Browsing by Author "Jia, Xiaowei"

Now showing 1 - 5 of 5
  • Thumbnail Image
    Item
    Automated Plantation Mapping in Southeast Asia Using Remote Sensing Data
    (2016-08-16) Jia, Xiaowei; Khandelwal, Ankush; Gerber, James; Carlson, Kimberly; West, Paul; Samberg, Leah; Kumar, Vipin
    Plantation mapping is critical for understanding and addressing deforestation, a key driver of climate change and ecosystem degradation. Unfortunately, most plantation maps are limited to small areas for specific years because they rely on visual inspection of imagery. Here we propose an automated approach for annual mapping of plantations using remote sensing data. Due to the heterogeneity of land cover classes, we propose a novel ensemble learning method that simultaneously uses training samples from multiple land cover classes over different years. After the ensemble learning, we further improve the performance by post-processing using a Hidden Markov Model. With the experiments on MODIS data, we demonstrate the superiority of the proposed method over multiple baselines. In addition, we conduct extensive validation by comparing the detected plantation by our approach with the existing datasets developed through visual interpretation by expert observers. Based on the random sampling and the comparison with high-resolution images, the precision (i.e. user’s accuracy) and recall (i.e. producer’s accuracy) of our generated map are around 85.53% and 81.51%, respectively, and the overall accuracy is 95.20%.
  • Thumbnail Image
    Item
    Classifying multivariate time series by learning sequence-level discriminative patterns
    (2018-01-23) Nayak, Guruprasad; Mithal, Varun; Jia, Xiaowei; Kumar, Vipin
    Time series classification algorithms designed to use local context do not work on landcover classification problems where the instances of the two classes may often exhibit similar feature values due to the large natural variations in other land covers across the year and unrelated phenomena that they undergo. In this paper, we propose to learn discriminative patterns from the entire length of the time series, and use them as predictive features to identify the class of interest. We propose a novel neural network algorithm to learn the key signature of the class of interest as a function of the feature values together with the discriminative pattern made from that signature through the entire time series in a joint framework. We demonstrate the utility of this technique on the landcover classification application of burned area mapping that is of considerable societal importance.
  • Thumbnail Image
    Item
    Integrating Physics into Machine Learning for Monitoring Scientific Systems
    (2020-07) Jia, Xiaowei
    Machine learning (ML) has transformed all aspects of our life including how we make decisions, entertain ourselves, and interact with each other. The power of ML models lies in their ability to automatically extract useful patterns from complex data. Given the well-known success of ML in commercial domains, there is an increasing interest in using ML models for advancing scientific discovery. However, direct application of “black-box” ML models has met with limited success in scientific domains given that the data available for many scientific problems is far smaller than what is needed to effectively train advanced ML models. Moreover, in the absence of adequate information about the physical mechanisms of real-world processes, ML approaches are prone to false discoveries of patterns which look deceptively good on training data but cannot generalize to unseen scenarios. This thesis introduces a new generation of machine learning approaches which leverage accumulated scientific knowledge to solve problems of great scientific and societal relevance. We investigate multiple ways in which physical knowledge can be used in the design of ML models for effectively capturing underlying physical processes that are evolving and interacting and multiple scales. We also introduce new optimization strategies for ML models so that they can achieve higher accuracy with limited data and also preserve the correctness from a physical perspective. We will describe our technical innovations and show how they help address real-world challenges by focusing on applications from two disciplines: aquatic science and monitoring crops at scale. We first introduce a physics-guided machine learning framework, which explores a deep coupling of ML methods with scientific knowledge. We show this approach can significantly outperform the state-of-the-art physics-based models and machine learning models in monitoring lake systems and river networks using limited training data while also maintaining consistency to known physical laws. Also, we greatly advance existing deep learning methods so that they can learn patterns from real-world data of greater complexity. These techniques have shown a lot of success in detecting primary crops in US and tree crop plantations in Southeast Asia.
  • Thumbnail Image
    Item
    Physics Guided RNNs for Modeling Dynamical Systems: A Case Study in Simulating Lake Temperature Profiles
    (2019-01-31) Jia, Xiaowei; Willard, Jared; Karpatne, Anuj; Read, Jordan; Zwart, Jacob; Steinbach, Michael; Kumar, Vipin
    This paper proposes a physics-guided recurrent neural network model (PGRNN) that combines RNNs and physics-based models to leverage their complementary strengths and improve the modeling of physical processes. Specifically, we show that a PGRNN can improve prediction accuracy over that of physical models, while generating outputs consistent with physical laws, and achieving good generalizability. Standard RNNs, even when producing superior prediction accuracy, often produce physically inconsistent results and lack generalizability.  We further enhance this approach by using a pre-training method that leverages the simulated data from a physics-based model to address the scarcity of observed data. Although we present and evaluate this methodology in the context of modeling the dynamics of temperature in lakes, it is applicable more widely to a range of scientific and engineering disciplines where mechanistic (also known as process-based) models are used, e.g., power engineering, climate science, materials science, computational chemistry, and biomedicine.
  • Thumbnail Image
    Item
    Source Aware Modulation for leveraging limited data from heterogeneous sources
    (2021) Li, Xiang; Khandelwal, Ankush; Ghosh, Rahul; Renganathan, Arvind; Willard, Jared; Xu, Shaoming; Jia, Xiaowei; Shu, Lele; Teng, Victor; Steinbach, Michael; Nieber, John; Duffy, Christopher; Kumar, Vipin
    In many personalized prediction applications, sharing information between entities/tasks/sources is critical to address data scarcity. Furthermore, inherent characteristics of sources distinguish relationships between input drivers and response variables across entities. For example, for the same amount of rainfall (input driver), two different basins will have very different streamflow (response variable) values depending on the basin characteristics (e.g., soil porosity, slope, …). Given such heterogeneity, a trivial merging of data without source characteristics would lead to poor personalized predictions. In recent years, meta-learning has become a very popular framework to learn generalized global models that can be easily adapted (fine-tuned) for individual sources. In this talk, we present an exhaustive analysis of the source-aware modulation based meta-learning approach. Source-aware modulation adjusts the shared hidden features based on source characteristics. The adjusted hidden features are then used to calculate the response variable for individual sources. Although this strategy shows promising prediction improvement, its applicability is limited in certain applications where source characteristics might not be available (especially due to privacy concerns). In this work, we show that robust personalized predictions can be achieved even in the absence of explicit source characteristics. We investigated the performance of different modulation strategies under various data sparsity settings on two datasets. We demonstrate that source-aware modulation is a very viable solution (with or without known characteristics) compared to traditional meta-learning methods such as model agnostic meta-learning.