arXiv subjects

Zhihan Gao

Publications and source records attributed to Zhihan Gao.

15 recordsLinked to original sources

arXiv2025

STLDM: Spatio-Temporal Latent Diffusion Model for Precipitation Nowcasting

Precipitation nowcasting is a critical spatio-temporal prediction task for society to prevent severe damage owing to extreme weather events. Despite the advances in this field, the complex and stochastic nature of this task still poses challenges to existing approaches. Specifically, deterministic models tend to produce blurry predictions while generative models often struggle with poor accuracy. In this paper, we present a simple yet effective model architecture termed STLDM, a diffusion-based model that learns the latent representation from end to end alongside both the Variational Autoencoder and the conditioning network. STLDM decomposes this task into two stages: a deterministic forecasting stage handled by the conditioning network, and an enhancement stage performed by the latent diffusion model. Experimental results on multiple radar datasets demonstrate that STLDM achieves superior performance compared to the state of the art, while also improving inference efficiency. The code is available in https://github.com/sqfoo/stldm_official.

cs.LG↗

arXiv2024

Multi-modal learning for geospatial vegetation forecasting

The innovative application of precise geospatial vegetation forecasting holds immense potential across diverse sectors, including agriculture, forestry, humanitarian aid, and carbon accounting. To leverage the vast availability of satellite imagery for this task, various works have applied deep neural networks for predicting multispectral images in photorealistic quality. However, the important area of vegetation dynamics has not been thoroughly explored. Our study breaks new ground by introducing GreenEarthNet, the first dataset specifically designed for high-resolution vegetation forecasting, and Contextformer, a novel deep learning approach for predicting vegetation greenness from Sentinel 2 satellite images with fine resolution across Europe. Our multi-modal transformer model Contextformer leverages spatial context through a vision backbone and predicts the temporal dynamics on local context patches incorporating meteorological time series in a parameter-efficient manner. The GreenEarthNet dataset features a learned cloud mask and an appropriate evaluation scheme for vegetation modeling. It also maintains compatibility with the existing satellite imagery forecasting dataset EarthNet2021, enabling cross-dataset model comparisons. Our extensive qualitative and quantitative analyses reveal that our methods outperform a broad range of baseline techniques. This includes surpassing previous state-of-the-art models on EarthNet2021, as well as adapted models from time series forecasting and video prediction. To the best of our knowledge, this work presents the first models for continental-scale vegetation modeling at fine resolution able to capture anomalies beyond the seasonal cycle, thereby paving the way for predicting vegetation health and behaviour in response to climate variability and extremes.

cs.CV↗

arXiv2023

PreDiff: Precipitation Nowcasting with Latent Diffusion Models

Earth system forecasting has traditionally relied on complex physical models that are computationally expensive and require significant domain expertise. In the past decade, the unprecedented increase in spatiotemporal Earth observation data has enabled data-driven forecasting models using deep learning techniques. These models have shown promise for diverse Earth system forecasting tasks but either struggle with handling uncertainty or neglect domain-specific prior knowledge, resulting in averaging possible futures to blurred forecasts or generating physically implausible predictions. To address these limitations, we propose a two-stage pipeline for probabilistic spatiotemporal forecasting: 1) We develop PreDiff, a conditional latent diffusion model capable of probabilistic forecasts. 2) We incorporate an explicit knowledge alignment mechanism to align forecasts with domain-specific physical constraints. This is achieved by estimating the deviation from imposed constraints at each denoising step and adjusting the transition distribution accordingly. We conduct empirical studies on two datasets: N-body MNIST, a synthetic dataset with chaotic behavior, and SEVIR, a real-world precipitation nowcasting dataset. Specifically, we impose the law of conservation of energy in N-body MNIST and anticipated precipitation intensity in SEVIR. Experiments demonstrate the effectiveness of PreDiff in handling uncertainty, incorporating domain-specific prior knowledge, and generating forecasts that exhibit high operational utility.

