Advertisement
Advertisement

More Related Content

Similar to Method for automated classification with INSPIRE data and Sentinel-2 satellite imagery: case remote crop monitoring(20)

Advertisement

Method for automated classification with INSPIRE data and Sentinel-2 satellite imagery: case remote crop monitoring

  1. 1 Method for automated classification with INSPIRE data and Sentinel-2 satellite imagery: case remote crop monitoring www.spatineo.com Joona Laine / Spatineo INSPIRE CONFERENCE 2018 Antwerp
  2. Background and Introduction ● EU is supporting farmers with Common Agricultural Policy (CAP), including direct subsidy payments for farmers ● EU member countries have to control the validity of CAP subsidy applications ● Crop type identification is one of the tasks ● According to EC1 , Crop monitoring could be carried out using remote sensing imagery, such as Sentinel-2 ● The overall accuracy (OA) should be >95%2 1. European Comission (2014). COMMISSION IMPLEMENTING REGULATION (EU) No 809/2014 2. JRC TECHNICAL REPORTS: 1st draft of the technical guidance on the decision to go for substitution of OTSC by monitoring
  3. Objective The main objectives of our work: ● to generate method for automated classification with various kind of vector or raster spatial data ● to investigate whether it was possible to reliably identify the crop growing in land parcels by using machine learning methods and Sentinel-2 satellite imagery in Finland
  4. Datasets ● Sentinel-2 products ○ For preprocessing: L1C images and cloud mask, L2A snow mask ○ For classification: L2A images for 10m and 20m bands ○ All available L2A products during the thermal growing season of from the area of whole Finland covering land parcels ● Agricultural land parcels obtained from Finnish Agency for Rural Affairs (INSPIRE land cover) ○ Land parcels from CAP subsidy applications 2017 and 2018 ○ Supervised land parcels, ~5% of the CAP application parcels ○ Formed 10 crop type classes according to suggestion of Finnish Agency for Rural Affairs
  5. Challenges: weather ● Challenging weather conditions for optical sensors ● In Finland partly cloudy images have to be used as well
  6. Challenges: class imbalance ● Class distribution highly imbalanced ● Two dominating classes ● Possible solutions: ○ Resampling ○ Model class weighting
  7. Challenges: class imbalance ● Class distribution highly imbalanced ● Two dominating classes ● Possible solutions: ○ Resampling ○ Model class weighting
  8. Crop monitoring: overview
  9. Masks ● 4 masks used: ○ Sentinel-2 cloud mask ○ Generated cloud mask3 ○ Sentinel-2 snow mask ○ Generated cloud shadow mask4 ● Masks filter out non-clear pixels from the images 3. S2cloudless algorithm https://medium.com/sentinel-hub/improving-cloud-detection-with-machine-learning-c09dc5d7cf13 4. Algorithm presented at https://github.com/samsammurphy/cloud-masking-sentinel2/blob/master/cloud-masking-sentinel2.ipynb Basemap by National Land Survey of Finland
  10. Extracting and Preprocessing ● Calculating the bandwise statistical features of the parcels from each available image during set time period ● Temporal interpolation of the extracted values ● Filtering out parcels with insufficient data ● Selection and calculation of the variables that produce highest accuracy => Make the data usable for machine learning algorithms
  11. Classification algorithm selection ● Multiple different ML algorithms tested ● MLP and SVM produced some of the best results ● New methods tested and constantly Multilayer Perceptron (MLP)5 Support Vector Machines (SVM)6 5. Gardner, M.W and S.R Dorling (1998). “Artifcial neural networks (the multilayer perceptron)—a review of applications in the atmospheric sciences” 6. Mountrakis, Giorgos, Jungho Im, and Caesar Ogole (2011). “Support vector machines in remote sensing: A review”
  12. Results for 2017 ● OA: 89%, K: 0.80 ● Supervised parcels used for training ● Time period: 1st of May to 1st of September 1:Broad bean, 2:Pea, 3:Beet, 4:Fallow, 5:Spring rapeseed, 6:Spring cereal, 7:Grass, 8:Potato, 9:Turnip rape, 10:Winter cereal
  13. Results for 2018 ● OA: 93%, K: 0.88 ● Two models: ○ One for labels 5 and 6 ○ One for other labels ● Trained and evaluated with CAP subsidy application parcels using 8-fold cross validation ● Time period: 1st of May to 1st of August
  14. Results for 2018: calibrated ● OA: 95%, K: 0.90 ● Models calibrated with conf. level 0.807 ● 97.9% (913126 out of 932920) parcels classified after calibration 7. Schmedtmann, J. and M. L. Campagnolo (2015). “Reliable crop identification with satellite imagery in the context of Common Agriculture Policy subsidy control”
  15. Conclusions ● Developed method works even under challenging conditions ● Imbalance of class distribution is major problem but it can be solved ● Crop classes should be grouped based on biological and phenological similarities if possible => policy needs to take this into account ● No method will be perfect ○ Not all parcels can be classified ○ Timely results are required to allow farmers to react to false negatives ○ No ground truth available
  16. Future directions ● Modifying the workflow to further meet the EC technical guidance suggestions8 ● Classifying with more ML algorithms ● Chaining multiple different ML algorithms ● Classifying with different crop class division and class formation ● Using other remote sensing sources, such as Sentinel-1 ● Utilizing the method for other applications with INSPIRE data 8. JRC TECHNICAL REPORTS: 1st draft of the technical guidance on the decision to go for substitution of OTSC by monitoring
  17. 17www.spatineo.com Thank you for your attention! .. and thanks to
Advertisement