About

Dr. Zhou Zhang leads research at the Digital Agriculture Research Lab within the Biological Systems Engineering department at the University of Wisconsin–Madison. His interdisciplinary work focuses on integrating advanced remote sensing technologies with machine learning techniques to address challenges in agricultural applications. Specifically, his research involves the fusion of multi-source remote sensing data, including hyperspectral and multispectral imagery, RGB imagery, and LiDAR, to enhance data analysis capabilities in agriculture. Dr. Zhang applies machine learning methods for high-dimensional data analysis, encompassing feature extraction, classification, and regression tasks, to improve the understanding and prediction of crop growth and yield. Additionally, he develops UAV-based imaging platforms aimed at precision agriculture, facilitating high-throughput image-based plant phenotyping and enabling more accurate crop monitoring and yield prediction through the combination of remote sensing and machine learning approaches.

Research topics

• Physics
• Particle physics
• Nuclear physics
• Computer Science
• Optics
• Quantum mechanics
• Astrophysics
• Mathematics
• Algorithm
• Database
• Operating system
• Statistics
• Geology
• Simulation
• Condensed matter physics
• Biology
• Combinatorics
• Telecommunications
• Engineering
• Aerospace engineering
• Atomic physics
• Computer hardware
• Real-time computing
• Computational science

Selected publications

• An open-set domain adaptation approach for cross-region crop mapping

  International Journal of Remote Sensing · 2026-03-26

  articleSenior authorCorresponding
  PublisherDOI

• Knowledge-Guided Machine Learning for County-Level Corn Yield Prediction Under Drought

  SSRN Electronic Journal · 2025-01-01 · 1 citations

  preprintOpen accessSenior author
  PublisherDOI

• Symbolic regression for discovering vegetation indices: a case study using alfalfa biomass estimation

  2025-09-11

  preprintOpen accessSenior author

  Predicting crop yield accurately is crucial for agricultural management. Traditional methods often rely on predefined vegetation indices derived from remote sensing data, which may not fully capture the complex relationships present in high dimensional hyperspectral datasets. We explored the potential of symbolic regression (SR), a machine learning technique that discovers best fit mathematical expressions from data, to develop a novel and interpretable model for predicting alfalfa biomass. Using hyperspectral reflectance data with 274 bands in the visible to near infrared, and corresponding alfalfa biomass measurements from 87 distinct sampling points, the PySR library was employed to search for optimal equations. Analyses revealed a trade-off between model complexity and accuracy. A moderately complex equation (complexity 22) involving key wavelengths in the red-edge (~779 nm) and near-infrared (~932 nm, ~952 nm, ~968 nm) regions, associated with chlorophyll content, canopy structure, and water content, achieved a coefficient of determination (r²) of approximately 0.71 on the training data. While demonstrating the capability of SR to uncover potentially meaningful relationships, the results also highlighted limitations such as systematic under-prediction at lower yields and potential feature saturation at higher yields. Selected models tended to include only three or four spectral bands, and a search for simpler expressions that follow normalized difference formulations revealed multiple options that further improved fit. SR may be useful for discovery in hyperspectral remote sensing but must be balanced against the simplicity and explainability of existing formulae.

  PublisherOA PDFDOI

• Dynamic Modelling and Experimental Investigation of an Active–Passive Variable Stiffness Actuator

  Actuators · 2025-03-29 · 3 citations

  articleOpen access

  To overcome the limitations imposed by the low flexible angle of conventional robots, an active–passive variable stiffness elastic actuator (APVSA) is investigated and a nonlinear dynamic model for the APVSA is established, considering the factors of the moment of inertia, stiffness and damping of elastic elements, meshing stiffness of gear systems, nonlinear backlash, nonlinear meshing damping, and comprehensive transmission error. The established dynamic model is discretized by the forward Euler method, and the variable stiffness performance and the influence of nonlinear factors on the APVSA are analysed by Adams and Simulink simulations, respectively. A physical prototype and an experimental platform were assembled, and the dynamic and static variable stiffness experiments were conducted. The experimental results realized the expected stiffness adjustment target and provided the foundation for the next step of control.

  PublisherOA PDFDOI

• P‐68: Double‐Data PHM Drive System Based on MicroLED Display

  SID Symposium Digest of Technical Papers · 2025-06-01

  articleOpen access

  This paper introduces a Double Data PHM drive system, focusing on PHM dimming technology, system design and verification The system is implemented on FPGA and applied to the Micro LED display based on LTPS glass substrate. Compared with PWM dimming technology and PAM dimming technology, the amplitude and width of LED driving pulse in PHM dimming technology can be adjusted, and the control of Micro LED display brightness is more flexible [1] K At present, the feasibility of Double Data PHM drive system has been verified on TCL CSOT 6.79 inch Micro LED products.

