AgriSentinel: Crop Disease & Sentinel Monitoring

• AgriSentinel is a dual-purpose framework that combines a privacy-enhanced crop disease alerting system with a Sentinel-based agricultural monitoring architecture.
• It integrates differential privacy image obfuscation, lightweight CNN classification, and embedded language models to generate tailored agronomic recommendations.
• The broader architecture transforms Sentinel-derived multispectral time series into field-level inferences for crop mapping, compliance, and operational interventions.

Searching arXiv for the named topic and closely related supporting work. AgriSentinel is a term that can denote a class of agricultural monitoring systems centered on Sentinel satellite data and associated analytics, but it also names a specific “Privacy-Enhanced Embedded-LLM Crop Disease Alerting System” that combines differential-privacy image obfuscation, lightweight on-device disease classification, and an embedded LLM for recommendation generation (Mylay et al., 11 Sep 2025). Across the broader literature, the term is also a useful descriptor for operational workflows that integrate Sentinel-1, Sentinel-2, parcel data, machine learning, and downstream decision support for crop mapping, compliance monitoring, and agronomic intervention (Christiansen et al., 2018, Zhao et al., 2023). In this broader sense, AgriSentinel refers to a multi-scale, multi-modal agricultural intelligence stack in which Sentinel-derived time series are transformed into analysis-ready features, then into field-level or pixel-level inferences, and finally into actionable outputs such as disease alerts, crop maps, phenology indicators, weed-management maps, or inspection priorities (Drivas et al., 2022, Choumos et al., 2022).

1. Conceptual scope and definitions

In the most specific sense, AgriSentinel is the system introduced in “AgriSentinel: Privacy-Enhanced Embedded-LLM Crop Disease Alerting System,” described as the first Privacy-Enhanced Embedded-LLM Crop Disease Alerting System (Mylay et al., 11 Sep 2025). That system targets crop disease alerting on mobile or edge devices and addresses three coupled issues: protection of sensitive crop image data, usability for farmers under mobile constraints, and generation of specific management guidance rather than mere classification labels (Mylay et al., 11 Sep 2025).

In a broader remote-sensing and agricultural informatics sense, AgriSentinel can also denote an operational paradigm built around Sentinel missions for agricultural monitoring. The literature provides several such blueprints. One Sentinel-1-oriented example is Fieldbabel, a national-scale service for Denmark that preprocesses Sentinel-1 GRD data into georeferenced, speckle-filtered, terrain-corrected backscatter products, then packages them with LPIS field boundaries and crop information for agronomy researchers (Christiansen et al., 2018). A Sentinel-2-oriented example uses Sentinel-2 L2A imagery, vegetation indices, and both deep learning and pixel-based machine learning for automated lavender field segmentation as a precision-agriculture use case (Zhao et al., 2023). Another generalized architecture is the Agriculture monitoring Data Cube, which automates Sentinel-1 and Sentinel-2 ingestion, preprocessing, indexing, and SITS analytics for CAP monitoring (Drivas et al., 2022).

Taken together, these uses indicate two complementary meanings. The narrow meaning concerns a farmer-facing edge-AI disease alerting stack (Mylay et al., 11 Sep 2025). The broad meaning concerns a Sentinel-centered agricultural monitoring architecture that supports crop mapping, field-boundary extraction, phenology monitoring, management detection, and compliance workflows (Christiansen et al., 2018, Drivas et al., 2022). This suggests that “AgriSentinel” is best understood as both a named system and an architectural pattern.

2. Core architectural patterns

The embedded disease-alerting AgriSentinel has a compact on-device pipeline. A mobile client captures crop images, applies a Gaussian-noise differential privacy mechanism, passes the obfuscated image to a lightweight CNN classifier, and then conditions a fine-tuned GPT-2-based on-device LLM on the predicted disease and a curated knowledge pool to generate actionable recommendations (Mylay et al., 11 Sep 2025). The image obfuscation is defined as

$x_{\text{obf}} = x + N(0,\sigma^2),$

and the paper uses a privacy-loss proxy

$\epsilon = \frac{\Delta f}{\sigma},$

where $\Delta f$ is sensitivity and $\sigma$ is the Gaussian noise scale (Mylay et al., 11 Sep 2025).

The broader AgriSentinel pattern is typically modular and distributed. In the Sentinel-1 service architecture, raw GRD products are calibrated to $\sigma^0$, speckle-filtered with Lee Sigma, terrain-corrected with Range-Doppler Terrain Correction using SRTM 1Sec HGT, and delivered as GIS-ready layers, including VV, VH, and a VV(dB)/VH(dB) ratio visualization layer, plus LPIS vector overlays (Christiansen et al., 2018). In the Agriculture monitoring Data Cube, data discovery, ARD generation, COG conversion, indexing into Open Data Cube, and xarray-based SITS analytics are separated into an end-to-end cloud-resident workflow (Drivas et al., 2022).