cs.LG↗

arXiv2023

Earthformer: Exploring Space-Time Transformers for Earth System Forecasting

Conventionally, Earth system (e.g., weather and climate) forecasting relies on numerical simulation with complex physical models and are hence both expensive in computation and demanding on domain expertise. With the explosive growth of the spatiotemporal Earth observation data in the past decade, data-driven models that apply Deep Learning (DL) are demonstrating impressive potential for various Earth system forecasting tasks. The Transformer as an emerging DL architecture, despite its broad success in other domains, has limited adoption in this area. In this paper, we propose Earthformer, a space-time Transformer for Earth system forecasting. Earthformer is based on a generic, flexible and efficient space-time attention block, named Cuboid Attention. The idea is to decompose the data into cuboids and apply cuboid-level self-attention in parallel. These cuboids are further connected with a collection of global vectors. We conduct experiments on the MovingMNIST dataset and a newly proposed chaotic N-body MNIST dataset to verify the effectiveness of cuboid attention and figure out the best design of Earthformer. Experiments on two real-world benchmarks about precipitation nowcasting and El Nino/Southern Oscillation (ENSO) forecasting show Earthformer achieves state-of-the-art performance. Code is available: https://github.com/amazon-science/earth-forecasting-transformer .

cs.LG↗

arXiv2020

A Study on Nice Open Covers in Constructive Analysis

Mathematicians like Markov and Bishop made an effort to develop constructive mathematics and extended many theorems in classical mathematical analysis. Heine Borel theorem tells us that a closed bounded subset of Euclidean space R is compact, but in constructive mathematics, Tseitin and Zaslavskii showed that the set of all constructive real numbers between 0 and 1 is not compact. We are going to show that when giving certain restriction to the open cover on [0,1], we can however always choose a finite sub-cover.

math.LO↗

arXiv2019

Fully Using Classifiers for Weakly Supervised Semantic Segmentation with Modified Cues

This paper proposes a novel weakly-supervised semantic segmentation method using image-level label only. The class-specific activation maps from the well-trained classifiers are used as cues to train a segmentation network. The well-known defects of these cues are coarseness and incompleteness. We use super-pixel to refine them, and fuse the cues extracted from both a color image trained classifier and a gray image trained classifier to compensate for their incompleteness. The conditional random field is adapted to regulate the training process and to refine the outputs further. Besides initializing the segmentation network, the previously trained classifier is also used in the testing phase to suppress the non-existing classes. Experimental results on the PASCAL VOC 2012 dataset illustrate the effectiveness of our method.

cs.CV↗

arXiv2017

Deep Learning for Precipitation Nowcasting: A Benchmark and A New Model

With the goal of making high-resolution forecasts of regional rainfall, precipitation nowcasting has become an important and fundamental technology underlying various public services ranging from rainstorm warnings to flight safety. Recently, the Convolutional LSTM (ConvLSTM) model has been shown to outperform traditional optical flow based methods for precipitation nowcasting, suggesting that deep learning models have a huge potential for solving the problem. However, the convolutional recurrence structure in ConvLSTM-based models is location-invariant while natural motion and transformation (e.g., rotation) are location-variant in general. Furthermore, since deep-learning-based precipitation nowcasting is a newly emerging area, clear evaluation protocols have not yet been established. To address these problems, we propose both a new model and a benchmark for precipitation nowcasting. Specifically, we go beyond ConvLSTM and propose the Trajectory GRU (TrajGRU) model that can actively learn the location-variant structure for recurrent connections. Besides, we provide a benchmark that includes a real-world large-scale dataset from the Hong Kong Observatory, a new training loss, and a comprehensive evaluation protocol to facilitate future research and gauge the state of the art.

cs.CV↗

arXiv2016

Approximating (Unweighted) Tree Augmentation via Lift-and-Project (Part 0: $1.8+ε$ approximation for (Unweighted) TAP)

We study the unweighted Tree Augmentation Problem (TAP) via the Lasserre (Sum of Squares) system. We prove an approximation guarantee of ($1.8+ε$) relative to an SDP relaxation, which matches the combinatorial approximation guarantee of Even, Feldman, Kortsarz and Nutov in ACM TALG (2009), where $ε>0$ is a constant. We generalize the combinatorial analysis of integral solutions of Even, et al., to fractional solutions by identifying some properties of fractional solutions of the Lasserre system via the decomposition result of Rothvoß (arXiv:1111.5473, 2011) and Karlin, Mathieu and Nguyen (IPCO 2011).

