Field Mapping

Using QGIS to support hyperspectral remote sensing.

‘Ground Truth’ is often required to build and validate predictive models, and so is the foundation of many remote sensing projects. In this article we explore the methods 2Excel Geo use to accurately map data in the field. Interested readers may also wish to read my related two-part series which discusses effective methodologies for field spectroscopy (Part One & Part Two).

‘Ground truth’ is a short-hand term used to describe the metadata (location and attributes) of known samples. Even if samples can be resolved in the imagery itself, the attributes can often only be determined by a domain expert on the ground. This then requires field work, involving staff with the training and equipment to collect accurate ground truth. As the starting point for building predictive models, metadata and spatial accuracy is essential, since any errors may be propagated throughout the endeavour.

In order to facilitate the capture of ground truth, we utilise QGIS (an open source Geographic Information System), installed on a ruggedised, GPS-enabled, touch-screen tablet, which is resistant to rain and dirt. The resulting set up gives a field worker access to any imagery or metadata that can assist in locating in-scene targets, and powerful customisable or bespoke tools to aid in capture of metadata.

Figure 1

Mapping heather (Calluna vulgaris) using custom data entry on a rugged tablet.

Spatial Accuracy

In the context of large-scale mapping projects, the field workers’ task is to collect samples of a number of land-cover classes. The number of classes, the number of samples per class and the spatial extent of the field collection are all project and landscape dependent. The minimum and optimum requirements are discussed and agreed with the project team before field work commences.

As discussed in the articles on field spectroscopy, the requirement to match ground truth with our hyperspectral airborne data, with a typical resolution of 30 cm, places a stressing demand on the spatial accuracy of field mapping. Hence, our approach is to collect the airborne imagery prior to field collection. This provides the field worker with the best chance of associating ground features with individual pixels in the hyperspectral imagery.

Here, the ruggedised field tablet is invaluable. The field worker can use pre-loaded datasets – maps, surface models, databases, as well as the hyperspectral imagery – to assist in the selection and location of ground truth samples. Initially, the built-in GPS may be used to provide location to within a few metres. The field worker must then interpret the imagery to recognise colours and textures to locate the exact pixels of interest. If it is not possible to do this with sufficient confidence, then this sample is omitted from the ground truth. Back in the office, to avoid any mislabelled ground truth, pixels around the boundary of each mapped polygon are discarded and the spectra of the remaining pixels are inspected for any outliers.

Attribute Accuracy

While matching real-life features to exact pixels is difficult and sometimes tedious work, recording accurate attributes for each sample is equally important. The attributes required can be both qualitative (observational data such as vegetative species) and quantitative (measured data such as Leaf Area Index using our SunScan Canopy Analyser). Domain expertise is often required to provide accurate observational ground truth, and so, suitably qualified individuals will often accompany our field mapper.

Another important consideration is standardisation of the protocols used in recording metadata, and of the data itself. Poorly executed field data collection will inevitably lead to poor outcomes. To ensure all data is consistent, accurate and unambiguous, we develop and adapt standards for every domain of interest. An example of this is our campaign to classify tree species.

A list of trees in the United Kingdom was compiled and a standard name agreed to describe both scientific and common names. A database was implemented with each record describing an individual tree, including a bounding polygon and taxonomic classification (species, genus and family). This provides a reliable and immutable reference for field workers and analysts alike.

To populate this database we designed custom data entry forms in QGIS (figure 2), with ease of use in the field of paramount importance. These forms ensure data entry is conducted in a standardised form and prevents incomplete entries.

Figure 2

An example of a simple QGIS user form which aids in ensuring metadata quality and integrity.

By using QGIS in the field, with domain experts and skilled mapping staff, we have been able to successfully map a range of different target sets in support of hyperspectral remote sensing research campaigns across several different domains.

ARTICLES

Starting an Internship During Lockdown
Which Remote Sensing Approach is Right for Me?
Life as an Intern at 2Excel Geo