ÍâÁ÷ӰƬ

SOIL MOISTURE FORECASTING WITH GENETICALLY OPTIMIZED DEPTH-AWARE LSTM

Master of Science (Computer Science) candidate: Md Ahad Hasan
1 April 2026

3:00 PM Atlantic

Join:
Meeting ID: 251 081 160 649 60
Passcode: js6ti6PL

Thesis Committee:
Drs. Sazia Mahfuz & Nasimul Noman, Supervisors
Dr. Saeid Homayouni, Eau Terre Environment Research Centre, National Scientific Research Institute (INRS), External Examiner
Dr. Amir Eaman, Internal Examiner
Dr. Allison Walker, Chair of the defence

Abstract

Accurate multistep soil moisture forecasting across the vertical profile is important for irrigation planning, drought monitoring, and hydrological decision support. In this thesis, we formulate a multi-depth, sequence-to-sequence forecasting problem using RISMA station data from Saskatchewan, Canada. We use hourly observations from 2014 to 2021 and predict soil moisture up to 7 days ahead for three aggregated layers: Shallow (0–5 cm, 5 cm), Mid (20 cm, 50 cm), and Deep (100 cm, 150 cm). We propose DALSTM, a depth-aware architecture that combines a Triple-stream BiLSTM encoder, Cross-Depth Attention, and an encoder-decoder design conditioned on future meteorological drivers. To improve long-horizon stability, we train with Curriculum learning (1-day to 3-day to 7-day progression), apply depth-weighted optimization, and include a lightweight physics-informed regularizer. We further apply multi-objective NSGA-II optimization to jointly improve predictive skill and model efficiency. Experimental results show that DALSTM achieves strong 7-day performance on the primary SK4 site (RMSE = 0.0122, R2 = 0.8730). At the same horizon, depth-wise performance reaches R2 = 0.8301 (Shallow), 0.9219 (Mid), and 0.8810 (Deep). Compared with reproduced unified multi-depth studies, DALSTM shows about 49–53% reduction in 7-day RMSE and retains 92.83% of its 1-day R2 at 7 days. NSGA-II also reduces DALSTM parameters from 836,387 to 83,991 while improving accuracy. For cross-site generalization, transfer learning from SK4 to SK2 improves 7-day overall R2 from 0.8122 to 0.8666 with a 15.0% RMSE reduction. These findings show that depth-aware representation learning, horizon-aware training, and compact multi-objective optimization can produce accurate, stable, and transferable soil moisture forecasts for cold-region agricultural systems.

About Md Ahad …

I am a master’s student in Computer Science at ÍâÁ÷ӰƬ, with a focus on machine learning and environmental data science. My research explores how artificial intelligence can improve soil moisture forecasting by combining multi-depth sensor data with deep learning models. Through this work, I developed a unified depth-aware LSTM (DALSTM) architecture designed to better forecast soil moisture across different depth layers over time.
Before graduate school, I built nearly a decade of experience as a software engineer, working on backend systems and data-driven applications. My goal is to contribute to more reliable forecasting tools that support sustainable agriculture and smarter environmental decision-making.