Precision agriculture demands continuous and accurate monitoring of soil moisture M and key macronutrients, including nitrogen N, phosphorus P, and potassium K, to optimize yields and conserve resources. Wireless soil sensing has been explored to measure these four components; however, current solutions require recalibration (i.e., re-train the data processing model) to handle variations in soil texture (characterized by aluminosilicates Al and organic carbon C, limiting their practicality. To address this, we introduce SoilX, a calibration-free soil sensing system that jointly measures six key components: {M, N, P, K, C, Al}. By explicitly modeling C and Al, SoilX eliminates texture- and carbon-dependent recalibration. SoilX incorporates Contrastive Cross-Component Learning (3CL), with two customized terms: the Orthogonality Regularizer and Separation Loss, to effectively disentangle cross-component interference. Additionally, we design a novel Tetrahedral Antenna Array with an antenna-switching mechanism, which can robustly measure soil dielectric permittivity independent of device placement. Extensive experiments demonstrate that SoilX reduces estimation errors by 23.8% to 31.5% over baselines and generalizes well to unseen fields.

Spatiotemporal reasoning plays a key role in Cyber-Physical Systems (CPS). Despite advances in Large Language Models (LLMs) and Large Reasoning Models (LRMs), their capacity to reason about complex spatiotemporal signals remains underexplored. This paper proposes a hierarchical SpatioTemporal reAsoning benchmaRK, STARK, to systematically evaluate LLMs across three levels of reasoning complexity: state estimation (e.g., predicting field variables, localizing and tracking events in space and time), spatiotemporal reasoning over states (e.g., inferring spatial-temporal relationships), and world-knowledge-aware reasoning that integrates contextual and domain knowledge (e.g., intent prediction, landmark-aware navigation). We curate 26 distinct spatiotemporal tasks with diverse sensor modalities, comprising 14,552 challenges where models answer directly or by Python Code Interpreter. Evaluating 3 LRMs and 8 LLMs, we find LLMs achieve limited success in tasks requiring geometric reasoning (e.g., multilateration or triangulation), particularly as complexity increases. Surprisingly, LRMs show robust performance across tasks with various levels of difficulty, often competing or surpassing traditional first-principle-based methods. Our results show that in reasoning tasks requiring world knowledge, the performance gap between LLMs and LRMs narrows, with some LLMs even surpassing LRMs. However, the LRM o3 model continues to achieve leading performance across all evaluated tasks, a result attributed primarily to the larger size of the reasoning models. STARK motivates future innovations in model architectures and reasoning paradigms for intelligent CPS by providing a structured framework to identify limitations in the spatiotemporal reasoning of LLMs and LRMs.

Multimodal deep learning systems are deployed in dynamic scenarios due to the robustness afforded by multiple sensing modalities. Nevertheless, they struggle with varying compute resource availability (due to multi-tenancy, device heterogeneity, etc.) and fluctuating quality of inputs (from sensor feed corruption, environmental noise, etc.). Statically provisioned multimodal systems cannot adapt when compute resources change over time, while existing dynamic networks struggle with strict compute budgets. Additionally, both systems often neglect the impact of variations in modality quality. Consequently, modalities suffering substantial corruption may needlessly consume resources better allocated towards other modalities. We propose ADMN, a layer-wise Adaptive Depth Multimodal Network capable of tackling both challenges - it adjusts the total number of active layers across all modalities to meet compute resource constraints, and continually reallocates layers across input modalities according to their modality quality. Our evaluations showcase ADMN can match the accuracy of state-of-the-art networks while reducing up to 75% of their floating-point operations.

Synthesizing radio-frequency (RF) data given the transmitter and receiver positions, e.g., received signal strength indicator (RSSI), is critical for wireless networking and sensing applications, such as indoor localization. However, it remains challenging due to complex propagation interactions, including reflection, diffraction, and scattering. State-of-the-art neural radiance field (NeRF)-based methods achieve high-fidelity RF data synthesis but are limited by long training times and high inference latency. We introduce GSRF, a framework that extends 3D Gaussian Splatting (3DGS) from the optical domain to the RF domain, enabling efficient RF data synthesis. GSRF realizes this adaptation through three key innovations: First, it introduces complex-valued 3D Gaussians with a hybrid Fourier–Legendre basis to model directional and phase-dependent radiance. Second, it employs orthographic splatting for efficient ray–Gaussian intersection identification. Third, it incorporates a complex-valued ray tracing algorithm, executed on RF-customized CUDA kernels and grounded in wavefront propagation principles, to synthesize RF data in real time. Evaluated across various RF technologies, GSRF preserves high-fidelity RF data synthesis while achieving significant improvements in training efficiency, shorter training time, and reduced inference latency.

