Source: Kaggle
Avoiding condensation trails, clouds of ice crystals that form in the exhaust of aircraft engines, is an effective sustainability strategy for airlines. Persistent contrails contribute to global warming by trapping heat in the atmosphere. Scientists have developed models to predict the formation and impact of condensation trails and are validating their predictions with satellite images.
The goal of this contest was to improve corresponding prediction models by verifying the accuracy of those models using satellite imagery. Therefore, this was a semantic segmentation task.
From the segmentation-models-pytorch library, we considered the following nine base models:
- Unet: A popular model for semantic segmentation tasks, consisting of a contract phase and an extension phase to capture context and localization.
- Unet++: An extended version of Unet that adds nesting and dense connections for improved segmentation performance.
- MAnet: A model based on multiple layers of attention to better highlight different features of the input data.
- Linknet: A lightweight, fast, and accurate model for recovering high-resolution features from low-resolution features.
- Feature Pyramid Network (FPN): A model that generates a pyramid of fast features with different scales for object detection and segmentation.
- Pyramid Scene Parsing Network (PSPNet): A network that aims to accurately segment complex scenes by using a pyramid pooling layer to aggregate contextual information.
- Path Aggregation Network (PAN): A model that aims to improve the flow of information in the network by connecting lower level features to higher levels to achieve better segmentation results.
- DeepLabV3: An improved version of the DeepLab model that uses dilated convolution and a new module structure (Atrous Spatial Pyramid Pooling) to achieve segmentations with high accuracy.
- DeepLabV3+: A further improvement of DeepLab that introduces an encoder-decoder structure to achieve more precise segmentation, especially at object edges.
There are basically four ways to combine these models::
- Majority Voting: In this approach, the class with the most "votes" is chosen for each pixel. If more than half of the models say that a pixel is a contrail, then it is a contrail, otherwise it is not. This approach is easy to implement and can work well if the models are relatively independent.
- Average Prediction: This approach averages the output of the models for each pixel. This can be useful when the models have similar outputs and you want to avoid overpredictions from a single model.
- Weighted Average Prediction: Similar to Average Predictions, but each model is assigned a weight corresponding to its performance. The better performing models are given more weight.
- Generalized Stacking: Another model is trained that learns to combine the predictions of the individual models. This can be more complex to implement, but can also lead to better results if done well.