SITS: WIRELESS SOIL SENSING AND ADVANCED DATA-DRIVEN ANALYSIS FOR POST-WILDFIRE HAZARDS AND RECOVERY IN MOUNTAINOUS LANDSCAPES

• Sponsoring Institution: National Institute of Food and Agriculture
• Project Status: ACTIVE
• Funding Source: AFRI COMPETITIVE GRANT
• Reporting Frequency: Annual
• Accession No.: 1029640
• Grant No.: 2023-67019-38829
• Cumulative Award Amt.: $250,000.00
• Proposal No.: 2022-09950
• Project Start Date: Jan 1, 2023
• Project End Date: Dec 31, 2025
• Grant Year: 2023
• Program Code: [A1401]- Foundational Program: Soil Health

Recipient Organization
UNIV OF SOUTH FLORIDA
(N/A)
TAMPA,FL 33620

Performing Department
(N/A)

Non Technical Summary
After wildfire, burned hillslopes are left in a precarious and often unstable state. However, without sufficient monitoring on site, the timeframe for which site conditions are predisposed to erosion and landslides is largely unknown. The proposed Burned Area Monitoring System (BAMS) will monitor soil state using passive wireless tracers in combination with smart wireless interrogation to collect data on soil displacement, soil temperature, and soil moisture. These data will support development and validation of models for assessing post-wildfire landslide or debris flow potential of burnt landscapes in temperate climates. As tracers move downslope with the soil and can be expected to become buried and unburied. In the former case, the tracers provide an opportunity to garner soil temperature and moisture data in addition to soil displacement information. BAMS is a highly asymmetric design where energy and costs are concentrated with a small number of interrogators serving a deployment site. This design allows the system to adapt dynamically, through machine learning, to changing spatial and temporal sensing needs. Not only does this allow the passive wireless tracers to be low-cost, but their simplistic functionality enables a significant advancement in in situ monitoring in that these devices will be biodegradable, leaving only trace inert elements (i.e., silicon, silver) in the environment.

Animal Health Component

40%

Research Effort Categories

Basic

20%

Applied

40%

Developmental

40%

Classification

Knowledge Area (KA)	Subject of Investigation (SOI)	Field of Science (FOS)	Percent
102	0110	2020	50%
112	0199	2020	50%

Knowledge Area
102 - Soil, Plant, Water, Nutrient Relationships; 112 - Watershed Protection and Management;

Subject Of Investigation
0110 - Soil; 0199 - Soil and land, general;

Field Of Science
2020 - Engineering;

