The major inspiration behind this project was to develop a faster and simpler model for generating features, connecting inputs and visualizing the elements of a temperature profiled dataset. This project looks at the use of traditional, ensemble and GAN based methods for the prediction of temperature profiles by targeting other variables.
Real life implications of this project will help in the interpretation of climatic effects and the consequences of unabated carbon emissions on natural environments by the reproduction of GAN based images which take inputs from precise models.
What it does
The code base is split up into a series of notebooks which concern the creation, testing and modelling of different algorithms for the prediction of temperature profiles based on predetermined and measured topographic and meteorological data. The notebooks can be executed to generate the models and process the images to create powerful results from the GANs. It can also serve as a stepping stone to build automated applications for image pre-processing and generating meaningful predicted images that capture the effects of inclement climate change on regions. Link to repository:- https://github.com/AmDeep/AI_LaunchLab_Team3
How we built it
The project was using a bundle of softwares and packages including Jupyter, Google Collab, AWS, GANs, keras and tensorflow. The models built are also designed to address the adaptation approach which can provide precise measures for mitigating regions that are likely to be most affected in the future as a result of climate change.
Challenges we ran into
It was challenging to rescale and reorient the data in a format that would make the modelling less cumbersome while also generating meaningful insights from the data analysis and exploratory analysis. Connecting the information for timestamps and dimensions to the various regions of the images and applying the best algorithms was also a challenge.
Accomplishments that we're proud of
We have a list of regression models, LSTMs and GANs that perform amazingly and offer a wide choice to users to select the best models. We were also able to generate GAN results from the images as well as a list of the feature importances for different inputs as a way of determining which models are best. This has also provided better context for which models are the fastest and simpler than their counterparts and therefore, more appropriate for deployment.
What we learned
Certain parameters and features are far more important and can offer great means of building adaptable models for tracking and reviewing the consequences of climate crisis on regions. This can be accomplished from a mix of GANs and ensemble methods which perform exceptionally with the right amount of fine tuning. There is a greater emphasis on changing parameters annually and how they interplay with each other which does give rise to anomalies. The accuracies for the testing performances(some of the models) have been listed below:- LSTM score:- [0.000696140865329653, 0.000696140865329653, 0.018852591514587402] Depth tree scores:- 1 -9.780439412970836 2 -5.963767675701363 3 -4.774158928268536 4 -4.442155006197374 5 -4.094581830923138 6 -3.7695465469683826 7 -3.7735580122522108 8 -3.776314958861915 9 -3.7690939010429303
ElasticNetCV(alphas=None, copy_X=True, cv=10, eps=0.001, fit_intercept=True, l1_ratio=0.5, max_iter=1000, n_alphas=100, n_jobs=None, normalize=False, positive=False, precompute='auto', random_state=0, selection='cyclic', tol=0.0001, verbose=0) 0.8449874104691709
XGBOOST RMSE: 1.264600
What's next for Super-resolution using GANs
We will be aiming for a release of the application for researchers and startups to use for future modelling purposes.