Guidelines on the use of technology for sample frame development
4.3. REMOTE SENSING
In this Handbook, RS refers to images acquired with a conventional camera or electronic sensors from aircraft or satellites. The scenes record radiation in several ranges of the electromagnetic spectrum, including the normal visual range, microwave radar, infrared, and ultraviolet. The techniques applied to process and interpret remote sensing imagery include visual photo-interpretation and a wide range of numeric algorithms. This section provides guidance on the use and choice of technology to develop a sampling frame and later in the estimation process.
4.3.1. Main types of satellite images
Most satellite images are produced by optical sensors that measure earth reflectance. To date, other types of images, such as Synthetic Aperture Radar (SAR) or Laser Imaging, Detection and Ranging (LIDAR) have had little impact on agricultural statistics. SAR images are linked with the roughness of the land cover, and LIDAR provides very accurate measures of distance. SAR images have the advantage of measuring through clouds. This should be a major advantage in areas with persistent cloud cover, but the high noise/signal ratio has limited their usability thusfar, except for the delineation of areas cultivated with paddy rice. It appears that the SAR images recently made available from the Sentinel 1 satellite present a major improvement, and pre-operational applications may be soon available. Currently, the most popular optical images are very high resolution (VHR) images, with a pixel size between 0.5 and 2.5 m. However, these images are often affected by strong limitations, for the purposes of agricultural statistics: full coverage of a given country tends to be too expensive and complex to manage, if built specifically to define a sampling frame or to produce estimates in a given year. However, if a full ortho-photographic coverage has been produced for other purposes and is available, it can be an excellent basis for stratifying an area sampling frame. An alternative that could be considered, in theory, is to construct an area sampling frame on the basis of a sample of VHR images. This option would requireusing PSUs of a size and shape similar to the VHR images and would lead to inefficient sampling schemes (Gallego, 2012; Gallego and Stibig, 2013).
Some leading companies in the information technology sector produce public-access image mosaics with global coverage (Google Earth, Bing). There are potential uses of these images for agricultural statistics. A major advantage is that they are available and easily accessible, with an efficient interface. Most agricultural areas of the world are covered byVHR images; this is a significant asset, especially for countries that seldom have recent homogeneous ortho-image coverage. However, these public image layers have some limitations:
• Image geometry. For example, the so-called Google projection is not an equal-area projection. It was conceived to optimize display speed at variable scales, but it is not optimal to provide comparable area measurements in different locations. In any case, the distortion introduced by the projection is likely to have a minor impact, for the purposes of agricultural statistics.
• Image overlay: when an older image is substituted with a recent one, the overlay between the two images displays a shift of up to 20m. This may introduce inconsistencies, if an older image is used to define a master frame and a different image is used to produce support documents to locate a given point or plot in a specific survey. • Heterogeneous image dates. Neighbouring areas may be covered by images that have been taken from five to
eight years apart. The impact of using images with such heterogeneous dates to define a master frame could be moderate if the field boundaries are relatively stable, but may be difficult to assess otherwise. Image viewing tools (Google Earth, Bing) report a date for the image on the screen, but this does not always coincide with the date on which the image was acquired.
Images with a resolution of 10-50m were called “high-resolution images” until the 1990s. At the time of writing, they are referred to as medium-resolution images. The most popular satellite series that provides medium- resolution images is Landsat. The Landsat-TM images have a resolution of 30 m, which since Landsat-7 has been complemented with a panchromatic (black-and-white) band with a resolution of 15 m. Traditionally, these are the most widely used for land cover mapping at national or subnational level, and especially for cropland identification. An important characteristic of satellite images is the swath, i.e. the width of the area covered by each pass of the
satellite. TM images have a swath of 180 km, compared to the 60 km of SPOT, another widely used earth observation series of satellites. A non-negligible difference between Landsat and SPOT is the distribution policy, since Landsat images can be distributed free of charge. There is a large number of satellites and sensors of the medium-resolution type, the availability of which is generally more heterogeneous. Hopefully, the June 2015 launch of Sentinel 2 with 10 m resolution and a 290 km swath will bring significant improvements.
Coarse resolution images with a pixel size between 250 m and several km generally do not enable single agricultural fields to be distinguished, except in countries with very large plots. They have the advantage of high frequency (usually, new images are available on a daily basis) and are a major tool in monitoring the status of vegetation, yield forecasting, and early warning; however, they are of limited interest in defining an MSF.
4.3.2. Aerial photographs
Since a wide range of satellite images is now available, the more traditional aerial photographs are receiving less attention by users, even though the popular Bing image database is mainly based on aerial ortho-photographs,which are usually clearer than satellite images.
In many cases, aerial photographs still present substantial advantages over satellite images. When a photogrammetric flight over a country is carried out, the entire area is usually covered in a relatively short period, with fewer problems due to cloudy areas. When a large area must be covered, they are often cheaper than equivalent VHR satellite images. Over the last few decades, ortho-correction algorithms have dramatically improved, which has enabled reductions in cost.
For the purpose of agricultural area frame surveys, aerial ortho-photographs generally provide the best field survey documents (even when the photographs are not recent), unless the landscape has undergone significant changes since the date when the images were acquired. For the purposes of stratification, they can constitute an excellent alternative to satellite images; however, the advantage of aerial photographyis not very clear, because very high resolution is usually not essential for stratification.
In recent years, the potential use of drones (unmanned aerial vehicles) is being widely discussed. The ortho- rectification and mosaicking technique is sufficiently developed to produce documents with acceptable accuracy, and the dates for image acquisition can be chosen with a flexibility similar to that presented by a field survey. A limitation of drones in many countries comes from flight regulations, that often forbid the flight beyond the sight of the operator. This limits the size and the shape of the area that can be covered by a single flight, preventing in particular the acquisition of long and thin stripes of images that would be much more efficient than compact areas. Thus, the limitation on the efficiency of drones may be similar to that of VHR images. The area covered by a drone during a flight can be seen as a surrogate of a segment. Drones are capable of providing images with a spatial resolution of approximately 5 cm. If the great majority of crops cultivated in a region can be identified with images having a resolution of 5 cm, drones could substitute field surveys. In areas with a complex crop pattern – in particular, with a significant amount of mixed crops – this is unlikely to happen.
Small low-altitude piloted aircraft provide images with a similar resolution (approximately 5 cm). They have the advantage of being more frequently authorized to fly long stripes (around 100 km by 100 m), much more efficient in terms of variance than approximately square units with the same area. Efficient stripe-based sampling plans can be defined for small aircrafts, especially for the estimation of nomadic livestock.
5. SUMMARY
This chapter provides an overview of the technology available for agricultural statistics and indicates those that are most effective for developing sample frames. Most of this technology supports the development of sample frames; however, it can also be used effectively for land cover classifications that can be linked to census or administrative areas.