Keywords

hydrological modeling

landslides

machine learning

wildfires

wireless sensing

Goals / Objectives
The proposed research will advance an energy-efficient, low-cost, and long-term sensing platform for in situ soil measurements that will inform our understanding of the dynamics of burned forested hillslopes, their recovery, and the risk of post-wildfire hazards. Wildfires are becoming more frequent and severe worldwide, often resulting in devastating social, economic, and environmental consequences. Some of these impacts include erosion and mass wasting on burned hillslopes driven by altered mechanical, hydrological, and vegetation conditions. With revegetation and time, these site conditions return to a pre-fire equilibrium, but the timing and trajectory of this recovery is poorly-constrained. The proposed soil sensing system will monitor moisture, temperature, and displacement within the soil matrix to support development of a thermo-hydro-mechanical framework of post-wildfire processes induced from climatic controls. Development and validation of a site-scale nondestructive wireless sensing system are proposed at a variety of scales: laboratory soil specimens, laboratory physical models, and at wildfire sites in Oregon. Novel data analysis and modeling using machine learning for collecting and interpreting field observations and data will be developed.Thrust 1: Wireless Tracers and InterrogationThe goal of this thrust is to investigate new devices and methods for high-resolution, long-term, and environmentally-safe embedded soil sensing. The objectives related to the passive wireless tracers (PWT) are to establish design methodologies to achieve orientation-independent and dual-frequency operation; to demonstrate controlled, temperature-dependent operation; and, to explore manufacturing techniques that use biodegradable materials. The objectives for the interrogation research are to leverage a software-defined radio architecture to explore methodologies for remotely sensing soil movement, temperature and moisture using the in situ wireless tracers.Thrust 2: Modeling of Evolving Post-Fire Soil Behavior and ErosionThe objectives of this thrust are to characterize feedbacks between wildfire, altered soil properties, and potential for erosion (both surficial and landslide-driven) as hydrological conditions evolve over time, achieved through development of a thermo-hydro-mechanical model (THM), and quantification of uncertainties in site conditions through strategic sensor placement and in-situ and laboratory soil testing.Thrust 3: Machine Learning-based Data Analysis and ModelingThe objectives of this thrust are to investigate machine learning (ML) algorithms to (i) help process collected PWT data, (ii) quantify uncertainty in THM model, (iii) fingerprint the site for evolving hydrological conditions and instability, and (iv) detect anomalies in the observed data as a potential precursor for landslides.Thrust 4: Validation and DeploymentThe objectives of this thrust include validation and deployment of sensors to a field location that has a mix of burned and unburned conditions from prior wildfire. Sensors will be deployed in two catchments that reflect a gradient of soil burn severity and observed mass wasting events, such as the 2017 Eagle Creek wildfire area in Oregon.

Project Methods
Thrust 1: Wireless Tracers and InterrogationOrientation-Independent, Dual-Frequency Design: We will investigate passive wireless tracer designs that incorporate four pairs of antennas in a compact, 3D form factor. Each pair of antennas includes one receive and one transmit antenna (operating at two different frequencies). The antenna pairs will be oriented such that the tracer can be interrogated from any direction using different polarizations. An optimization strategy to accelerate the design process will be investigated.Temperature Dependency: We will investigate the use of bi-material beams that flex under temperature change and are integrated into the impedance-matching networks of the wireless tracers.Manufacturing and Biodegradable Design: We will investigate the use of biodegradable thermoplastics such as poly(lactide) (PLA) as the substrate and packaging for the wireless tracers.Localization: Device displacement will be monitored through remote localization over time. The multi-antenna interrogator design will allow the signal returned to be received through multiple ports. The signals from these ports will be added and then subtracted and the location will identified by the peak difference. Data collected from multiple interrogator positions will enable tracer depth to be determined.Resonant Frequency: Temperature of the passive wireless tracer will dictate its resonant frequency, which will be determined remotely by sweeping the interrogation signal over the band of operation.Link Attenuation: Soil moisture changes will change the attenuation loss both to and from the tracer. Exploiting the non-linearity of this device, we will investigate methods to remotely ascertain the operating point of the device, thereby isolating the forward link loss, the device conversion loss, and the return link loss. Link losses at two frequencies will provide the data from which soil moisture can be found.Software Design Radio Architecture: The various interrogator operations require a flexible architecture for which we will investigate the use of a software defined radio platform.Thrust 2: Soil Thermo-Hydro-Mechanical (THM) ModelingCharacterization of soil properties: We will gather soil samples from the study site to characterize soil properties for THM modeling and to characterize heterogeneity in site conditions.THM 1D Boundary Value Modeling: Develop THM models of plausible site soil-columns to characterize water fluxes and evolving stresses based on observed pedologic, climatic and vegetative conditions.Catchment-Scale Model for Slope Instability and Landscape Recovery: Extend THM framework to a grid-based model for susceptibility of erosion and mass-wasting to characterize evolving controls on instability at a catchment scale.Thrust 3: Machine Learning-based Data Analysis and ModelingTHM Uncertainty Quantification: We will use a Bayesian network to quantify uncertainty in the parameters and associated behavior of the proposed THM model and its relationship to PWT observations.PWT Data Analysis: We will train ML methods to determine the location, temperature, and moisture of a single PWT from its returned signal.Area Fingerprinting: To characterize an area of interest with heat maps for better understanding and guidance to future PWT deployments, we will build interpolation methods for the observed (soil temperature, moisture, and movement) and inferred (susceptibility index for landslides and debris flows, landscape recovery index) variables.Anomaly Detection: We will also monitor the time series data observed from the PWTs to detect anomalies, which might be precursors of a landslide event and response to climatic conditions (i.e., a comparative signal over time for recovery).Thrust 4: Validation and Demonstration of BAMSPlacement of sensors in two catchments of varying burn severity: We will place a distributed set of sensors throughout catchments of contrasting burn severity (severe and modest/no soil burn severity) to monitor evolving soil moisture conditions as well as movement of sediment.Compare BAMS observations to other models: We will compare observed changes from BAMS to other metrics for change, including differencing of aerial and terrestrial lidar, comparison to climate and hydrologic monitoring stations, and field interpretation.

