Nigeria for CSSL
Part of the STARS work portfolio is the development of the Crop Spectrotemporal Signature Library, the CSSL. The purpose of the CSSL is to become a repository of crop-characteristic data, as crops are followed through the growing season. Such data in part will be derived from satellite and UAV images, and in part from fieldwork. For the initiation of the CSSL within STARS, we have opted to make the farm management unit (fmu) the central notion of study, and so this first version of the CSSL stores data per fmu.
Within the STARS project, our ambition is to have the CSSL include the fields that are covered by the respective teams in Mali, Nigeria, Tanzania, Uganda and Bangladesh. In most of these regions, intensive in situ monitoring has taken place, with image acquisitions at roughly the same time. Much of that data is now under scrutiny of the various research teams, and by end of project we hope to bring together all their findings.
The CSSL will start to prove more useful as its number of entries, say counted as fmu-seasons, grows; this is the first law of statistics. (For those who want to go deeper: Mosteller and Diaconis are the names implied with this …) But this also gives friction: intensive fmu monitoring is both expensive and time-consuming, and does not build up quickly datasets with high numbers of fmu-seasons. Where one strives for larger geographic coverage in smallholder agriculture monitoring, the in situ monitoring cannot be sustained for all fields under study.
It is from this perspective, that we welcomed so dearly the efforts made by the ICRISAT team in Nigeria in the 2015 season. Their intensive campaign accumulated data about 5100+ farm fields around the city of Kofa, Nigeria. This was not an intensive per-field monitoring effort, but one with a single visit to characterize each field, and after which that characterization was registered on the JotBi platform.
The ITC team saw great potential for this dataset in context of helping to grow the CSSL, because of the large number of fields. Perhaps less studied, less well-monitored but sufficiently attributed as candidates to be included in the CSSL. Rather typical for datasets collected in this rapid assessment way, the spatial data quality gave concerns, and we felt that human interpretation was needed to identify systematically the actual fields with the polygons obtained in situ. In this first phase, we geometrically corrected field boundaries manually.
- Figure 1. On left, original spatial data from ICRISAT field campaign; on right, geometrically corrected data from nearby location to the west. Boundary colours reflect croptype classes.
We have almost completed this first phase now, and only some 500 fields are left to be corrected. In the second phase, we will perform quality control on the resulting polygons, and we will identify those fields that are actually composite: they have multiple fmus inside them. The plan is to use the second ITC crowd-source mapping party to have volunteers digitize split lines for such field composites, allowing subsequent splitting to single fmus.
Most of you are well aware of the overall image processing workflow that we will soon make public in STARS context, as an open-source endeavor. It delivers images fully pre-processed for many practical follow-up analysis purposes. More on that workflow in an upcoming newsletter item, so let me not spoil that bit of news.
At the same time, our work on the image workflows aims to bring an extension of a few more processing steps to the STARS image workflow, specifically designed to feed the CSSL. They will allow us to derive as pure as possible spectral and textural characteristics of crops per fmu in the dataset, which will then be stored in the CSSL. The image processing steps that we are currently working on include cloud masking, cloud shadow masking and tree masking. None of these are finished at time of writing, but we are zooming in to workable and fairly robust solutions for them. Once they are in place, we will derive the per-fmu statistics, and this, we expect, will allow us to enter the exciting phase of initial CSSL exploitation. At that stage, with large enough datasets to allow such exploitation.