cs.DS↗

arXiv2015

Approximating (Unweighted) Tree Augmentation via Lift-and-Project, Part II

In Part II, we study the unweighted Tree Augmentation Problem (TAP) via the Lasserre (Sum~of~Squares) system. We prove that the integrality ratio of an SDP relaxation (the Lasserre tightening of an LP relaxation) is $\leq \frac{3}{2}+ε$, where $ε>0$ can be any small constant. We obtain this result by designing a polynomial-time algorithm for TAP that achieves an approximation guarantee of ($\frac32+ε$) relative to the SDP relaxation. The algorithm is combinatorial and does not solve the SDP relaxation, but our analysis relies on the SDP relaxation. We generalize the combinatorial analysis of integral solutions from the previous literature to fractional solutions by identifying some properties of fractional solutions of the Lasserre system via the decomposition result of Karlin, Mathieu and Nguyen (IPCO 2011).

cs.DS↗

arXiv2015

Approximating (Unweighted) Tree Augmentation via Lift-and-Project, Part I: Stemless TAP

In Part I, we study a special case of the unweighted Tree Augmentation Problem (TAP) via the Lasserre (Sum of Squares) system. In the special case, we forbid so-called stems; these are a particular type of subtree configuration. For stemless TAP, we prove that the integrality ratio of an SDP relaxation (the Lasserre tightening of an LP relaxation) is $\leq \frac{3}{2}+ε$, where $ε>0$ can be any small constant. We obtain this result by designing a polynomial-time algorithm for stemless TAP that achieves an approximation guarantee of ($\frac32+ε$) relative to the SDP relaxation. The algorithm is combinatorial and does not solve the SDP relaxation, but our analysis relies on the SDP relaxation. We generalize the combinatorial analysis of integral solutions from the previous literature to fractional solutions by identifying some properties of fractional solutions of the Lasserre system via the decomposition result of Karlin, Mathieu and Nguyen (IPCO 2011). Also, we present an example of stemless TAP such that the approximation guarantee of $\frac32$ is tight for the algorithm. In Part II of this paper, we extend the methods of Part I to prove the same results relative to the same SDP relaxation for TAP.

cs.DS↗

arXiv2015

On the metric s-t path Traveling Salesman Problem

We study the metric $s$-$t$ path Traveling Salesman Problem (TSP). [An, Kleinberg, and Shmoys, STOC 2012] improved on the long standing $\frac{5}{3}$-approximation factor and presented an algorithm that achieves an approximation factor of $\frac{1+\sqrt{5}}{2}\approx1.61803$. Later [Sebő, IPCO 2013] further improved the approximation factor to $\frac{8}{5}$. We present a simple, self-contained analysis that unifies both results; our main contribution is a \emph{unified correction vector}. We compare two different linear programming (LP) relaxations of the $s$-$t$ path TSP, namely, the path version of the Held-Karp LP relaxation for TSP and a weaker LP relaxation, and we show that both LPs have the same (fractional) optimal value. Also, we show that the minimum-cost of integral solutions of the two LPs are within a factor of $\frac{3}{2}$ of each other. We prove that a half-integral solution of the stronger LP-relaxation of cost $c$ can be rounded to an integral solution of cost at most $\frac{3}{2}c$. Additionally, we give a bad instance that presents obstructions to two obvious methods that aim for an approximation factor of $\frac{3}{2}$.

cs.DS↗

arXiv2014

Ultrafast Terahertz Probe of Transient Evolution of Charged and Neutral Phase of Photoexcited Electron-hole Gas in Monolayer Semiconductor

We investigate the dynamical formation of excitons from photoexcited electron-hole plasma and its subsequent decay dynamics in monolayer MoS2 grown by chemical vapor deposition using ultrafast pump and terahertz probe spectroscopy. Different photoexcited electron-hole states are resolved based on their distinct responses to THz photon and decay lifetime. The observed transient THz transmission can be fit with two decay components: a fast component with decay lifetime of 20 ps, which is attributed to exciton life time including the exciton formation and subsequent intraexciton relaxation; a slow component with extremely long decay lifetime of several ns due to either localized exciton state or a long live dark exciton state which is uncovered for the first time. The relaxation dynamics is further verified by temperature and pump fluence dependent studies of the decay time constants.