Soil composition sensing is essential for precision agriculture, sustainable land management, and optimizing crop yields. However, existing sensing systems face major limitations, including extensive calibration needs to account for soil variability, a narrow focus on measurements for specific properties, and sensitivity to the placement of the device. These challenges hinder practical deployment. This demo shows SoilX, a comprehensive wireless soil sensing system that quantifies all major soil components---including aluminosilicates, water, organic carbon, and micronutrients---using RF and VNIR sensing technologies. To enable the generalizability, SoilX employs contrastive pretraining to mitigate cross-component interference. Additionally, a tetrahedron-based antenna geometry ensures robustness to device placements. Extensive evaluations in both lab and field settings demonstrate that SoilX achieves state-of-the-art accuracy in soil composition analysis with low costs.

Effective communication and sensing in next-generation wireless technologies require resource-intensive site surveys for data collection. These surveys capture key RF characteristics at various positions, including the Received Signal Strength Indicator (RSSI) and spatial spectrum (RSSI measured from all directions around the receiver). An alternative approach is Radio-Frequency (RF) signal spatial propagation modeling, which predicts received signals based on transceiver positions. Existing Neural Radiance Field (NeRF)-based methods exhibit a fundamental trade-off between scalability and fidelity. To address this challenge, we explore leveraging 3D Gaussian Splatting, an advanced technique for real-time image synthesis of 3D scenes from arbitrary camera poses. This work develops RFSPM, an end-to-end 3D Gaussian distribution-based framework for scalable RF signal Spatial Propagation Modeling. We evaluate RFSPM in spatial spectrum synthesis, demonstrating its learning efficiency compared to NeRF-based methods.

Precise robot localization at the tree level is essential for smart agriculture applications such as precision disease management and targeted nutrient distribution. Existing methods fail to achieve the required accuracy. We propose OrchLoc, a fingerprinting-based localization solution that achieves treelevel precision using a single Long Range (LoRa) gateway. Our approach utilizes channel state information (CSI) across eight channels as a localization fingerprint. To minimize labor-intensive site surveys for fingerprint database construction and maintenance, we develop a CSI generative model (CGM) that learns the relationship between CSI vectors and their corresponding locations. The CGM is fine-tuned using CSI data from static agricultural LoRa sensor nodes, enabling continuous fingerprint database updates. Extensive experiments in two orchards demonstrate that OrchLoc effectively achieves accurate tree-level localization with minimal overhead, improving robot navigation

LoRa networks have been deployed in many orchards for environmental monitoring and crop management. An accurate propagation model is essential for efficiently deploying a LoRa network in orchards, e.g., determining gateway coverage and sensor placement. Although some propagation models have been studied for LoRa networks, they are not suitable for orchard environments, because they do not consider the shadowing effect on wireless propagation caused by the ground and tree canopies. This article presents FLog , a propagation model for LoRa signals in orchard environments. FLog leverages a unique feature of orchards, i.e., all trees have similar shapes and are planted regularly in space. We develop a three-dimensional model of orchards. Once we have the location of a sensor and a gateway, we know the media that the wireless signal traverses. Based on this knowledge, we generate the First Fresnel Zone (FFZ) between the sender and the receiver. The intrinsic path loss exponents of all media can be combined into a classic Log-Normal Shadowing model in the FFZ. Extensive experiments in almond orchards show that FLog reduces the link quality estimation error by 42.7% and improves gateway coverage estimation accuracy by 70.3%, compared with a widely used propagation model. The source codes and dataset are released at https://github.com/ycucm/Flog .