Progress 01/01/24 to 12/31/24

Outputs
Target Audience:Undergraduate and graduate students. Changes/Problems: Nothing Reported What opportunities for training and professional development has the project provided?Md Mahmuddun Nabi Murad, a second-year PhD student, continued working on the project. He collaborated with a PhD student from Dr. Thomas Weller's group(our project partner in the project)at Oregeon State University (OSU) to develop the deep learningalgorithm for wireless sensor tracking. The OSU team has provided a simulator to generate data for total received signal at interrogators from several passive wireless sensors buried under the soil. Mr. Murad, under Dr. Yilmaz's guidance, developed a time series forecasting algorithm which can accurately track the sensor locations. Moreover, he collaborated with Dr. Ben Leshchinsky's group from OSU on the soil moisture monitoring problem using data collected from weather stations in Corvallis, Oregon. How have the results been disseminated to communities of interest?Our work on time series forecasting which presents a novelalgorithm with superior performance on several datasetsis accepted by The 39th Annual AAAI Conference on Artificial Intelligence, one of the top machine learning/AI conferences. Mr. Murad will present it in Philadelphia in February. Mr. Murad will also present our work on wireless sensor tracking at The 2025 IEEE Wireless Communications and Networking Conference (WCNC)in Milan, Italy, in March. What do you plan to do during the next reporting period to accomplish the goals?In the last year, we obtained state-of-the-art results on different time series forecasting datasets using different deep neural network architectures. For example, while wavelet decomposition improved the performance in electric power forecasting and sensor tracking, it did not help with ship route prediction.We will investigate the reasons why certain architectures work better with certain datasets with the aim of developing a framework for automated optimization of deep neural networkarchitecture based on the considered dataset. We plan to extend our time series forecasting algorithms to the time series anomaly detection problem. We will investigate ways to improve the state-of-the-art methods in terms of quick and accurate detection of anomalies. We also plan to write a journal paper on ship route prediction by expanding our results on the M3 dataset. To this end, we are collecting ship movement data from surveillance cameralive streams around the world. We will continue our collaborations with the OSU team on sensor tracking and soil moisture monitoring. The objective is to apply our algorihtms to real data that will be collected in Oregon this year. Dr. Weller's team is planning to perform field experiments with their sensors and interrogators. Dr. Leshchinsky will also provide us with more data from his weather stations for a detailed work on moisture forecasting.

Impacts
What was accomplished under these goals? Under Thrust 3, we developed different time series forecasting methods for soil moisture monitoring andwireless sensor tracking, as explained below. Although we started with thosespecific problems motivated by this project, the developed deep learning algorithms are applicable to other time series forecasting problems in different fields, such as predicting ship routes, electric power, weather, and traffic. Several different deep neural network architecures, e.g., encoder-only and encoder-decoder, and deep learning techniques such asattention-based transformers andmixing-based multilayer perceptrons,have been investigated to improve the state-of-the-art forecasting performance in various benchmark datasets. Moreover, preproressing techniques such as wavelet decomposition have been investigated. Our algorithms signficantlyoutperformed the state-of-the-art results on soil moisture monitoring and tracking of passive wireless sensors used for soil moisture monitoring, as well asseveralbenchmark datasets in other domains, including M3 ship ETTh1, ETTh2, ETTm1, ETTm2, Weather, Electricity, and Traffic.