A further architectural motif is parcel-centric data fusion. The “space-to-ground data availability” framework combines Sentinel-1 and Sentinel-2 parcel time series with Mapillary street-level imagery linked through LPIS parcel geometry, then trains separate models per modality and fuses them at decision level (Choumos et al., 2022). This suggests that an AgriSentinel deployment need not be satellite-only; rather, Sentinel data can serve as the backbone onto which ground imagery, UAV imagery, and expert knowledge are attached.

A plausible implication is that contemporary AgriSentinel systems are converging on a layered design: acquisition and preprocessing, geospatial indexing, parcel- or patch-level feature generation, modality-specific inference, and a decision-support or alerting layer. That general structure appears across disease alerting (Mylay et al., 11 Sep 2025), SAR-based field monitoring (Christiansen et al., 2018), crop mapping (Zhao et al., 2023), and CAP-oriented data-cube systems (Drivas et al., 2022).

3. Sentinel data foundations and preprocessing workflows

Sentinel-2 is repeatedly treated as a cornerstone sensor for AgriSentinel-style applications because it offers 10–20 m spatial resolution, approximately 5-day revisit with the constellation, and bands spanning visible, red-edge, NIR, and SWIR regions (Zhao et al., 2023). In the lavender-field study, the workflow uses Sentinel-2 L2A imagery with 12 spectral bands, excluding B1 and B10, together with NDVI and NDMI, yielding 14-channel inputs in some experiments (Zhao et al., 2023). NDVI and NDMI are defined as

$$\text{NDVI} = \frac{NIR - RED}{NIR + RED}, \qquad \text{NDMI} = \frac{NIR - SWIR}{NIR + SWIR}.$$

These indices are then stacked with the spectral bands (Zhao et al., 2023).

Sentinel-1-based AgriSentinel systems emphasize reproducibility and preprocessing standardization. Fieldbabel explicitly implements a SNAP graph consisting of read, radiometric calibration to $\sigma^0$, Lee Sigma speckle filtering with single look, $7 \times 7$ window size, sigma 0.9, target window $3 \times 3$, then terrain correction to 10 m output spacing (Christiansen et al., 2018). The products are then exported as GeoTIFFs and assembled into QGIS projects with LPIS parcel overlays (Christiansen et al., 2018).

At larger scale, the Agriculture monitoring Data Cube harmonizes Sentinel-1 and Sentinel-2 to 10 m, converts them to Cloud Optimized GeoTIFFs, and indexes them into a multidimensional cube with xarray access (Drivas et al., 2022). This enables parcel-level group-by operations after rasterizing parcel IDs and supports filtering, interpolation, smoothing, feature extraction, and downstream ML (Drivas et al., 2022). Such cube-based infrastructure is critical when AgriSentinel is expected to serve national-scale monitoring, not only bespoke field studies.

A separate line of work extends the Sentinel basis to crop suitability mapping. AgriPotential constructs a benchmark dataset from Sentinel-2 L2A imagery over Southern France, super-resolves ten bands to 5 m/px, normalizes reflectances to $[0,1]$, and aligns them with pixel-level agricultural-potential labels for three crop sectors across 11 monthly images in 2019 (Sakka et al., 13 Jun 2025). Although not itself named AgriSentinel, it is almost exactly the kind of data backbone that such a system would require for land-use planning and suitability inference (Sakka et al., 13 Jun 2025).

4. Analytical tasks and model families

AgriSentinel-style systems support several distinct classes of inference. One is crop mapping and field segmentation. The lavender Sentinel-2 study compares deep learning semantic segmentation with pixel-based machine learning for lavender-vs-background discrimination (Zhao et al., 2023). The tuned U-Net takes $\epsilon = \frac{\Delta f}{\sigma},$0 patches with 14 channels, uses four downsampling stages, skip connections, dropout, binary cross-entropy, learning rate 0.001, lecun_normal initialization, dropout 0.1, and L2 regularization 0.001, trained for 150 epochs with batch size 16 and early stopping patience 20 (Zhao et al., 2023). The fine-tuned model reaches Dice $\epsilon = \frac{\Delta f}{\sigma},$1 averaged over 10 runs on the test set, while a Random Forest baseline attains Dice $\epsilon = \frac{\Delta f}{\sigma},$2 in the model-comparison table (Zhao et al., 2023). This indicates that AgriSentinel need not be uniformly deep-learning-centric; pixel-based models remain competitive when multispectral information is rich.

