In-situ data for EO in agriculture

In-situ data for EO in agriculture can be crop type data, phenology, agricultural practices, biophysical variables (e.g., LAI, LAIeff, FAPAR or FCOVER, soil moisture, yield, crop height, density,…), or other data collected on the ground or obtained from data analysis (e.g. photo-interpretation) (ESA Land Training).

Calibration and validation field campaigns for cropland and crop type can be all combined, but the sampling design should be explicitly different to be independent. In agricultural mapping, sometimes two field campaigns are needed for early and end-of-season maps when all crop types cannot be identified during the mid-season campaign.

Requirements for in-situ samples

Gathering in-situ data, for example, to validate a crop classification map, is not a trivial task. The sampling design should fulfill the following criteria (Stehman & Foody, 2019):

1. Satisfy the conditions defining a probability sampling design
2. Easy to implement for both selecting the sample and producing estimates for accuracy
3. Cost-effectiveness
4. Readily allows for increasing or decreasing the sample size
5. Precise in the sense that estimates have small standard errors
6. Unbiased estimator of variance
7. Spatially well-distributed samples across the study area

A typical approach can be conducted in the following three steps:

Stratification

Stratification according to existing agroecological zoning to sample the range of diversity.

On-screen visual interpretation

On-screen visual interpretation to select samples (min. 1 ha) of land cover types different than cropland on recent aerial photographs, Google Earth, or Bing imagery to capture the diversity of non-cropland land cover types. The sample distribution between strata could also consider the stratum size and their respective diversity.

Ground survey

Ground survey to delineate crop type samples (min. 1 ha, but larger is better): For each stratum. No strict sampling design but need to capture each crop’s diversity. Visual delineation could use the most recent color composite.

Innovative methods to gather in-situ data

Challenges to data collection

While a sampling design may be straightforward for a smaller study area in a controlled environment, a larger study site size, inaccessibility of parts of the area, and temporal variability (crop phenology, management practices), can complicate things. Thus, challenges for in-situ data collection are:

• To capture the entire diversity of the agriculture patterns
• To cover the full diversity of the crop types of interest and all the minor crop types (it is important to define the targeted crop types to emphasize the field data collection on them and to capture all the gradients in crop growing conditions of the crop types of interest)
• To integrate the quality of the in-situ data (labeling error, location error, editing error, etc.) as the in-data quality is of paramount importance
• To ensure the spatial distribution of the in-situ data across the whole study area to avoid local overfitting of the classification model to specific growing conditions
• To use spatially independent in situ data for robust accuracy assessment (the closer the calibration and validation data set are in space, the more the accuracy metrics represent the quality of the classification model rather than the quality of the entire map)

Novel ways of data collection

Where, from a statistical point of view, traditional sampling with random or stratified sampling schemes may be the most adequate ones to fulfill the above-mentioned criteria, conventional ways to sample ground data often lack the scale that is needed for modern applications. Therefore, novel ways to survey agriculture in a resource-efficient manner are increasingly explored. These include new ways of conducting surveys, using administration data, as well as crowdsourcing, for example through social networks, and farmers‘ data from farm management tools. Find some examples of each of these in-situ data types below.

The windshield survey

One example is the windshield survey, which was tested along with random and transect sampling schemes in Belgium, Russia, South Africa, and Ukraine in the framework of the Joint Experiment on Crop Assessment and Monitoring (JECAM). A sampling methodology as shown in the figure below was applied.

The results below show that roadside sampling did provide representative and accurate results, which critically simplifies field data collection, as roadside data collection is an easy and cheap type of survey to collect crop-type information.

The Land Use/Cover Area frame Survey (LUCAS)

Administration data is data that is collected by public authorities. It can often be used for a different purpose than it was initially collected for.
LUCAS is a land use/land cover data collection exercise of the European Union and spans its full coverage evenly. Data has been collected in 2006, 2009, 2012, 2015, and 2018, with the 2022 campaign being in preparation.
Initially designed for statistical purposes, the LUCAS data collection protocol was complemented by a Copernicus module for the first time in 2018, which corresponds to the Copernicus satellites’ spatial resolution of up to 10 m and can be used for agricultural applications like crop type mapping with Sentinel-1 imagery.

Crowdsourcing

Crowdsourcing refers to the creation of data by a large group of non-professionals. It can be active crowdsourcing, where the crowd is asked to collect specific information. Passive or opportunistic crowdsourcing means that people share information without knowing what the data is used for (see the paper from Minet et al. (2017)).
An example of both active and opportunistic crowdsourcing was the campaign #YellowFlowersEU, where people were asked to share pictures of flowering rapeseed in 2018. Citizens, scientists, farmers, and photographers contributed to a database with rapeseed flowering dates across the EU by posting pictures of yellow flowers on Twitter.

Farmer’s data

Farm management apps can offer the possibility to share data and provide in-situ information that can potentially be used in combination with remote sensing data. However, the question of data ownership can be a hindrance: Who owns and can share the data? The farmer, the platform provider, or the equipment provider?

The CEOS Cal/Val working group has assembled a repository with the scientific literature on the calibration and validation of good practices, guidelines, and standards for specific topics here (d’Andrimont, 2019; d’Andrimont et al., 2021; Defourny, 2019; Gallego & Delincé, 2010).