Exploring DeepSense Neural Network Architecture for Farming Events Detection
Name
Kyrylo Medianovskyi
Abstract
Nowadays satellite imagery became widely available and found to be applicable in a range of different areas. Agriculture is one of those domains. With the help of imagery data there is a set of processes that can be automatized. Thousands of people across the European Union are involved in field inspection. They are checking the crop types and take a record of mowing events that happen on the parcels. Estonia has a relatively high level of cloud coverage and rains during a vegetation season. That leads to interruptions and noises in satellite imagery data. A noise tolerating automated mowing event detection system is required.
For this thesis Sentinel-1 coherence for VV and VH polarisation together with Sentinel-2 normalized difference vegetation index were chosen as the main features to build a mowing event recognition system. The architecture DeepSense is implemented and evaluated as a mowing event detection mechanism.
The system was trained on Estonia 2018 labeled data containing information about over 1700 fields. An optimal configuration of hyper-parameters was obtained based on experiments with the architecture. Proposed modification of the DeepSense framework allowed to reach 94% event accuracy and 93\\% end of season event accuracy obtained from 5-fold cross-validation.
The DeepSense implementation allowed to outperform a purely convolutional model based on the end of season accuracy metric (93% against 90%).
The proposed architecture can be adopted for the mowing event detection tasks.
For this thesis Sentinel-1 coherence for VV and VH polarisation together with Sentinel-2 normalized difference vegetation index were chosen as the main features to build a mowing event recognition system. The architecture DeepSense is implemented and evaluated as a mowing event detection mechanism.
The system was trained on Estonia 2018 labeled data containing information about over 1700 fields. An optimal configuration of hyper-parameters was obtained based on experiments with the architecture. Proposed modification of the DeepSense framework allowed to reach 94% event accuracy and 93\\% end of season event accuracy obtained from 5-fold cross-validation.
The DeepSense implementation allowed to outperform a purely convolutional model based on the end of season accuracy metric (93% against 90%).
The proposed architecture can be adopted for the mowing event detection tasks.
Graduation Thesis language
English
Graduation Thesis type
Master - Computer Science
Supervisor(s)
Amnir Hadachi
Defence year
2020