Another major task is parcel-level time-series classification. In the space-to-ground dataset for Utrecht grassland monitoring, parcel time series from Sentinel-1 and Sentinel-2 are modeled with SVM, Random Forest, TempCNN, LSTM, and LSTM with attention (Choumos et al., 2022). Reported results show 95.22% accuracy and 89.96% F1 for TempCNN, and 95.20% accuracy and 90.05% F1 for LSTM with Attention, while InceptionV3 on street-level patches reaches 85% accuracy (Choumos et al., 2022). Late fusion yields marginal overall-accuracy improvement but notably increased confidence in decisions (Choumos et al., 2022). This supports a broader AgriSentinel principle: the value of extra modalities may lie as much in decision reliability as in average accuracy.

Temporal sequence modeling also appears in disease and stress detection. For broomrape-infested tomato fields, a Sentinel-2 time-series pipeline derives 12 bands, 20 vegetation indices, and five plant traits, aligns phenology using growing degree days, segments vegetation pixels, and trains a two-layer LSTM on $\epsilon = \frac{\Delta f}{\sigma},$3 pixels across 48 GDD time points (Narimani et al., 13 Sep 2025). The resulting model reaches 88 percent training accuracy and 87 percent test accuracy, with precision 0.86, recall 0.92, and F1 0.89 (Narimani et al., 13 Sep 2025). Permutation importance ranks NDMI, CCC, FAPAR, and a chlorophyll red-edge index highest, consistent with the physiology of infestation (Narimani et al., 13 Sep 2025). This suggests that AgriSentinel benefits from physiologically interpretable features, not only raw spectra.

Management detection is another emerging domain. A grazing-detection system built on Sentinel-2 L2A time series and a CNN-LSTM ensemble classifies Swedish pastures as grazed or not grazed and achieves an average F1 score of 77 percent across five validation splits, with 90 percent recall on grazed pastures (Pirinen et al., 16 Oct 2025). Operationally, when only 4 percent of sites can be visited, prioritizing fields predicted as non-grazed yields 17.2 times more confirmed non-grazing sites than random inspection (Pirinen et al., 16 Oct 2025). This provides a concrete AgriSentinel use case in which model utility is measured not only in predictive metrics but in inspection-resource efficiency.

Weed-management mapping in orchards illustrates the importance of sensor choice. Parcel-based RF, XGB, and KNN models trained on 4-month SITS from Sentinel-2 and PlanetScope indicate that PlanetScope clearly outperforms Sentinel-2, with the best PlanetScope RF achieving weighted F1 $\epsilon = \frac{\Delta f}{\sigma},$4, versus 0.49 for the best Sentinel-2 RF (Kontogiorgakis et al., 28 Apr 2025). Mowing and tillage are more separable than chemical-spraying and no practice (Kontogiorgakis et al., 28 Apr 2025). This suggests that coarse Sentinel data may suffice for some AgriSentinel tasks, whereas understory-sensitive management detection may need finer commercial imagery.

5. The embedded disease-alerting system named AgriSentinel

The named AgriSentinel system specializes in crop disease alerting under privacy and deployment constraints (Mylay et al., 11 Sep 2025). Its CNN classifier is deliberately shallow: input $\epsilon = \frac{\Delta f}{\sigma},$5, two Conv2D-ReLU-MaxPooling blocks with 32 and 64 filters, a 128-unit dense layer with ReLU and dropout, and a 3-class softmax output for Rice Blast, Brown Spot, and Bacterial Leaf Blight (Mylay et al., 11 Sep 2025). The model is trained on 1,200 rice disease images, roughly 400 per class, with 80/20 train/test split, Adam optimization, batch size 16, early stopping, learning-rate reduction on plateau, and categorical cross-entropy

$\epsilon = \frac{\Delta f}{\sigma},$6

(Mylay et al., 11 Sep 2025)

The privacy mechanism adds Gaussian noise directly to input images. The paper presents five obfuscation levels, from Very High obfuscation with $\epsilon = \frac{\Delta f}{\sigma},$7 to Very Low obfuscation with $\epsilon = \frac{\Delta f}{\sigma},$8 (Mylay et al., 11 Sep 2025). Classification performance degrades at strong obfuscation but approaches or exceeds baseline at medium to weak obfuscation. For Rice Blast, baseline accuracy is 0.84; Very High obfuscation yields 0.65, Medium yields 0.85, Low 0.88, and Very Low 0.91 (Mylay et al., 11 Sep 2025). For Brown Spot, baseline is 0.87 and Very Low obfuscation reaches 0.90 (Mylay et al., 11 Sep 2025). This indicates a usable privacy–utility operating region in which non-zero obfuscation does not meaningfully destroy predictive performance.