The rapid decline in groundwater around the world poses a significant challenge to sustainable agriculture. To address this issue, agricultural managed aquifer recharge (Ag-MAR) is proposed to recharge the aquifer by artificially flooding agricultural lands using surface water. Ag-MAR requires a carefully selected flooding schedule to avoid affecting the oxygen absorption of crop roots. However, current Ag-MAR scheduling does not take into account complex environmental factors such as weather and soil oxygen, resulting in crop damage and insufficient recharging amounts. This paper proposes MARLP, the first end-to-end data-driven control system for Ag-MAR. We first formulate Ag-MAR as an optimization problem. To that end, we analyze four-year in-field datasets, which reveal the multi-periodicity feature of the soil oxygen level trends and the opportunity to use external weather forecasts and flooding proposals as exogenous clues for soil oxygen prediction. Then, we design a two-stage forecasting framework. In the first stage, it extracts both the cross-variate dependency and the periodic patterns from historical data to conduct preliminary forecasting. In the second stage, it uses weather-soil and flooding-soil causality to facilitate an accurate prediction of soil oxygen levels. Finally, we conduct model predictive control (MPC) for Ag-MAR flooding. To address the challenge of large action spaces, we devise a heuristic planning module to reduce the number of flooding proposals to enable the search for optimal solutions. Real-world experiments show that MARLP reduces the oxygen deficit ratio by 86.8% while improving the recharging amount in unit time by 35.8%, compared with the previous four years.

In orchards, tree-level localization of robots is critical for smart agriculture applications like precision disease management and targeted nutrient dispensing. However, prior solutions cannot provide adequate accuracy. We develop our system, a fingerprinting-based localization system that can provide tree-level accuracy with only one LoRa gateway. We extract channel state information (CSI) measured over eight channels as the fingerprint. To avoid labor-intensive site surveys for building and updating the fingerprint database, we design a CSI Generative Model (CGM) that learns the relationship between CSIs and their corresponding locations. The CGM is fine-tuned using CSIs from static LoRa sensor nodes to build and update the fingerprint database. Extensive experiments in two orchards validate our system's effectiveness in achieving tree-level localization with minimal overhead and enhancing robot navigation accuracy.

This paper presents a novel map matching framework that adopts deep learning techniques to map a sequence of cell tower locations to a trajectory on a road network. Map matching is an essential pre-processing step for many applications, such as traffic optimization and human mobility analysis. However, most recent approaches are based on hidden Markov models (HMMs) or neural networks that are hard to consider high-order location information or heuristics observed from real driving scenarios. In this paper, we develop a deep reinforcement learning based map matching framework for cellular data, named as DMM, which adopts a recurrent neural network (RNN) coupled with a reinforcement learning scheme to identify the most-likely trajectory of roads given a sequence of cell towers. To transform DMM into a practical system, several challenges are addressed by developing a set of techniques, including spatial-aware representation of input cell tower sequences, an encoder-decoder based RNN network for map matching model with variable-length input and output, and a global heuristics-driven reinforcement learning based scheme for optimizing the parameters of the encoder-decoder map matching model. Extensive experiments on a large-scale anonymized cellular dataset reveal that DMM provides high map matching accuracy and fast inference time.

Both our experiments and previous studies show that LoRa links vary dynamically, which makes data transmission unreliable and consumes much energy of sensor nodes by retransmissions. This paper presents, a Rateless-enabled link Adaptation system for LoRa networks. Rateless coding approaches the optimal data rate of a link by continuously transmitting encoded data with an initial data rate. However, LoRa’s modulation, Chirp Spread Spectrum (CSS), introduces unique challenges to rateless-enabled transmissions. With CSS, Spreading Factor (SF) simultaneously determines the initial data rate and the concurrent transmissions of multiple links. This dual role of SF requires the co-design of coding and networking. We thus formulate an optimization problem for allocating network resources (SF, frequency channels, and transmission power) and coding parameters (block size and packet size) to all sensor nodes. A key component of this formulation is a rateless-aware network model that estimates the data transmission results of sensor nodes based on their link quality and transmission setting. Given that the optimization problem for obtaining the best transmission setting of all sensor nodes is NP-complete, a two-stage heuristic algorithm is designed. A Kalman filter-based link quality predictor is developed to capture the link quality variation. We implement on commodity LoRa hardware. Extensive experiments on a real testbed show that extends the lifetime of LoRaWAN by 66.1 %.

