Corrective Labeling
Once the crop mask for the region of interest meets the metric condition [F-1 score> 0.7], it qualifies as a good map. However, if the metric falls below 0.7, then more data must be added to increase the training samples hence improving the model's performance. This process of choosing informative samples which a user subsequently annotates and incorporates into the training set is known as Corrective labeling.
This technique, in contrast to active learning, necessitates human intervention to opportunistically spot misclassified locations on the predicted map. With this strategy, the labelers are free to make use of their extensive domain expertise to determine which points need to be corrected. The labeled data are fed into the data pipeline to retrain the model.
Prerequisite
- Crop mask
The Corrective Labeling App was developed in Google Earth Engine to facilitate the collection of additional labels. Labelers have access to a synchronized view of the crop mask and the high-resolution satellite imagery which enhances the understanding of the landscape and physical features on the ground.
From the image above, the green color on the cropland map represents croplands predicted by the model and the yellow color shows non-croplands. The image on the right is a Sentinel-2 true-color composite.
In the image above, the crop mask on the left shows a road that was predicted as cropland. The labeler identified this and dropped a marker on it.
The image reveals a built-up area on the Sentinel-2 image that was predicted as a crop in the crop mask.
This non-crop area as highlighted in the Sentinel-2 image was predicted as cropland in the crop mask
Link to the App with detailed instructions: