Crowdsourcing experiment to enrich field data collection
After the successful first crowdsourcing mapping party last year, a second mapping party was held on 20th April 2016. The purpose of the mapping party was to extract field-level crop data from fine spatial resolution imagery through visual interpretation and feature digitization. Crowd sourced data would then be used as ground truth data and also for validating automatic field data extraction algorithms developed by ITC researchers. All these activities were geared towards achieving the STARS project goal of looking for ways of using remote sensing technology to improve small-scale holder farming practices in Sub-Saharan Africa and South Asia.
The focus area of the mapping event was Kofa district, northern Nigeria (Fig1); which was one of the STARS project flagship sites. The field data collection was carried out using mobile-based data collection devices. During the exercise, enumerators digitized field boundaries and entered information obtained from farmer interviews for each of the fields visited. This was carried out during the 2015 cropping season and covered an area of approximately 1020 hectares. The challenge with such a field data collection exercise is that errors are bound to be made during field boundary digitization and to some extent incorrect field attributes entry. This made the process of data cleaning tedious and time consuming and therefore crowd sourcing was considered an option in speeding up the data cleaning process.
The target participants were ITC’s students and staff. The mapping party was opened by ITC’s STARS Project manager Mr. Stroeven who gave an introduction to the STARS project and highlighted the goal of the mapping event including specific tasks to be performed. Thereafter, Mr. Kibet who organized the mapping event logistics took the participants through each of the tasks and provided materials with instruction for each task to be performed. The key tasks were to digitize crop row width, roads and paths as well as identify crop labels within the digitized field polygons.
The digitization of crop row width was considered an important input for modeling crop yield and crop identification. Likewise, digitizing roads and paths would be useful in determining field accessibility, which is one of the indicators in computing cost of production at field level. The third task was crop identification and field labeling which was undertaken to independently collect crop type by carrying out visual interpretation of field texture, crop row spacing and color differences based on multi-temporal pan-sharpened and multi-spectral WorldView imagery. The target crops which could be visually identified included maize, millet, sorghum, legumes, rice and onions. Field labeling was included in the task so as to identify crop field status if pure (one crop within a polygon) or composite crop (more than one crop with distinct boundaries contained in one polygon). In addition, field status identification aimed at distinguishing between a finely delineated crop field (the digitized boundary follows the correct crop field boundary) and a coarsely delineated field. Printed and digital copies of the protocol of each activity were issued to the participants. All the tasks were performed using QGIS open source software.
Participants of the mapping event using printed and digital copies of the protocol.
A total of 23 participants attended the mappinp event.
The initial plan was to allocate tasks in such a way that 14 participants would digitize crop row width, 6 would digitize roads and paths and 36 would visually identify crops and label fields. However, considering the number of participants who attended the event, the focus was shifted to crop identification and field status labeling. Total of 23 participants attended the mapping event. They were grouped into 6 teams of 3 participants; with each team working on a single crop. The remaining 5 participants were distributed randomly to each of the groups. Every team member was given a section of the study area to work on. The idea was to avoid duplication of effort, except at the border of two areas split from the main study area where an overlap of the crops fields was provided. This was to allow acquisition of data for assessing accuracy of the crowdsourced data from more than two participants working independently.
Participants demographic data obtained from an online feedback of 16 participants who filed the form indicate that the participants were drawn from 9 countries. In terms of gender, 11 men and 5 women participated with most of the participants being bachelor degree holders (9) against 7 master’s holders. Majority of the participants (14/16) participate in a crowd sourcing event for the first time. In terms of their level of experience with using QGIS, 63 % of the participants indicated little or no experience while 38% indicated basic experience while 9% were experienced. Total of 94% of the participants recommended the use of QGIS for future crowdsourcing mapping events with only 6% having a dissenting opinion. In overall, the participants showed interest in the event and commented to have gained some level of experience in carrying out crop identification as well as using QGIS software. However, there were challenges with using different version of the software, loading data to the flash drives and providing guidance to the participants as most of them were not conversant with QGIS.