This article presents a novel system, LLDPC , 1 which brings Low-Density Parity-Check (LDPC) codes into Long Range (LoRa) networks to improve Forward Error Correction, a task currently managed by less efficient Hamming codes. Three challenges in achieving this are addressed: First, Chirp Spread Spectrum (CSS) modulation used by LoRa produces only hard demodulation outcomes, whereas LDPC decoding requires Log-Likelihood Ratios (LLR) for each bit. We solve this by developing a CSS-specific LLR extractor. Second, we improve LDPC decoding efficiency by using symbol-level information to fine-tune LLRs of error-prone bits. Finally, to minimize the decoding latency caused by the computationally heavy Soft Belief Propagation (SBP) algorithm typically used in LDPC decoding, we apply graph neural networks to accelerate the process. Our results show that LLDPC extends default LoRa’s lifetime by 86.7% and reduces SBP algorithm decoding latency by 58.09×.

LoRa networks have been deployed in many orchards for environmental monitoring and crop management. An accurate propagation model is essential for efficiently deploying a LoRa network in orchards, e.g., determining gateway coverage and sensor placement. Although some propagation models have been studied for LoRa networks, they are not suitable for orchard environments, because they do not consider the shadowing effect on wireless propagation caused by the ground and tree canopies. This paper presents FLog, a propagation model for LoRa signals in orchard environments. FLog leverages a unique feature of orchards, i.e., all trees have similar shapes and are planted regularly in space. We develop a 3D model of the orchards. Once we have the location of a sensor and a gateway, we know the mediums that the wireless signal traverse. Based on this knowledge, we generate the First Fresnel Zone (FFZ) between the sender and the receiver. The intrinsic path loss exponents (PLE) of all mediums can be combined into a classic Log-Normal Shadowing model in the FFZ. Extensive experiments in almond orchards show that FLog reduces the link quality estimation error by 42.7% and improves gateway coverage estimation accuracy by 70.3%, compared with a widely-used propagation model.

This paper aims to predict a set of apps a user will open on her mobile device in the next time slot. Such an information is essential for many smartphone operations, e.g., app pre-loading and content pre-caching, to improve user experience. However, it is hard to build an explicit model that accurately captures the complex environment context and predicts a set of apps at one time. This paper presents a deep reinforcement learning framework, named as DeepAPP, which learns a model-free predictive neural network from historical app usage data. Meanwhile, an online updating strategy is designed to adapt the predictive network to the time-varying app usage behavior. To transform DeepAPP into a practical deep reinforcement learning system, several challenges are addressed by developing a context representation method for complex contextual environment, a general agent for overcoming data sparsity and a lightweight personalized agent for minimizing the prediction time. Extensive experiments on a large-scale anonymized app usage dataset reveal that DeepAPP provides high accuracy (precision 70.6 percent and recall of 62.4 percent) and reduces the prediction time of the state-of-the-art by 6.58×. A field experiment of 29 participants demonstrates DeepAPP can effectively reduce launch time of apps.

This paper presents GeoDMA , which processes the GPS data from multiple vehicles to detect anomalous driving maneuvers, such as rapid acceleration, sudden braking, and rapid swerving. First, an unsupervised deep auto-encoder is designed to learn a set of unique features from the normal historical GPS data of all drivers. We consider the temporal dependency of the driving data for individual drivers and the spatial correlation among different drivers. Second, to incorporate the peer dependency of drivers in local regions, we develop a geographical partitioning algorithm to partition a city into several sub-regions to do the driving anomaly detection. Specifically, we extend the vehicle-vehicle dependency to road-road dependency and formulate the geographical partitioning problem into an optimization problem. The objective of the optimization problem is to maximize the dependency of roads within each sub-region and minimize the dependency of roads between any two different sub-regions. Finally, we train a specific driving anomaly detection model for each sub-region and perform in-situ updating of these models by incremental training. We implement GeoDMA in Pytorch and evaluate its performance using a large real-world GPS trajectories. The experiment results demonstrate that GeoDMA achieves up to 8.5% higher detection accuracy than the baseline methods.

