Protected agriculture polygons derived using freely available EO data in a variety of agroecological regions

Protected Agriculture (PA) is a practice used globally to increase productivity of certain agricultural crops. PA keeps the temperature more constant throughout the year, meaning growing temperatures are more favourable for plants leading to faster and more frequent crops. The issue with using this practice is that the conditions that PA infrastructure foster are extremely conducive to pest survival and propagation. The humidity in PA structures is generally much higher than the surrounding environment, all year round, which allows for the infection of fungal pathogens that can infect the crop inside the PA structures, as well as provide a critical overwintering location for fungal pathogens in the winter months which allowing pathogens to spread in the spring. 

Due to this fact, efforts are being made to ingest information relating to PA into epidemiological pest models, so that this emerging risk factor can be included when calculating pest risk. To date, there is very little information available on the spatial distribution of PA, where Earth Observation (EO) data offers invaluable insight to its distribution, given the scale of area capable in analysis and sensitivity to specific landcovers. 

There are several methodologies that have configured spectral indexes capable of vaguely distinguishing PA from other landcover types, but these often do not have the capabilities of modelling PA distribution consistently across different agroecological regions. There is also a methodology in the literature which is a global dataset of PA derived by using a combination of Sentinel 2 and commercially available Planetlabs multispectral data for the year 2019. This is an extensive and robust methodology for detecting PA, however it uses commercially expensive data which is prohibitive for many, as well as it is a ‘one off’ processing job and would be difficult to replicate for other years and obfuscating the inter-annual change of PA. 

The workflow derived to fit this goal uses a combination of Sentinel 1 and Sentinel 2 annual temporal composites derived on the GEE, where a two tiered classification workflow was trained using shapefiles hand drawn over PA infrastructure, regions of other landcover commonly found around PA, as well as regions where the first classifier was falsely identifying PA as NPA (termed a false positive). These polygons were drawn over a variety of agroecological regions in Australia, so that the trained models were robust and adaptable in a variety of spatial settings. The results from the 70/30 train test split were very good, and when the model was validated against an independent validation source (generated by CABI) showed that the overall accuracy of the workflow were also very good. But the ‘hit-rate’ was slightly lower – indicating that the workflow tends to miss PA structures. These were found to be thin PA structures, where the smallest axis of the structure was smaller than the spatial resolution of the input data, leading to the conclusion that this workflow is generally reliable and accurate but can miss smaller PA structures.  

This workflow represents an exciting prospect for assisting biosecurity and epidemiological modelling, as it is scalable, reliable across agricultural regions globally, and interannually applicable. The workflow has to date been applied to Adelaide, Perth, Sydney, Melbourne, Bundaberg, the entire state of Victoria in South East Australia, Crete and the ‘Mar de Plástico’ in southern Spain. 

This dataset represents results in all Australian examplar regions merged into a single shapefile, representing polygon derivations of PA structures within these regions for the year of 2023. 

Funding for the Allocated Work under this Agreement is made available by Science and Technology Facilities Council Grant: EO4AgroClimate Using Earth Observation data to improve datasets for biosecurity risk mapping of pest and disease and biocontrol suitability (Ref ST/Y00017X/1)