SALSA crop type maps for small farms

For the purpose of the project, small farms were defined for statistical purposes as those with less than 5 hectares or with 8 economic units.
The project used also a participatory approach that ended up with a more complete, not statistical, definition depending on the position of the farm in the food chain/system. The approach was used to select small farms to be included in the SALSA survey, and it was flexible to consider what in the context of each region and in relation to its farm structures, was relevant to be classified as a small farm, and was accepted as such.

The first step involved intense fieldwork for crop type data collection to be used for the calibration/validation datasets in the image classification procedure. Due to the very high costs of field campaigns to collect enough calibration/validation data, a methodological approach to collect reference crop data was developed and implemented to select a minimum number of points to be collected in each region. 

A six-stage methodology to define the spatial distribution of the sample points was used with the aim to combine the agricultural landscape diversity and the accessibility of each square: 

1. Create fishnet with 2x2 km size using the ArcGis tool ‘Create Fishnet’ in Data Management Tools. 
2. Unsupervised Random Forest classification. 
3. Compute Shannon Evenness Index (SEI) to assess agricultural landscape heterogeneity - for the definition of the Index, see:  
    https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Glossary:Shannon_evenness_index_(SEI). 
4. Compute road distances to assess the degree of accessibility in the squares. 
5. Square selection, ensuring high diversity and good accessibility.   
6. Point selection, for each of the above squares. 

Different teams in the field were used in the different regions. The aim was to produce a common output, consistent across reference regions, the guarantee of a quality product was a key step. For the quality control, an average of 8.5% (std = 2.75%) of the sample points (min= 5.1% and max= 15.8%) were checked in 16 out of the 21 reference regions. The control points were randomly selected from the previous ~500 points selected for each region. At least one point per square was selected.

The methodological workflow applied to produce the crop type map through image processing involved 10 steps: 

1. Download of the multi-temporal Sentinel images and pre-processing;
2. Vegetation indices computation to be used as auxiliary information in the classification;
3. Create multitemporal NDVI stack for the segmentation stage;
4. Image segmentation;
5. Prepare the crop dataset to be used in the classification;
6. Select the segments intersected by the crop field points;
7. Extract all the pixels within the segments to be used in the classification procedure;
8. Image classification (pixel based) using a Random Forest machine learning algorithm;
9. Accuracy assessment of the classification;
10. Build the small farms crop type classification map. 

For the crop type map image processing and classification, all the available Sentinel-2 images for each reference region between April and September 2017 were downloaded from the ESA’s Sentinel Scientific Hub (ESA, 2017).