Monsoon Shift Detection & Rainfall Regime Modeling
Thirty-one years of daily rainfall in Birendranagar, analysed with classical trend tests and an LSTM model — a hybrid statistical and deep-learning study of a Himalayan mid-hill town.
The question
Has the monsoon over Birendranagar, Nepal changed — and can a deep-learning model capture its day-to-day rainfall regime? I investigated 31 years of daily precipitation (1993–2023; 11,322 records) from the Surkhet Airport meteorological station.
Method
- Quality control, gap handling and temporal aggregation to build an analysis-ready precipitation dataset
- Mann–Kendall trend testing with Theil–Sen slope estimation
- Pettitt change-point detection on annual and monsoon-season totals
- A two-layer LSTM using 30-day rainfall windows, with chronological train / validation / test splits to avoid temporal leakage
Findings
- No statistically significant trend in annual rainfall (τ = 0.028, p = 0.838; +1.15 mm/yr) or monsoon rainfall (τ = −0.024, p = 0.865; −1.05 mm/yr)
- No significant abrupt change point in annual (p = 1.000) or monsoon-season rainfall (p = 0.899)
- The LSTM achieved RMSE 11.78 mm and MAE 4.97 mm on a 2,259-day held-out test period — a 22.8% RMSE reduction over a persistence baseline
Publication
Hybrid Deep Learning and Statistical Approach for Monsoon Rainfall Regime Modeling in a Himalayan Mid-Hill Town (1993–2023): A Case Study of Birendranagar, Nepal — submitted to Theoretical and Applied Climatology (Springer Nature).