Taimoor Akhtar

Geo-spatiotemporal Analytics

Data apps for hydrology and environmental science

Good quality climate data is available in abundance via multiple platforms and resources (e.g., NASA Earthdata, ESGF, etc.). However, leveraging such data can be a barrier and challenge for environmental scientists. I have created numerous web applications and services that source, interpret and manage climate data for applicable use in hydrology and earth sciences. These applications incorporate data pipelines that periodically access climate data from numerous sources (e.g., see my blog post on accessing NASA’s earthdata).

---

Can-GLWS is a Data as a Service (DaaS) platform I co-developed that allows application users to download SWAT-ready climate data (historical, climate change scenarios and weather statistics) of a pre-specified region within the Canadian Great Lakes watersheds. SWAT is a highly popular tool for hydrologic and watershed modelling.

---

The APWS application provides weather statistics for the Asia Pacific region that can be used in WXGN weather generator, for generating synthetic weather time-series for any location (or multiple locations) in the Asia Pacific region.

---

The technology stack used for developing these applications includes R Shiny, PostgreSQL, Python, Apache Airflow and Docker. All applications are deployed on a dedicated server at the University of Guelph.

Close

Operations Research

Algorithms and Mathematica Models for Calibration, Hyperparameter Tuning and Water Management

I specialize in developing novel and efficient heuristic and surrogate optimization algorithms for applications in model calibration, hyperparameter tuning of Deep Neural Networks and environmental management. In my current and prior optimization projects I have developed and applied numerous such algorithms and also large-scale mathematical models (e.g., Linear and Non-Linear Programming), to environmental and computer science applications. Some of these models have been successfully integrated into production systems.

---

GOMORS is a surrogate-assisted Multi-Objective Optimization (MO) algorithm, designed for computational expensive MO problems, e.g., expensive environmental simulation optimization problems, hyperparameter tuning of Deep Neural Networks etc. It is implemented in the pySOT library (python) and framework, and uses Radial Basis Functions (RBFs), as surrogates.

---

HORD is an efficient optimization method designed for hyperparameter optimization of Deep Learning (DL) algorithm. HORD has been extensively evaluated on DL benchmarks and performs well in comparison to popular Bayesian optimization methods. It is implemented in the pySOT library (python), and applied to DL problems implemented in torch, with tests conducted on GPUs.

---

The technology stack I use for optimization includes numerous python libraries, e.g., pySOT , platypus , pyomo , scikit-learn and scipy etc.

Close

Machine Learning

Applications in earth sciences and spatio-temporal forecasting

Earth science datasets and data acquisition tools are readily available in abundance these days and present novel opportunities for application of Machine Learning (ML) to unearth patterns and derive insights. I apply ML methods to open earth science datasets with application focus in hydrology, watershed management and soil classification. While my primary ML work relates to data exploration and model building, training & testing, I am also intricately involved in the process of pushing these models into production at scale.

---

Streamflow Prediction: I have recently applied different supervised machine learning algorithms for modeling streamflow in Ontario’s (Canada) catchments. Observed streamflows from across Ontario are being used in training and testing a regional streamflow prediction model, where input features include observed weather, land-surface and land-use data (geo-processed). Comparison algorithms include XGBoost, Random Forests, and Recurrent Neural Networks (e.g., LSTM). This work is being conducted in collaboration with Greenland Consulting . The final model is being pushed into production within Greenland’s THREATS platform, for early warning against flood and drought events. Some preliminary findings related to the ML prediction system can be accessed at the following link. This streamflow prediction product has multiple use-cases including early flood & drought warning, water resource assessment in ungaged streams and time-series imputations at gaged locations.

---

Watershed Quality Prediction: Predicting how river quality is affected by agricultural practices and land-use is a major challenge that typically requires development of highly complex watershed simulation models. Working within an interdisciplinary project focused on developing agricultural management plans for reducing phosphorus in Ontario’s watersheds, I am developing a Machine Learning algorithm for predicting river water quality in response to agricultural management. This ML method will act as a surrogate for the highly intricate physical watershed model, SWAT, and thus, is being trained on the output of SWAT.

---

The technology stack I typically use in machine learning includes scikit-learn, xgboost keras, pytorch and pytorch forecasting.

Close

Dashboards

Data Pipelines, DataViz and Decision Support Systems for Water and Energy Management

I develop innovative dashboards and production-ready Decision Support Systems (DSS) to derive insights from environmental data and models. My specialization within dashboard and DSS development is web scraping, data pipelining, DataViz and deployment.

