costly. Further, the persistent fall in groundwater level means that any rapid change in number of these structures must be detected early on for the government to take regulatory actions to prevent competitive extraction and storage of ground water in farm ponds, leading to the tragedy of the commons (Prasad and Sohoni, 2018). This scenario calls for the development of an au- tomatic system for detection, mapping, and record- ing of irrigation structures. Using Remote Sensing, Computer Vision, and Deep Learning techniques, we have built an online web based object detection sys- tem that assists users to get locations and counts of these structures on a GIS based interface. Currently, our system provides detection of two important struc- tures, dug wells and farm ponds and we have incor- porated three deep learning architectures - YOLO, FasterRCNN and RetinaNet. We next discuss the mo- tivation behind choosing these structures. 208 Tundia, C., Tank, P. and Damani, O. Aiding Irrigation Census in Developing Countries by Detecting Minor Irrigation Structures from Satellite Imagery. DOI: 10.5220/0009421302080215 In Proceedings of the 6th International Conference on Geographical Information Systems Theory, Applications and Management (GISTAM 2020), pages 208-215 ISBN: 978-989-758-425-1 Copyright c  2020 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved