Modern precision agronomy depends on data. While IoT devices (Khanna and Kaur, 2019; Farooq et al., 2020) facilitate continuous monitoring and UAV-based hyperspectral sensors (Adão et al., 2017) allow for detailed field mapping, they have limitations. As noted by Mahlein (2016), conventional RGB and visual imaging techniques identify disease only after physical changes appear. As noted in Section 1, a primary limitation of conventional methods is the reliance on lagging indicators, where visual or environmental signals typically arrive only after stress has initiated. The opportunity for early action lies in detecting pre-symptomatic physiological changes. Electronic-nose technologies have been tested (Wilson, 2013; Wang et al., 2024), but they often lack the field-ready capability to distinguish specific stress signals within complex chemical mixtures, a challenge the proposed QDSS model seeks to address within the simulation. High humidity can interfere with sensor readings, and sensor drift, where the baseline reading shifts over time due to contamination or aging, is a major problem.

Recent work integrating UAV-mounted spectroscopy and FBG sensors (Sliti et al., 2025) proposed a framework to enhance forest monitoring by fusing strain metrics with spectral data. However, while they rely on hydraulic and spectral markers for early detection, this work targets volatile organic compounds (VOCs) to identify specific biochemical signatures. Additionally, recent work on reconfigurable holographic surfaces (Li et al., 2024) demonstrates that transceiver hardware impairments limit spectral efficiency. This limitation supports the QAL strategy of processing data at the Edge Layer, reducing reliance on raw data transmission that is susceptible to signal degradation.

Deep learning has gained significant momentum in agriculture (Kamilaris and Prenafeta-Boldú, 2018), but most deep learning models use centralized data, raising privacy concerns for farmers. Federated Learning (FL) (McMahan et al., 2017; Wen et al., 2023) addresses data privacy and accessibility challenges. However, AI performance depends on input data. A key limitation in current agricultural AI is data quality. Current advanced multi-modal networks, such as the DeepMMCropYNet proposed by Pramela (2025), have achieved high accuracy by fusing static and temporal features. However, the authors note that the model remains limited by overlapping data across feature, temporal, and spatial dimensions, which makes it difficult to learn distinct crop patterns. The FSPM is designed to address this by combining the FL structure with modeled high-fidelity, biochemical input data that offers greater feature separability.

Blockchain is often suggested for food traceability (Kamilaris et al., 2019; Lin et al., 2019). While moving from energy-heavy Proof-of-Work (PoW) (De Vries, 2018) to Proof-of-Stake (PoS) (Saleh, 2021) helps sustainability, the oracle problem remains (Pasdar et al., 2023): transferring real-world data onto the blockchain. Conventional systems often function as passive logs dependent on manual entry, which can be falsified. They record claims but cannot self-verify them. The dPoS-SS mechanism is designed to address this by integrating signed sensor data into the verification process, turning the ledger into an active verification tool.

Table 4 benchmarks the proposed system against current commercial platforms, highlighting the shift toward preemptive diagnostics.

This layer handles in-field data collection. In the proposed physical architecture, it consists of autonomous, low-power IoT nodes that monitor crop health. A key part is the QDSS Array, designed to detect specific VOCs and soil nutrients. This creates a spectral fingerprint of plant stress for early detection. The nodes also use standard sensors for soil moisture, temperature, humidity, and macronutrients (NPK) to provide environmental context. To ensure data authenticity, each node is embedded with a unique cryptographic key pair for hardware-level digital signing of all transmitted data.

The Edge Layer is the on-farm computational hub. It connects sensors to the cloud, handling real-time preprocessing and privacy. The Edge Gateway Device gathers data from the Perception Layer. By processing raw data into a refined, compressed, and encrypted format locally, it reduces bandwidth usage. Crucially, this device runs the local FL client, keeping raw farm data on-site.

The Platform Layer is the cloud backend for analysis. The FL Server manages the learning process. It sends a global model to Edge Gateways and collects only model updates (not raw data) to improve the global system. External Oracle Services provide independent data, like weather reports, to validate blockchain claims. To prevent manipulation, data is aggregated from multiple independent providers before being finalized on the ledger.

While the QDSS provides the granular data necessary for detection, and FL preserves privacy, a trust mechanism is required to ensure that this self-reported data is not spoofed for financial gain. To address this, the Blockchain Layer acts as the immutable backbone of the QAL framework. It uses the PoSP consensus mechanism (dPoS-SS) to record verified sustainable actions. These actions trigger Smart Contracts that issue rewards upon validation. Validators verify claims by comparing farm data against trusted oracle data, using a voting process designed to maximize Byzantine Fault Tolerance (BFT) while minimizing computational overhead.

The central sensing tool is the QDSS. For this study, the QDSS is a theoretical model based on the physical properties of quantum dots, validated through simulation rather than physical prototyping.

A quantum dot (QD) is a semiconductor nanocrystal with properties governed by quantum confinement. This allows the dot’s color to be tuned by changing its size, creating an array where each sensor targets different spectral lines. The sensor operates via fluorescence quenching, as illustrated in Figure 2. QDs have a baseline glow. When target molecules (VOCs) bind to them, energy is diverted, dimming the glow. This quenching effect is proportional to the target molecule’s concentration.

This study uses a theoretical sensor model based on the established properties of quantum dots found in literature, acknowledging that physical implementation would face additional complexities not fully captured in a synthetic environment. Sol-gel encapsulation of polymer-coated ZnS/CdSe quantum dots has detected organic vapors (Hasani et al., 2010), and CdSe/ZnS crystals have detected phenolic compounds (Rahman et al., 2020).

The quenching effect is modeled using a modified Stern-Volmer equation (Equation 1). This generates a spectral fingerprint from a mixture of analytes ([Q_i]).

