Farming Events Detection from Sentinel-1 and -2 Satellite Imagery Time Series with Deep Learning
Name
Viacheslav Komisarenko
Abstract
Satellite imagery allows building applications in a variety of domains. Agriculture is an example with a lot of possibilities for automation. Thousands of inspectors visit fields across the European Union to check if mowing events were performed. Reliable automated detection system can free this work force for other needs.
Sentinel-1 (coherence in VH and VV polarization) and -2 (normalized difference
vegetation index) data was chosen as the base feature set in this thesis. Convolutional neural network model was created which is capable to detect mowing events based on satellite’s imagery time series. Using Estonia 2018 labeled data about 2000 fields, the model was trained and optimal configuration of hyperparameters was tuned. Transfer learning techniques were applied based on Swedish 2018, Danish 2018 and Estonian 2017 data. Weights of trained models were re-used to improve performance on target Estonian 2018 dataset. Based on the reject region method, an algorithm for finding a subset with highly confident and accurate predictions was proposed.
Proposed modifications allowed to obtain event accuracy of 76.1% and end of season accuracy of 96.6%. The proposed model architecture is suitable for practical use in the mowing detection system.
Sentinel-1 (coherence in VH and VV polarization) and -2 (normalized difference
vegetation index) data was chosen as the base feature set in this thesis. Convolutional neural network model was created which is capable to detect mowing events based on satellite’s imagery time series. Using Estonia 2018 labeled data about 2000 fields, the model was trained and optimal configuration of hyperparameters was tuned. Transfer learning techniques were applied based on Swedish 2018, Danish 2018 and Estonian 2017 data. Weights of trained models were re-used to improve performance on target Estonian 2018 dataset. Based on the reject region method, an algorithm for finding a subset with highly confident and accurate predictions was proposed.
Proposed modifications allowed to obtain event accuracy of 76.1% and end of season accuracy of 96.6%. The proposed model architecture is suitable for practical use in the mowing detection system.
Graduation Thesis language
English
Graduation Thesis type
Master - Computer Science
Supervisor(s)
Sherif Sakr, Kaupo Voormansik, Yousef Essam
Defence year
2019