  PublisherOA PDFDOI

• Interpretable data-driven chemometric approach for predicting non-optically active water quality parameters using ultraviolet-visible-near infrared absorption spectroscopy and physical-chemical measurements

  Spectrochimica Acta Part A Molecular and Biomolecular Spectroscopy · 2025-01-23 · 6 citations

  articleOpen access

  • Developed a data-driven chemometric approach for predicting NAWQPs. • An end-to-end ensemble learning algorithm without spectral preprocessing. • Integrated UV–Vis-NIR spectra and physical–chemical measurements for better accuracy. • Post-hoc model interpretation proves that the UV band is extremely important. • Provided a rapid, eco-friendly method for in-situ water quality monitoring . Non-optically active water quality parameters (NAWQPs) are essential for surface water quality assessments, although automated monitoring methods are time-consuming, include labor-intensive chemical pretreatment, and pose challenges for high spatiotemporal resolution monitoring. Advancements in spectroscopic techniques and machine learning may address these issues. We integrated ultraviolet–visible-near infrared absorption spectroscopy with physical-chemical measurements to predict total nitrogen (TN), dissolved oxygen (DO), and total phosphorus (TP) in the Yangtze River Basin, China. By combining the eXtreme Gradient Boosting algorithm with OPTUNA hyperparameter optimization and the SHapley Additive exPlanations interpretability framework, we developed an algorithm that yielded Nash–Sutcliffe efficiency values of 0.944, 0.934, and 0.835, and mean absolute percentage errors of 7.8 %, 8.2 %, and 7.7 % for TN, DO, and TP, respectively. The UV spectrum was significant in the NAWQPs prediction tasks. Our study offers a novel approach to water quality monitoring and resource management in complex aquatic environments.

  PublisherDOI

• Mitigating NDVI saturation in imagery of dense and healthy vegetation

  ISPRS Journal of Photogrammetry and Remote Sensing · 2025-06-18 · 18 citations

  articleSenior authorCorresponding
  PublisherDOI

• High-resolution surface and rootzone soil moisture over US cropland: A novel framework assimilating multi-source remote sensing data, machine learning, and the Layered Green and Ampt Infiltration with Redistribution model

  Remote Sensing of Environment · 2025-12-06 · 3 citations

  articleOpen access

  Accurate and high spatiotemporal resolution soil moisture (SM) monitoring in cropland is important for water resource management, drought forecasting, and nutrient transport estimation at the field scale for sustainable crop production. Although recent research has applied machine learning (ML) to downscale coarse-resolution satellite SM products, most of this past work has focused only on surface SM estimation, and the performance of rootzone SM products has not been intensively evaluated in cropland. This study introduces a novel framework that integrates multi-source satellite-based ML models with the Layered Green and Ampt Infiltration with Redistribution (LGAR) model to produce high-resolution (100 m, hourly) SM products for both the surface layer (0–5 cm) and rootzone (0–100 cm) across cropland in the contiguous United States (CONUS). First, six ML models were trained using multiple high-resolution remote sensing datasets (Sentinel-1, Sentinel-2, and Landsat) to predict surface and rootzone SM. These ML predictions were then assimilated into the LGAR model using the ensemble Kalman filter (EnKF). The framework was developed and validated using an eight-fold cross-validation scheme with in-situ data from 431 cropland sites across CONUS, sourced from three networks (SCAN, USCRN, and PSA). The 100-m hourly SM data from this framework surpasses existing products (9-km SMAP L4, SMAP-based 1-km thermal hydraulic disaggregation of SM product) in spatial and temporal resolution and captures rootzone SM that is not available in the SMAP-HydroBlocks SM product. It achieves good performance, with median bias-corrected root mean squared error (ubRMSE) of 0.053 m3/m3 and median Kling-Gupta efficiency (KGE) of 0.379 in the surface layer, and median ubRMSE of 0.027 m3/m3 and median KGE of 0.302 in the rootzone. While the framework demonstrates strong performance, its accuracy varies across climatic regimes, with surface SM performing better in non-humid areas (median KGE = 0.375 versus median KGE = 0.416) and rootzone SM in humid regions (median KGE = 0.313 versus median KGE = 0.127). This high-resolution cropland SM product can potentially benefit multiple agricultural applications, such as irrigation management and nutrient leaching estimation, and provide valuable insights to support farmers and land managers in decision-making processes.