Limitations of this model must be noted. While the standard Stern-Volmer equation predicts a linear response, real-world sensor arrays exhibit saturation at high analyte concentrations. To account for this, the simulation introduces a hyperbolic tangent nonlinearity (Equation 8) to rigorously stress-test the FSPM’s ability to disentangle complex, saturated VOC signals. Creating a physical QDSS faces challenges such as stability in UV light and interference from humidity. The simulation assumes ideal encapsulation but introduces noise to model these physical imperfections.

Table 5 lists the proposed array composition. While this selection lists the 8 primary quantum dot materials selected for their specific affinity to stress markers, the full sensor model generates a high-dimensional feature vector. This is achieved by hyperspectral sampling, where each of the 8 dots is sampled across 12 discrete spectral sub-bands to capture subtle shifts in fluorescence, resulting in a total feature vector of 96 dimensions.

Theoretical arrays allow for the simulation of unique spectral fingerprints for VOC mixtures, as shown in Figure 3.

In the proposed physical architecture, raw data is processed into a clean format using Algorithm 1 to handle outliers and drift. However, for this simulation study, this explicit calibration step was bypassed to rigorously stress-test the FSPM’s ability to learn directly from raw, noisy, simulated sensor data. To mitigate drift, a dynamic baseline is kept using an Exponentially Weighted Moving Average (EWMA), as shown in Equation 2. This protects the baseline from spikes.

The FSPM provides predictive capability, trained via FL to ensure privacy (Algorithm 2). In this protocol, K represents the total number of clients, and C represents the fraction of clients selected per round. It uses a feed-forward neural network with two hidden layers, Batch Normalization, and Dropout. A composite loss function (Equation 3) balances yield prediction and stress classification using a weight term, λ_t. Weights (W) are adjusted during local training using Adam optimization, with regularization enforced via the network’s Dropout layers (p = 0.5). λ_t was set to 0.5 for equal task importance.

The dPoS-SS consensus mechanism verifies and rewards sustainable actions (Algorithm 3).

The following algorithms conceptualize the decision-support logic used by the QAL framework to process the simulated data streams.

Anomaly Detection Algorithm 4 outlines the logic for the proposed early warning system, which is designed to calculate a Mahalanobis distance anomaly score. It triggers an alert only if the score stays high for a set window.

Nutrient Recommendation Algorithm 5 calculates a fertilizer blend to fix deficiencies without causing secondary imbalances, using weather data to avoid runoff. For the purpose of the simulation study, this algorithm was instantiated with a linear nitrogen depletion rate of 1.2 kg/ha/day and a critical stress threshold of 50 kg/ha, triggering intervention when deficiency probability exceeds 50%.

Adaptive Thresholding Algorithm 6 presents the theoretical logic for adaptive thresholding envisioned for the physical Edge Layer. While the simulation in this study strictly evaluates the FSPM (as detailed in Section 6), this algorithmic concept is proposed to address the requirement for standalone operation during potential network interruptions in real-world deployment.

A dataset of 20,000 data points was generated to model a growing season. To simulate the complexity of real-world sensing, a custom stochastic environmental noise function was applied to all sensor readings. This function introduces three types of errors common in field sensors:

To simulate a decentralized network, the data was partitioned among 20 simulated clients (farms). The data was split so that 80% was used for training and 20% for testing. To reflect the privacy-preserving nature of the system, each client trained the model locally on its own data partition for 5 epochs before sending only the weight updates to the central server.

Clients processed time-series data to generate temporal features, expanding the 96 spectral bands and conventional sensors into a high-dimensional feature vector. As detailed in Table 6, statistical moments (mean and standard deviation) were calculated across multiple rolling windows for every sensor channel to capture temporal dynamics.

The FSPM is a feed-forward neural network designed for efficiency on edge devices. It consists of two hidden layers with Batch Normalization (to stabilize learning) and Dropout (to prevent the model from memorizing the data). The hyperparameters, listed in Table 7, were optimized to balance learning speed and accuracy.

The simulation was repeated 10 independent times with different random seeds to ensuring that results were consistent and reproducible.

The following evaluation scenarios and methodological constraints were defined to validate the system:

To verify if the complex QDSS sensor is actually necessary, the QAL system was compared against a Conventional setup that used only standard temperature and humidity sensors. The results (Table 8, Figure 8) are significant.

When using only standard sensors, the accuracy was poor (≈ 33%), which yields performance comparable to random chance. This confirms that temperature and humidity are lagging indicators—they change too late to predict stress types accurately. However, when the QDSS (chemical) data was added, accuracy significantly improved across all models, peaking at 96.74% with the FSPM. Notably, while chemical data alone improved Random Forest performance to 65%, the combination of high-fidelity spectral data with the deep learning architecture was required to achieve near-perfect classification, suggesting that detecting VOCs is a necessary but not solely sufficient condition for early diagnosis; advanced non-linear modeling appears to be required.

Real-world farm environments are noisy—sensors get compromised, and weather fluctuates. To evaluate the model’s susceptibility to such conditions, a Noise Sensitivity Analysis was run. The sensor data was intentionally perturbed with increasing levels of random noise (σ).

As shown in Figure 9, the model demonstrates resilience at lower noise thresholds. The classification accuracy (orange line) remains high (> 90%) when noise levels are ≤ 0.10. However, performance degradation accelerates beyond this point, dropping to approximately 71% at σ = 0.20. This indicates that while the system can handle standard sensor drift, it requires maintenance protocols to prevent noise from reaching these critical instability thresholds. This suggests that the framework maintains theoretical efficacy even under simulated non-ideal conditions, though physical validation remains necessary to confirm these noise-tolerance thresholds.

Transitioning from a computational framework to a physical reality presents several challenges that must be addressed.