Open Access Open Access  Restricted Access Subscription Access

doi:10.3808/jei.202000427
Copyright © 2017 ISEIS. All rights reserved

Using Satellite Remote Sensing and Machine Learning Techniques Towards Precipitation Prediction and Vegetation Classification

D. Stampoulis1 *, H. G. Damavandi1, D. Boscovic2, and J. Sabo1

  1. Future H2O, Office of Knowledge Enterprise Development, Arizona State University, Tempe, AZ 85281, USA
  2. Center for Assured and Scalable Data Engineering, Arizona State University, Tempe, AZ 85281, USA

*Corresponding author. Tel.: +1 480 9653312. E-mail address: dstampou@asu.edu (D. Stampoulis).

Abstract


The spatial distribution, magnitude and timing of precipitation events are being altered globally, often leading to extreme hydrologic conditions with serious implications to ecosystem services, water, food and energy security, as well as the welfare of billions of people. Motivated by the pressing need to understand, from a hydro-ecological perspective, how the dynamic nature of the hydrologic cycle will impact the environment in water-stressed regions, we implemented a novel approach that predicts precipitation spatio-temporal trends over the drought-burdened region of East Africa, based on other major hydrological components, such as vegetation water content (VWC), soil moisture (SM) and surface temperature (ST). The spatial patterns and characteristics of the inter-relations among the four aforementioned hydrologic variables were investigated over regions of East Africa characterized by different vegetation types and for various precipitation intensity rates during 2003-2011. To this end, we analyzed multi-year satellite microwave remote sensing observations of SM, ST, and VWC (derived from Naval Research Laboratory's WindSat radiometer) as well as their response to precipitation patterns (derived from NASA's TRMM 3B42 V7). We categorized precipitation into four bins (ranges) of intensity and trained five different state-of-the-art machine learning models for each of these categories. The models were then applied to predict the spatiotemporal precipitation dynamics over this complex region. Specifically, the Random Forest and Linear Regression models outperformed the others with the normalized mean absolute error being less than 27% for all of the categories. The characteristics of the predicted precipitation were in turn used to classify vegetation regimes in East Africa. Our results indicate significant discrepancies in the performance of the models with varying values in the predicting skill as well as their ability to accurately classify vegetation into different types. Our predictive models were able to forecast the three vegetation regimes, i.e., Forest/Woody Savanna, Savanna/Grasslands and Shrubland, with precision rate of at least 81% for all of the aforementioned precipitation bins.

Keywords: machine learning, linear regression, passive microwave remote sensing, precipitation, random forest, soil moisture, surface temperature, vegetation water conten


Full Text:

PDF

Supplementary Files:

Refbacks

  • There are currently no refbacks.