  PublisherDOI

• Uncrewed Aerial Vehicle (UAV)-Based High-Throughput Phenotyping of Maize Silage Yield and Nutritive Values Using Multi-Sensory Feature Fusion and Multi-Task Learning with Attention Mechanism

  Remote Sensing · 2025-11-06 · 1 citations

  articleOpen accessSenior authorCorresponding

  Maize (Zea mays L.) silage’s forage quality significantly impacts dairy animal performance and the profitability of the livestock industry. Recently, using uncrewed aerial vehicles (UAVs) equipped with advanced sensors has become a research frontier in maize high-throughput phenotyping (HTP). However, extensive existing studies only consider a single sensor modality and models developed for estimating forage quality are single-task ones that fail to utilize the relatedness between each quality trait. To fill the research gap, we propose MUSTA, a MUlti-Sensory feature fusion model that utilizes MUlti-Task learning and the Attention mechanism to simultaneously estimate dry matter yield and multiple nutritive values for silage maize breeding hybrids in the field environment. Specifically, we conducted UAV flights over maize breeding sites and extracted multi-temporal optical- and LiDAR-based features from the UAV-deployed hyperspectral, RGB, and LiDAR sensors. Then, we constructed an attention-based feature fusion module, which included an attention convolutional layer and an attention bidirectional long short-term memory layer, to combine the multi-temporal features and discern the patterns within them. Subsequently, we employed multi-head attention mechanism to obtain comprehensive crop information. We trained MUSTA end-to-end and evaluated it on multiple quantitative metrics. Our results showed that it is capable of practical quality estimation results, as evidenced by the agreement between the estimated quality traits and the ground truth data, with weighted Kendall’s tau coefficients (τw) of 0.79 for dry matter yield, 0.74 for MILK2006, 0.68 for crude protein (CP), 0.42 for starch, 0.39 for neutral detergent fiber (NDF), and 0.51 for acid detergent fiber (ADF). Additionally, we implemented a retrieval-augmented method that enabled comparable prediction performance, even without certain costly features available. The comparison experiments showed that the proposed approach is effective in estimating maize silage yield and nutritional values, providing a digitized alternative to traditional field-based phenotyping.

  PublisherOA PDFDOI

• A Multimodal Deep Learning Approach for Soil Moisture Downscaling Using Remote Sensing and Weather Data

  Journal of Geophysical Research Machine Learning and Computation · 2025-07-11

  articleOpen accessSenior authorCorresponding

  Abstract Understanding soil moisture (SM) dynamics is crucial for environmental and agricultural applications. While satellite‐based SM products provide extensive coverage, their coarse spatial resolution often fails to capture local SM variability. This study presents a multimodal network (MMNet) that integrates remote sensing and weather data to downscale Soil Moisture Active Passive (SMAP) Level‐4 surface SM. We evaluated the performance of MMNet by comparing it with in situ SM observations from the Soil Climate Analysis Network (SCAN) and the United States Climate Reference Network (USCRN) under three scenarios. The results showed that (a) MMNet trained with on‐site data provided accurate SM estimates over time in withheld years; (b) MMNet demonstrated spatial transferability, capturing SM dynamics in regions with sparse or no in situ measurements; and (c) the integration of snapshot and time‐series data was crucial for maintaining the model's accuracy and generalizability across diverse scenarios. The downscaled SM maps demonstrated its potential for producing high‐resolution temporally and spatially continuous SM estimates, which could further support a broad range of environmental and agricultural applications.

  PublisherOA PDFDOI

Frequent coauthors

• T. Beau

  Consejo Nacional de Investigaciones Científicas y Técnicas

  10956 shared

• L. Roos

  Laboratoire de Physique Nucléaire et de Hautes Énergies

  9713 shared

• S. Trincaz-Duvoid

  Laboratoire de Physique Nucléaire et de Hautes Énergies

  9699 shared

• J. Ocariz

  Université Paris Cité

  9699 shared

• M. Ridel

  Université Paris Cité

  9683 shared

• B. Trocmé

  Laboratoire AstroParticule et Cosmologie

  8850 shared

• S. De Cecco

  Radboud University Nijmegen

  8218 shared

• P. A. Delsart

  Université Grenoble Alpes

  8174 shared

Education

• Ph.D., Biological Systems Engineering

  University of Wisconsin–Madison

  2010

• M.S., Biological Systems Engineering

  University of Wisconsin–Madison

  2006

• B.S., Agricultural and Biological Engineering

  University of Wisconsin–Madison

  2004

Awards & honors

• Vilas Faculty Early-Career Investigator Award, UW-Madison
• Alfred Toepfer Faculty Fellow Award, UW-Madison
• Madison Teaching and Learning Excellence Fellow, UW-Madison