The LLM component, AgriLLM, is a GPT-2-based on-device model fine-tuned on a curated Q&A knowledge pool covering symptoms, environmental conditions, chemical treatments, and cultural practices (Mylay et al., 11 Sep 2025). At inference, the system conditions on the predicted disease and a relevant question to generate recommendations: $\epsilon = \frac{\Delta f}{\sigma},$9 (Mylay et al., 11 Sep 2025) Qualitative evaluation shows that fine-tuned AgriLLM produces disease-specific answers more closely aligned with the curated knowledge pool than untuned GPT-2, including recommendations such as azoxystrobin, flutolanil, or prothioconazole for Brown Spot and copper-based fungicides or antibiotics for Bacterial Leaf Blight (Mylay et al., 11 Sep 2025). This suggests that the system is not merely a classifier with text generation appended, but a small expert-system hybrid in which the LLM surfaces domain knowledge in natural language.

A common misconception would be to treat this AgriSentinel as a satellite remote-sensing platform. The specific 2025 system is not Sentinel-based and instead operates on mobile crop images (Mylay et al., 11 Sep 2025). However, the name overlaps with a broader Sentinel-centered monitoring paradigm. The two meanings are complementary rather than contradictory: the disease-alerting AgriSentinel can be seen as a farmer-edge endpoint, while Sentinel-based AgriSentinel architectures constitute upstream regional sensing backbones.

6. Operational significance, limitations, and future directions

AgriSentinel’s significance lies in connecting observation to action. In crop mapping, segmentation masks can be mosaicked back to full images and aligned closely with field boundaries, supporting field-boundary extraction and condition monitoring (Zhao et al., 2023). In Sentinel-1 services, field-level backscatter time series and LPIS overlays allow crop-type distinction, sampling design, and phenology tracking (Christiansen et al., 2018). In CAP-oriented systems, data cubes can generate parcel-level feature spaces and feed downstream compliance queries (Drivas et al., 2022). In disease alerting, the embedded system provides natural-language guidance rather than labels alone (Mylay et al., 11 Sep 2025).

At the same time, several limitations recur. Many studies remain geographically narrow or crop-specific. The lavender segmentation work uses 13 Sentinel-2 images focused on lavender bloom and a single binary task (Zhao et al., 2023). The space-to-ground dataset is restricted to Utrecht, 2017, and grassland-vs-non-grassland (Choumos et al., 2022). The broomrape detection study uses only ten tomato fields in California and field-level infestation labels derived from farmer reports (Narimani et al., 13 Sep 2025). The disease-alerting AgriSentinel is limited to three rice diseases and uses a DP-inspired privacy metric rather than a full $\Delta f$0-DP training guarantee (Mylay et al., 11 Sep 2025). These constraints suggest that current AgriSentinel implementations are proofs of capability rather than finished universal platforms.

Another recurring challenge is generalization. Models trained in one region, season, or management context may not transfer cleanly without retraining or domain adaptation (Zhao et al., 2023, Sakka et al., 13 Jun 2025). Weather, cloud cover, and phenological shifts remain major confounders, and in optical settings, cloud masking and gap filling are often only partially addressed (Zhao et al., 2023, Narimani et al., 13 Sep 2025). This suggests that mature AgriSentinel systems will likely require multi-temporal and multi-sensor fusion, especially combining Sentinel-2 with Sentinel-1 for cloud robustness and structural sensitivity (Christiansen et al., 2018, Drivas et al., 2022).

Future directions are relatively consistent across the literature. They include transformer-based or hybrid CNN–Transformer models for remote sensing (Zhao et al., 2023), more sophisticated multimodal fusion beyond late fusion (Choumos et al., 2022), integration of Sentinel-1 with Sentinel-2 and UAV or street-level imagery (Choumos et al., 2022), multi-temporal suitability modeling and ordinal prediction (Sakka et al., 13 Jun 2025), and broader operational pipelines in which satellite screening schedules high-resolution interventions, as in satellite-to-drone weed management (Bansal et al., 2024). A plausible implication is that the next generation of AgriSentinel will be multi-layered: Sentinel data for coarse surveillance, local sensors or mobile imaging for confirmation, and language-model interfaces for explanation and recommendation.

In sum, AgriSentinel names both a concrete privacy-enhanced edge disease-alerting system (Mylay et al., 11 Sep 2025) and a larger family of Sentinel-driven agricultural intelligence architectures (Christiansen et al., 2018, Zhao et al., 2023, Drivas et al., 2022). As an encyclopedia topic, it is best situated at the intersection of Earth observation, edge AI, agronomic decision support, and compliance monitoring. Its defining characteristic is not any single model, sensor, or interface, but the systematic transformation of agricultural observations into actionable, operational intelligence.