Earth Observations for Crop Area Statistics in Africa. The GMFS Project Experience in Malawi
Ceccarelli, Tomaso1; Haub, Carsten2; Remotti, David1; Chirwa, Isaac3; Giovacchini, Aldo1; Gilliams, Sven4
1Consorzio-ITA, ITALY; 2EFTAS, GERMANY; 3Ministry of Agriculture and Food Security, Malawi, MALAWI; 4VITO, Flemish Institute for Technological Research, BELGIUM

There are four key areas where earth observations (EO) data can contribute to crop area statistics through geo-referenced data frames:

1) The design of the sampling frame (based on points or segments), and its stratification through pre-existing land cover maps or visual interpretation of satellite imagery, based on generalized land cover classes,
2) The ground survey preparation which includes optimization (through careful re-interpretation of the points to be visited on the ground), planning and cartographic preparation,
3) The orientation of the surveyors on the ground (done with the help of photomaps from satellite imagery),
4) The improvement of the crop area estimates (through reduction of their variance) by means of statistical estimators of different nature, relating the ground observations with satellite image classifications of individual crops.

For 1), 2) and 3) Very High Resolution (VHR) imagery would be preferable, although neither necessarily taken during the cropping season nor in the same year of the survey. For 4) it is a requirement that images are taken during the cropping season and preferably (also depending on the variability of the growing cycles) more passages over the season are needed. The spatial resolution depends on the size of the parcels or of clusters of parcels with similar cropping cycles and hence on the types of crops, farming and land tenure systems.

Geo-referenced sampling frames using EO have been applied only in a few cases in Africa until today (e.g. Morocco, South Africa, Kenya, Ethiopia), one of the main reasons being the high cost of the imagery and the skills required in their processing and analysis. However, new opportunities of accessing EO data are nowadays brining this approach closer to an operational and cost-effective perspective in the generation of crop area statistics. This is also recognized by the "Global Strategy to Improve Agricultural and Rural Statistics" developed, among others, by the FAO and the World Bank. In this context EO is regarded as a new opportunity, to be further researched and tested in the African context.

The Global Monitoring for Food Security (GMFS) project funded by the ESA as part of the GMES Programme includes, among the others, the "Agricultural Survey Optimisation service. This consultancy service, which is provided in Malawi and Sudan, has the objective of supporting the respective Ministries of Agriculture in improving the quality and cost-effectiveness of their agricultural statistics, particularly crop area estimates. The hypothesis is that the service can assist in achieving the mentioned objectives through geo-referenced sampling frames, combining EO data and ground observations. In Malawi, a pilot survey based on a point frame was carried-out in 2012 covering the two districts of Dowa and Ntchisi in the Central Region of the country. The survey was conducted by the Ministry of Agriculture and Food Security (MoAFS), with the technical assistance of Consorzio-ITA and EFTAS on behalf of the GMFS Consortium. For Dowa the completeness of the ground survey (in terms of points surveyed on the ground) was around 86%, which was considered as a very good achievement. On the opposite, for Ntchisi (due to logistic and organisation problems), results were regarded as being possibly biased, due to the high number of missing data (39%). The estimation of the crop acreages were therefore generated only for Dowa. A CV of 3% was obtained for maize and CVs of < 10% for other major crops (groundnuts, soybeans, tobacco), at District level. Simulations based on the 2012 survey and optimized according to Bethel's allocation algorithm indicate that a smaller sample would be sufficient to attain target CVs at regional level: with around 4,500 sampling observation points, estimates can be achieved for the Central Region with a CV of around 2.5% for major crops such as maize, 7-9% for major crops, and 12 % for other crops of importance. Lower precisions can be achieved for minor crops. Moreover, at the district level, a CV of 5-6% can be achieved for a key crop such as maize. These precisions in the crop acreage estimations are regarded as being sufficient for the main purpose of the annual crop estimates, which is food security.

EO data were used in all the four steps previously mentioned. Their utility has been positively assessed for steps 1) stratification, 2) ground survey preparation and 3) orientation of the surveyors on the ground. As far as 4) improvement of the crop area estimates, a specific activity was carried-out together with the Joint research Center in Ispra and SPOT 5 images ("in-season", "single-date", 10 and the 5 m resolution ORTHO products) were classified. The classification was used as an input in the estimation of the crop areas. The accuracy results of the classification were not satisfactory, due to the complexity in the phenology/cropping cycles of the crops in the study area, as well as the small size of the parcels. This was reflected in a relative efficiency (ratio between the variances of estimates with/without the contribution of the classification) of 1.07, which represents only a slight improvement.

A multi-temporal series of images during the cropping season is expected to improve the classification results, although this may still represent a challenge for optical satellites in Malawi due to persistent cloud cover during the growing season. In order to overcome this constraint GMFS is implementing another service in Malawi where SAR and optical satellites are integrated in order to attain a better discrimination of the major crop types.

In this perspective it is expected that the Sentinel Missions will bear a significant impact. This is in relation to the spatial resolution (10 meters for most bands of interest in the classification) and revisiting time (5 days) of Sentinel 2, as well as to the opportunity of integrating SAR technology through Sentinel 1.