---

GSAT-PK is an on-going collaborative open source project, initiated to develop a DSS for approximating and disseminating snow & glacial storage and river flow information of selected catchments within the transboundary Indus Basin. GSAT-PK relies completely on open-source data (e.g., satellite imagery from NASA’s MODIS dataset) to model latest snow cover and storage status. Moreover, latest river flow data is gathered via web scraping. Apache airflow is used to manage and maintain the entire GSAT-PK data pipeline that includes data processing and modeling scripts developed in python and R. Data visualizations and dashboard are developed in plotly and R shiny, whereas the DSS is dockerized and deployed on a DigitalOcean droplet.

---

CANWET-SWAT is a data visualization tool designed specifically to analyze the spatio-temporal outputs of the SWAT watershed modeling tool. SWAT is a highly complex watershed model with numerous spatio-temporal outputs. CANWET-SWAT uses flask APIs to export these outputs into a PostgreSQL / PostGIS database. DataViz APIs are then used to export outputs from the database into CANWET-SWAT’s data visualization interface (developed in the THREATS system). CANWET-SWAT is developed as a collaboration between University of Guelph and Greenland Consulting, with my work centered on development of the Postgres Database, flask APIs and interactive visualizations.

---

Pakistan’s Groundwater-Energy Nexus Exploration is an interactive data visualization and exploration application developed to understand the impact of electric tubewell pumping on Pakistan’s water-energy nexus. The application is developed using streamlit and plotly, whereas data is sourced from multiple public reports and repositories.

---

The technology stack I use in DSS and dashboard development includes Flask, PostgreSQL, PostGIS, plotly, R Shiny, Apache Airflow, streamlit, Docker and DigitalOcean Droplet.

Close

Parallel Programming

Frameworks to Speed-up Models and Algorithms

I am adept at developing programming frameworks for parallelization of complex workflows and optimization algorithms. I have extensive experience of developing parallel (synchronous and asynchronous) codes in C, python and R, and have worked with numerous parallelization libraries including multiprocessing, mpi4py & POAP in python, and snow & parallel in R.

---

SOP is a surrogate-assisted efficient parallel optimization algorithm, specifically designed for optimization of expensive problems on multi-node clusters. The algorithm achieves good parallel speed-up on up to 64 cores, and is implemented within the pySOT library in python.

---

LCC-SWAT is a cloud-based calibration and uncertainty analysis system, designed for calibration of SWAT models. The calibration system is deployed on a server within University of Guelph (UoG), and can be easily accessed by UoG modelers, via a web-interface, to fast-track their SWAT model development and calibration workflows.

---

The technologies and libraries I use in parallel programming include MPI, mpi4py, POAP , snow (R) and parallel (R).

Close

Statistical Analysis

Modeling and Analysis of Extreme Events and Time-Series Data

My core suite of expertise also includes statistical modeling and analysis, with application focus in time-series analysis and extreme events. My prior experience in statistical modeling and analysis includes (but is not limited to) Extreme Value Analysis (EVA) in hydrology, time-series imputation (for sensor data), statistical drought analysis, statistical downscaling and time-series trend analysis.

---

Drought Early Warning: This work, done in collaboration with the Global Change Impact Studies Centre (GCISC) Pakistan, devised a statistical framework for early prediction and monitoring of droughts within Pakistan. Using cross-correlations between standardized statistical indices of hydrologic and meteorologic patterns, numerous insights were derived for early warning against droughts in rivers within Pakistan.

---

Hazard Indexes in Hydrology: This work is part of the Streamflow Prediction product that I am working on for Greenland group. In this task I am using statistical analysis and dataviz to develop and visualize hazard indices for identifying floods and droughts, using predicted streamflow data (from Machine Learning models). Using Extreme Value Analysis (EVA) and flood frequency analysis, we are developing statistical indices for quantifying (with uncertainty) and qualifying magnitudes of flood and drought events from predicted streamflow data. The developed indices and related visualizations are being integrated into Greenland’s THREATS system. A prototype dataviz that includes EVA for defining extrene event thresholds is shown below.

---

Time-Series Modelling for Data Imputation: This is an exploratory project for analyzing different classical statistical models and their capabilities in time-series imputations using correlated streamflow sensor network data. I used numerous statistics models including MICE, for this analysis and results are summarized in the link below.

---

The technology stack I use for statistical modeling and analysis includes numerous python libraries, e.g., scikit-learn, scikit-extremes and statsmodels etc.

Close