As part of the final BETTER Hackathon, project partners prepared 4 hackathon exercises. WFP organised this exercise as the challenge promoter for the Food Security thematic area.
This open exercise featured the use of Binder and purposely provided cloud resources. Participants were expected to be familiar with the Jupyter environment (Python 3) and the most common EO libraries (e.g. GDAL) and were guided to use their favourite approach (e.g. pixel-based or object-based classification) to derive a crop type map for the region using the following combinations of datasets: S2 unfiltered – benchmark, S2 filtered, S2 unfiltered + SAR, S2 filtered + SAR. Libraries used included Rasterio / GDAL, pandas + numpy, scipy, numba, keras / tensorflow / opencv. The recorded part includes the introduction of the exercise in the context of the BETTER project.
1. Enhancing Agricultural Mapping with BETTER Pipelines
WFP - BETTER Hackathon – 23rd October 2020
Big-data Earth observation Technology and Tools Enhancing Research and development
http://ec-better.eu
This project has received funding from the European Union’s Horizon 2020 Research and Innovation
Programme under grant agreement no 776280
2. Relevance for WFP
Changes in Agricultural Area under Conflict
Crop type mapping in conflict areas:
Quantify losses of cropland and post-conflict
regeneration. First time cropland baseline
(e.g. South Sudan)
Technical assistance to governments:
In close cooperation with Min of Agriculture.
Smallholder support systems:
Assess seasonal performance potential,
likelihood of meeting procurement targets,
changes staple-cash crop in area of
intervention.
WFP Project Interventions:
Agriculture oriented FFA projects, resilience
building interventions, index insurance.
BETTER | Enhancing Agricultural Mapping with BETTER Pipelines
3. Sentinel-2
Data Assimilation Cloud
Storage and
Processing
Centre
Crop Type Map
ONA Platform
Assimilation, Quality
Control, Conversion
Sen2Agri System
BETTER | Enhancing Agricultural Mapping with BETTER Pipelines
Crop Type
Ground Data Collection
4. BETTER | Enhancing Agricultural Mapping with BETTER Pipelines
Preferred option: Geo-referenced perimeters
Field Data
Large Fields along roads:
Geo-referenced Transects
Field Capture
Office
Completion
Huge Fields: Single Points / Corners
5.
6. http://ec-better.eu
BETTER | Enhancing Agricultural Mapping with BETTER Pipelines
Sen2Agri has served its purpose well. However:
• It only uses S2 and L8, but no SAR.
• There is a module for atmospheric correction but significant atmospheric
interference may still have an impact on the quality of the classification
WFP ClEO Team has implemented a flexible,
data-driven, pixel-optimized filter:
• So far routinely applied to medium resolution
data, NDVI and LST – smooth, clean images
without mosaicking artifacts or cloud inflicted
gaps.
• Can we apply the same to Sentinel-2 data?
Good for hazard mapping, but is there a
benefit for crop type (or more generally land
cover) classification
8. http://ec-better.eu
BETTER | Enhancing Agricultural Mapping with BETTER Pipelines
Filter behavior is modulated by two
parameters:
• S – controls the amount of smoothing
that is applied. The smaller the S, the
closer to the original data the output is.
At the extremes the output will be a
polynomial of degree n (usually 1 or 2).
• S can be optimized (CV, V-curve)
9. http://ec-better.eu
BETTER | Enhancing Agricultural Mapping with BETTER Pipelines
Filter behavior is modulated by two parameters:
• p – controls the asymmetry of the filtered
output and is a value between 0 and 1. When
p is 0.5, filtered output balances noise. As p
tends to 1 the filter fits the output to the
higher envelope of the signal. As p tends to
0 the filter fits to the lower envelope of the
signal (baseline) .
• NDVI: use p > 0.5
• Reflectances: use p < 0.5
• p is not optimized, expert judgement
10. http://ec-better.eu
BETTER | Enhancing Agricultural Mapping with BETTER Pipelines
Data Set:
• 8800 time series of raw and smoothed Sentinel-2 reflectances and SAR
• Smoothing: S = optimized btn 3.0-4.0 and p=0.2 (p=0.3 for Band 8, 11 and 12)
• Nature: pixels selected inside each agronomic data sample, 25m distance
• Field data: Crop field outlines, collected early October 2020 (pre-harvest)
• Location: Adamawa State, NE Nigeria
• Dates: Time series from May to October 2020
• Crop Types: Maize, Millet, Sorghum, Beans, Soya and their mixes. Also fallow
(unused) and No-ID (abandoned field)
• 7000 labelled data and 1800 unlabelled
11. http://ec-better.eu
BETTER | Enhancing Agricultural Mapping with BETTER Pipelines
The Exercise:
• Devise a classifier for crop types to be run with the various datasets provided
and results compared
• You are free to choose combinations of data. E.g.
• Raw S2 vs Filtered S2
• Raw S2 + SAR vs Filtered S2
• SAR vs Filtered S2
• … whatever
• You can team up if you want
• Produce your predictions for the unlabeled data and submit
Google Hangouts for Bilateral Support:
https://meet.google.com/mde-useo-qrb