Publications

• Type: Conference Papers and Presentations Status: Accepted Year Published: 2025 Citation: M. Murad, D. Yirenya-Tawiah, T. Weller, Y. Yilmaz, "Multi-Resolution Mixer Network for Localization of Multiple Sensors from Cumulative Power Measurements", Proceedings of the IEEE Wireless Communications and Networking Conference, 2025
• Type: Conference Papers and Presentations Status: Accepted Year Published: 2025 Citation: M. Murad, M. Aktukmak, Y. Yilmaz, "WPMixer: Efficient Multi-Resolution Mixing for Long-Term Time Series Forecasting", Proceedings of the 39th Annual AAAI Conference on Artificial Intelligence, 2025

Progress 01/01/23 to 12/31/23

Outputs
Target Audience:Undergraduate and graduate students. Changes/Problems: Nothing Reported What opportunities for training and professional development has the project provided?A graduate student, Md Mahmudun Nabi Murad, was trained on the two problems explained above, soil moisture forecasting and sensor localization. He is a first-yearPhD student and started on working the project in August 2023. During the project meeting in summer a second-year PhD student, Justin McMillen, participated in the discussions. How have the results been disseminated to communities of interest?Project has been introduced to undergraduate and graduate students in the Signals and Systems class and the Advanced Data Analytics class by presenting the project objectives and pictures from the field trip in Oregon taken during the kick-off meeting. What do you plan to do during the next reporting period to accomplish the goals?Three papers are in progress and planned to be submitted during the next reporting period. One of them is on a novel deep learning technique, aBayesian transformer, for time series anomaly forecasting and anomaly detection. The second one is on soil moisture forecasting in collaboration with the project partner Dr. Ben Leshchinsky from the Oregon State University, College of Forestry. The third one is on the localization and tracking of multiple passive sensors in the soil, which reflect the signal transmitted from an interrogator,using received signal strength at the interrogators and probabilistic inference techniques. This work is in collaboration with the project partner Dr. Thomas Weller from theOregon State University, Department of Electrical Engineering and Computer Science.

Impacts
What was accomplished under these goals? Project meeting was held in Corvallis, Oregon between 8/14/23 and 8/16/23. The interdisciplinary project team visited several burned forested hillslope areas during the first two days to better understand the problem and discuss potential research strategies for studying the problem. The last day of the meeting was reserved for a lengthy discussion of future research directions. Project agenda for the first year was agreed on during that discussion. In accordance with the project agenda, our team focused on developing machine learning techniques for soil moisture forecasting and sensor localization under Thrust 3. With the help of a PhD student, PI Yilmaz have trained several deep learning methods, namely time series transformers, on the sensor data obtained from the project partner Dr. Ben Leshchinsky from the Oregon State University. We successfully demonstrated soil moisture forecasting capabilities on the hourly sampled and 15-minute sapmpled data. The achieved normalized root-mean-square forecasting error is on the order of 0.002 with the 15-minute data and 0.005 with the hourly data. These errors represent around 30% performance improvement over the traditional autoregressive (AR) models. We also worked on localizing the passive sensors under the soil to measure their displacement over time. Probabilistic inference methods, in particular Bayesian techniques, have been developed to accurately identify the location of multiple sensors under soil using the strength of reflected interrogator signals from the passive sensors. The probabilistic localization methods have been developed using a simulator provided by the project partner Dr. Thomas Wellerfrom the Oregon State University.

Publications