In this month of Planet’s Agile EO Webinar Series, we covered our newest Planetary Variable, Field Boundaries (FB). This is the first PV that we have built in Slovenia after the acquisition of Sinergise. FB contains a dataset of polygons that represents the boundaries of agricultural fields for your areas of interest anywhere around the globe. Agricultural field boundaries serve as the foundational framework for a wide range of analyses from food security to supply chain traceability to commodities trading.
You can watch the on-demand recording here.
If you want to learn more about Field Boundaries, please contact a representative in your region: North and South America - matthew.walkley@planet.com, Asia Pacific - janelle.ting@planet.com, and Europe/Middle East/Africa - lukas.aitken@planet.com.
Below you’ll find some FAQs and top questions from the live webinar.
Questions & Answers
Definition and characteristics of field boundaries
- What is the minimum field size that is reliably detected?
Fields with area lower than 50 sqm are removed from the output, while fields with Max Inscribed Circle diameter (MICD) lower than 30 m are flagged as unreliable, i.e. qa=1. The FB performance is dependent on the field’s shape and size, and MICD better represents these than area alone. However, FB can be reliable for fields with an area greater than 0.1 ha based on the shape of the field.
- Why do you base this product on Sentinel-2, and not on Planet Dove?
Historically, FBs were created by Sinergise within an European Commission funded research project on Sentinel-2, and the results for large fields are reliable even with using Sentinel-2 data.
- What is the output resolution of the field boundary maps?
The pixel size of the output (prediction) map is 2.5 m, which is then thresholded to derive FB contours.
- What is the availability around the globe?
Area of Interest (AOI) can be specified by customers worldwide, but it is subjected to availability of cloudless Sentinel-2 imagery.
- Is the data ground validated?
We perform curation of the training data via a combination of visual inspection and automated filtration using machine learning algorithms.
- You have a probability raster you convert to vector, what's the probability threshold to call it a 'field' in the vector dataset?
The probabilities for extent and boundaries are combined into a joint map, which does not represent probabilities anymore, but is nevertheless thresholded at 0.75 (on a scale from 0 to 1).
- For precision agriculture, we are more concerned with crop boundaries - which exclude buildings, ponds/lakes - occlusions can be at the edge or internal in the boundary. Is there a difference in your context?
From the validation of FB, we noticed that the deep learning model typically does not include in the estimated field extent non-arable land cover, like buildings, trees, water, windmills. The estimated FB therefore typically denotes only arable land.
Data processing and methodology
- Can you distinguish crop types?
Field Boundaries dataset contains vector polygons marking the extent of fields. This dataset does not provide information about specific crop types.
- Does your grid application follow PLSS in the US?
The grid used to parallelize the processing of FB is global and based on the UTM coordinate system.
- How well does the separation of cropland and non-cropland work in your model across regions?
Since the current model has been trained on data for Europe, its generalization to very different climatic regions is not optimal, in particular in desert or arid regions. The global demo results can help understand how the model performs for cropland and non-cropland across the globe.
- Which Deep Learning pre-defined model is used to create this product?
The deep learning model is based on a U-net architecture, which is trained from scratch on Sentinel-2 and EuroCrops data.
- Do you have any measure of the quality of these boundaries when compared to the actual field boundaries?
We evaluated the FB product extensively on the largest source of reference data in Europe, i.e. EuroCrops dataset. The dataset is based on GSAA polygons, which have known quality limitations (described in Utility of AI4Boundaries and EuroCrops as training datasets for field delineation blog post), leading to under-estimation of the delineation performance. However, such a dataset allows us to evaluate the algorithm on a diverse set of samples, in terms of sizes, shapes and semantic definitions. As can be expected due to the limitations of the spatial resolution of the input imagery, the performance increases as the size of the fields increases, with median IoU scores greater than 0.8 for fields having minimum side larger than 150m, reaching a IoU up to 0.98 for for fields with minimum side larger than 240m.
- What additional attributes (RGB values, NDVI, etc.) of the polygons can be extracted?
For the moment, only FB polygons and attributes describing their shape/size and quality are provided. However, these polygons can be directly used in other tools offered by Planet to derive image-based time-series data.
- Could you provide some information on the latency of this product? How fast can it generate the boundaries for a region?
In general, the larger the area the longer the required processing time, although the algorithm is highly parallelized. In general, the processing lasts from a few hours to a day.
- What are your model inputs?
The inputs to the model are the 10-meter Sentinel-2 bands - R, G, B, NIR.
- Is it possible to see the real satellite images and how often are they updated?
The FBs are created for the area of interest and time of interest that the customer selects. The FBs are generated once, therefore there is no updating unless the customer orders a new set of FBs. Only the FBs, not the imagery is delivered. However, customers can use the FBs to order other datasets such as PlanetScope imagery, Crop Biomass, Soil Water Content, etc.
Geographic coverage and performance evaluation
- The appearance of agricultural fields varies dramatically across the globe, particularly for example in smallholder farming systems in the Global South, in terms of shape, size, and spatial distribution. What is the performance in, say, India, where there are less well-defined wall-to-wall fields in an image?
The performance of the FB algorithm outside of Europe has been only qualitatively assessed due to lack of reference data. A demo application on test global areas shows how FB generalizes across the globe, including India, China, Africa, Australia. Please contact us for more information about your area of interest. South and North America - matthew.walkley@planet.com. Asia Pacific - janelle.ting@planet.com, and Europe/Middle East/Africa - lukas.aitken@planet.com.
- For those of us working in smallholder agriculture, can you translate that 30 m maximum inscribed circle diameter to an approximate range?
Maximum inscribed circle is the largest circle which can be inscribed within the polygon. Therefore when we say 'Field with Maximum Inscribed Circle Diameter less than 30 m,' it essentially means a field with the minimum/shorter (narrower) side of less than 30 meters.
Pricing, access, and restrictions
- How is the data priced?
Pricing is based on square kilometers, with volume discounts available for larger areas.
- Who can access the field boundaries product?
The Field Boundaries product is accessible to every Planet customer.
- Can we review the data before we buy?
Data cannot be reviewed prior to purchase since the product is generated on demand based on the specified area of interest provided by the customer. However, we provide a demo application to visually assess the performance of FB on test areas across the globe.
Customer demographics and usage
- Who is the typical customer? What problem does this solve?
Field Boundaries are a critical component to many agriculture solutions. Customers are generally coupling Field Boundaries with other Planet data such as PlanetScope imagery, Crop Biomass, Land Surface Temperature, and Soil Water Content. This supports work on large area analytics such as crop identification, yield estimation, food security, subsidies validation, and supply chain monitoring. We have also seen customers use this as an entry point for onboarding new growers into precision farming platforms or for benchmarking.
- Can we have access to just the ground-truthed original training polygons which were used to train the DL model?
EuroCrops dataset was chosen for the training and validation of our model because of its comprehensive coverage of publicly accessible GeoSpatial Aid Application (GSAA) data from across Europe. It is publicly available on Zenodo.
