Drought LSTM
A proof-of-concept LSTM that predicts next week’s drought index from eight weeks of weather.
Won’tNever trains on the future.
- Role
- Undergraduate research project, 3rd-year B.Tech
- When
- Jan 2026
- Built with
- PythonTensorFlow / KerasLSTMpandasscikit-learnJupyterUS Drought Monitor data
- Links
Illustration of the forecast shape, not the model’s actual output.
Why it exists
Drought early-warning systems need a signal before conditions turn bad. This project tests whether a small LSTM can predict next week’s drought level from recent weather.
What it does
- A single-layer LSTM (32 units) with a dense output: 4,641 parameters.
- Input: 8 weeks of precipitation, temperature and humidity. Output: a 0 to 1 drought index for week 9.
- A fully documented Jupyter notebook.
What it won’t do
- It never trains on the future: a time-aware 70/15/15 split replaces a random one.
- It doesn’t pretend to be production: the model stays small and readable, as an academic demonstration.
How it works
Sliding 8-week windows over US Drought Monitor data (about 1.1 GB) feed the LSTM. It trains on MAE with Adam (learning rate 0.001), batch size 16, for 50 epochs.
By the numbers
87%+Prediction accuracy
10K+Climate data points
4,641Parameters
8 weeksLookback window
Next build notesInsightAI
Answers plain-English questions about an e-commerce database with safe, read-only MongoDB queries.