cond-mat.mtrl-sci↗

arXiv2014

On Integrality Ratios for Asymmetric TSP in the Sherali-Adams Hierarchy

We study the ATSP (Asymmetric Traveling Salesman Problem), and our focus is on negative results in the framework of the Sherali-Adams (SA) Lift and Project method. Our main result pertains to the standard LP (linear programming) relaxation of ATSP, due to Dantzig, Fulkerson, and Johnson. For any fixed integer $t\geq 0$ and small $ε$, $0<ε\ll{1}$, there exists a digraph $G$ on $ν=ν(t,ε)=O(t/ε)$ vertices such that the integrality ratio for level~$t$ of the SA system starting with the standard LP on $G$ is $\ge 1+\frac{1-ε}{2t+3} \approx \frac43, \frac65, \frac87, \dots$. Thus, in terms of the input size, the result holds for any $t = 0,1,\dots,Θ(ν)$ levels. Our key contribution is to identify a structural property of digraphs that allows us to construct fractional feasible solutions for any level~$t$ of the SA system starting from the standard~LP. Our hard instances are simple and satisfy the structural property. There is a further relaxation of the standard LP called the balanced LP, and our methods simplify considerably when the starting LP for the SA system is the balanced~LP; in particular, the relevant structural property (of digraphs) simplifies such that it is satisfied by the digraphs given by the well-known construction of Charikar, Goemans and Karloff (CGK). Consequently, the CGK digraphs serve as hard instances, and we obtain an integrality ratio of $1 +\frac{1-ε}{t+1}$ for any level~$t$ of the SA system, where $0<ε\ll{1}$ and the number of vertices is $ν(t,ε)=O((t/ε)^{(t/ε)})$. Also, our results for the standard~LP extend to the Path-ATSP (find a min cost Hamiltonian dipath from a given source vertex to a given sink vertex).

cs.DS↗

arXiv2013

An LP-based 3/2-approximation algorithm for the graphic s-t path TSP

We design a new LP-based algorithm for the graphic $s$-$t$ path Traveling Salesman Problem (TSP), which achieves the best approximation factor of 1.5. The algorithm is based on the idea of narrow cuts due to An, Kleinberg, and Shmoys. It partly answers an open question of Sebő.

cs.DS↗

arXiv2012

Approximating Minimum-Cost Connected T-Joins

We design and analyse approximation algorithms for the minimum-cost connected T-join problem: given an undirected graph G = (V;E) with nonnegative costs on the edges, and a subset of nodes T, find (if it exists) a spanning connected subgraph H of minimum cost such that every node in T has odd degree and every node not in T has even degree; H may have multiple copies of any edge of G. Two well-known special cases are the TSP (|T| = 0) and the s-t path TSP (|T| = 2). Recently, An, Kleinberg, and Shmoys [STOC 2012] improved on the long-standing 5/3-approximation guarantee for the latter problem and presented an algorithm based on LP rounding that achieves an approximation guarantee of (1+sqrt(5))/2 < 1.6181. We show that the methods of An et al. extend to the minimum-cost connected T-join problem. They presented a new proof for a 5/3-approximation guarantee for the s-t path TSP; their proof extends easily to the minimum-cost connected T-join problem. Next, we improve on the approximation guarantee of 5/3 by extending their LP-rounding algorithm to get an approximation guarantee of 13/8 = 1.625 for all |T| >= 4. Finally, we focus on the prize-collecting version of the problem, and present a primal-dual algorithm that is "Lagrangian multiplier preserving" and that achieves an approximation guarantee 3 - 2/(|T|-1) when |T| >= 4. Our primal-dual algorithm is a generalization of the known primal-dual 2-approximation for the prize-collecting s-t path TSP. Furthermore, we show that our analysis is tight by presenting instances with |T| >= 4 such that the cost of the solution found by the algorithm is exactly 3 - 2/(|T|-1) times the cost of the constructed dual solution.

cs.DS↗