Methods
Summary
The methodology is also described at the peer reviewed paper.
...
The first step will be to collect images from various case studies and from various sources globally. It is an important step in order to ensure diversity.
Some artificial flood images will also be created and classified separately.
Then, the classification schema followed in v1.0 will be updated.
Image classification will be performed.
Sequentially, several thousands of images will be imported into Label Studio. The masks will be digitized and classified as described in the paper and according to the updated classification schema. AI models that can help in masking will be evaluated and used.
Challenges
The project is related to expanding in terms of volume and creating a more detailed classification schema. The process is already defined and presented in the scientific paper, so there is experience that can speed up the tasks. The process is time-consuming. In general, if AI models can help in masking within the Label Studio environment, that will save more time. Otherwise, the process will continue to be time-consuming.
It can be empirically assumed that the models will partially help due to the nature of the data. In that case, models trained in Deucalion and Pyrrha v1.0 could also be useful.
In any case, many checks will be needed to ensure quality.
Pre Analysis Plan
The dataset will be used to finetune various state-of-the-art models. The generated metrics can help to set more effective parameters, pay attention to specific classes, etc. The fine-tuned models will be further tested on other data to ensure generalization.
Protocols
Browse the protocols that are part of the experimental methods.