Predicting the next application (app) a user will open is essential for improving the user experience, e.g., app pre-loading and app recommendation. Unlike previous solutions that only predict which app the user will open, this article predicts both the next app and the time to open it. Time prediction is essential to avoid loading the next app too early and consuming unnecessary resources on smartphones. To predict the next app and open time jointly, we model the app usage sequence as a marked temporal point process (MTPP), whose conditional intensity function can capture the probability of a new app usage event. We develop a novel data-driven MTPP-based app prediction system, named ATPP (App Temporal Point Process), which adopts a recurrent neural network architecture to learn the MTPP conditional intensity function for app prediction. ATPP adopts a set of techniques to incorporate the unique features of app prediction in our RNN architecture, including learning the correlated usage behavior of different apps by representation learning, the temporal dependency of app usage by an attention mechanism, and the location-related app usage behavior by feature extraction and fusion layer. We conduct extensive experiments on a large-scale anonymized app usage dataset to verify ATPP’s effectiveness.

Low-density parity-check (LDPC) codes have been widely used for Forward Error Correction (FEC) in wireless networks because they can approach the capacity of wireless links with lightweight encoding complexity. Although LoRa networks have been developed for many applications, they still adopt simple FEC codes, i.e., Hamming codes, which provide limited FEC capacity, causing unreliable data transmissions and high energy consumption of LoRa nodes. To close this gap, this paper develops LLDPC, which realizes LDPC coding in LoRa networks. Three challenges are addressed. 1) LoRa employs Chirp Spread Spectrum (CSS) modulation, which only provides hard demodulation results without soft information. However, LDPC requires the Log-Likelihood Ratio (LLR) of each received bit for decoding. We develop an LLR extractor for LoRa CSS. 2) Some erroneous bits may have high LLRs (i.e., wrongly confident in their correctness), significantly affecting the LDPC decoding efficiency. We use symbol-level information to fine-tune the LLRs of some bits to improve the LDPC decoding efficiency. 3) Soft Belief Propagation (SBP) is typically used as the LDPC decoding algorithm. It involves heavy iterative computation, resulting in a long decoding latency, which prevents the gateway from sending timely an acknowledgment. We take advantage of recent advances in graph neural networks for fast belief propagation in LDPC decoding. Extensive simulations on a large-scale synthetic dataset and in-filed experiments reveal that LLDPC can extend the lifetime of the default LoRa by 86.7% and reduce the decoding latency of the SBP algorithm by 58.09×.

Predicting the app that a user will open next is essential for improving user experience, e.g., app pre-loading. Unlike previous solutions that only predict next app’s ID, this work also predicts the time to open next app. Time prediction is important to avoid loading the next app too early, consuming too much memory and energy on smartphones. To predict next app’s ID and open time jointly, we model app usage as a marked temporal point process (MTPP), whose conditional intensity function can capture the probability of a new app usage event. We develop a novel data-driven MTPP-based app prediction system, named as ATPP (App Temporal Point Process), which adopts a recurrent neural network architecture to learn the MTPP conditional intensity function for app prediction. ATPP adopts a set of techniques to incorporate the unique features of app prediction in our RNN architecture, including learning the correlated usage behavior of different apps by representation learning, the temporal dependency of app usage events by an attention mechanism, and the location-related app usage behavior by a feature extraction and fusion layer. We conduct extensive experiments on a large-scale anonymized app usage dataset from 443 users over 21 days. The experiment results demonstrate that ATPP outperforms the state-of-the-art app prediction method by 6.0% in the prediction accuracy of next app ID and 2.09× reduction in the prediction error of next app open time. A field experiment of 22 users reveals that ATPP can reduce the app loading time by 78%.

This paper aims to predict the apps a user will open on her mobile device next. Such an information is essential for many smartphone operations, e.g., app pre-loading and content pre-caching, to save mobile energy. However, it is hard to build an explicit model that accurately depicts the affecting factors and their affecting mechanism of time-varying app usage behavior. This paper presents a deep reinforcement learning framework, named as DeepAPP, which learns a model-free predictive neural network from historical app usage data. Meanwhile, an online updating strategy is designed to adapt the predictive network to the time-varying app usage behavior. To transform DeepAPP into a practical deep reinforcement learning system, several challenges are addressed by developing a context representation method for complex contextual environment, a general agent for overcoming data sparsity and a lightweight personalized agent for minimizing the prediction time. Extensive experiments on a large-scale anonymized app usage dataset reveal that DeepAPP provides high accuracy (precision 70.6% and recall of 62.4%) and reduces the prediction time of the state-of-the-art by 6.58×. A field experiment of 29 participants also demonstrates DeepAPP can effectively reduce time of loading apps.