In the context of SHA, high spatial resolution RS data is required for mapping farm boundaries due to the small sizes of farms. Individual farm sizes typically range between 0.2 and 2 ha. Previous studies have shown that satellite images with a spatial resolution of 5 m (or less) can adequately capture the typically small farm sizes and subsequently allow an accurate estimation of cropped area (Forkuor et al., 2014).
A simple approach to delineating farm boundaries is to manually draw its outline on the image – the so-called on-screen digitizing (De Wit and Clevers, 2004; Möller et al., 2007). The prerequisite to this is an accurately orthorectified satellite image of the area of interest and being able to clearly see farm boundaries on the image. Sometimes, it is difficult to identify clear boundaries between farms when several farmers cultivate the same crop in an area. Although some remote sensing applications/software permit on-screen digitizing, this process is best done in a Geographic Information System (GIS) environment. Despite being simple, this approach is suitable when the boundaries of only a few plots are intended to be digitized. In other words, it could be time-consuming if one is dealing with a large area.
In case of large areas, automated delineation of farm boundaries can be performed on RS data using standard algorithms. Image segmentation is a popular approach in this regard (Aksoy et al., 2012; B. Chen et al., 2015). The approach delineates field boundaries based on multi-temporal information, spectral contrast between fields as well as shape, size and texture of the fields. Accurate results can be obtained when there is sufficient spectral and/or textural contrast between fields and the spatial resolution of the RS image permits identification of the typical field boundaries in the area of interest (Yuan et al., 2014).
However, depending on the nature of the landscape, this approach may not always lead to optimal results. In most of Africa, for example, a relatively high tree coverage and/or poorly defined field boundaries in some areas limit the application of such approaches (segmentation) in delineating field boundaries (J. Chen et al., 2015; Forkuor et al., 2014).
Segmentation algorithms can be applied to UAV and satellite images to delineate field boundaries. For example, in Tanzania, delineation of agricultural field boundaries using UAV imagery has shown that segmentation parameters have to be chosen carefully. Even when the minimum polygon size is set to the average field size present in the landscape, many actual field boundaries are not detected due to the spectral and structural similarity of neighboring fields (Figure A). Decreasing the minimum polygon size results in capturing smaller agricultural fields but also causes larger fields to be segmented into smaller polygons as well (Figure B). An additional challenge are cloud shadows, cutting through agricultural fields. Shadow edges are often confused with field boundaries by the segmentation algorithms. Furthermore the minimal reflectance in shaded areas leads to a loss of information, which is then lost to the classification algorithm. The results have shown that field boundary delineation requires a semi-automated approach, where image segmentation is the first step, which then needs to be corrected manually by an experienced analyst who is familiar with local conditions.
In recent years, crowd-sourcing approaches have become another means of delineating farm boundaries. This approach allow its participants to interpret snippets of satellite imagery for land cover type and size of agricultural fields in an area they are familiar with, via the internet. High spatial resolution images (e.g. from Google Earth) form the basis for the delineation. Due to high internet penetration in many parts of the world and the increasing availability of high spatial resolution, this approach is increasingly becoming a popular option for obtaining boundary and land cover information over large areas (Fritz et al., 2009).
Obtaining accurate statistics for cropland and other land cover types from remotely sensed data is difficult for many nations. The dramatic increase in the availability of very high resolution satellite imagery with 2 m pixel resolution or finer, and the significant reduction in prices has enabled new opportunities for scientific analysis and practical applications, however the sheer data volume and level or detail presents challenges. The use of fully automated algorithms quickly reaches methodological limits over large regions and require significant expertise and computing power. The Geo-Wiki platform provides an innovative solution to the problem of interpreting large volumes of very high resolution imagery. Geo-Wiki provides citizens with the means to engage in environmental monitoring of the earth by providing feedback on existing spatial information overlaid on satellite imagery or by contributing entirely new data. Data can be input via the traditional desktop platform or mobile devices, with campaigns and games used to incentivize input.
In a study case of Tanzania, DigitalGlobe provided a web service of high resolution satellite imagery for parts of Tanzania. Geo-Wiki analysts from IIASA (International Institute for Applied Systems Analysis) filtered through the archive of images (dated between 2010 and 2014) and extracted 30,000 1 km x 1 km tiles depicting cropland and woodland from within the growing season. The district of Kilosa was selected as the focus area with wall-to-wall coverage. A random sample for the rest of Tanzania made up the remaining tiles. These images were downloaded into an offline client application developed by IIASA for visualization and field validation.
During two workshops, students of Sokoine University of Agriculture in Tanzania were presented 1 km x 1 km image tiles and asked to determine whether the center area of the tile was covered by cropland, woodland, both or neither, the dominant field size and the cultivation stage (cultivated, fallow, mixed, unclear or uncultivated). In order to entice the students to classify as many images as accurately as possible, a competition was used whereby the top 3 candidates with the most accurately classified images would win a prize.