AgroSense: Precision Agriculture Systems

• AgroSense is a framework that integrates sensing hardware, embedded processing, wireless communication, and cloud analytics to enable precision agriculture.
• It leverages diverse modalities such as insect monitoring, soil moisture and chemical analysis, and multimedia imaging to support targeted agronomic interventions.
• The system architecture adapts to constrained connectivity with delay-tolerant networking and low-power modes, ensuring reliable performance in remote fields.

AgroSense is a label used in the cited literature for several closely related precision-agriculture architectures that couple in-situ sensing, embedded or edge computation, wireless communications, cloud data infrastructure, and decision support. Across these works, AgroSense denotes systems for distributed insect monitoring based on dual-wavelength near-infrared backscatter, wireless multimedia sensor networks for disease and insect detection, plug-and-play electrochemical salicylic acid sensing, delay-tolerant participatory agro-advisory, IoT irrigation control, smartphone-based soil chemical analysis, multimodal crop recommendation, and joint sensing-and-communication RF nodes for soil moisture measurement (Rydhmer et al., 2021, Shi et al., 2018, Kashyap et al., 2021, Gupta et al., 2016, Hasib et al., 22 Jan 2026, Silva et al., 2022, Pandey et al., 1 Sep 2025, Raza et al., 2024).

1. Architectural scope and system composition

A recurring architectural pattern in AgroSense systems is the decomposition into sensing hardware, local processing, communications, storage or analytics, and downstream alerting. The MiLive Wireless Multimedia Sensor Network (WMSN) platform instantiates this as a physically co-located scalar node (iLive), multimedia node (MWiFi), and Power Management Unit (PMU), with heterogeneous buses including I²C, SPI, ADC, and GPIO for agronomic sensors, plus dual-radio operation using IEEE 802.15.4 for low-rate wake-up and control and IEEE 802.11 for high-rate image or video transfer (Shi et al., 2018). The insect-monitoring implementation uses a solar-powered, weather-sealed field module that records insect events, appends metadata including timestamp, sensor ID, GPS, temperature, humidity, light, and camera snapshot URL, and pushes the resulting records over GSM via AMQP to a central broker (Rydhmer et al., 2021). The irrigation-oriented system centers on an ESP32-WROOM-32 with DHT22, HC-SR04, capacitive soil-moisture probe, SSD1306 OLED, buzzer, RGB LED, relay, and pump, while the salicylic-acid reader combines a screen-printed three-electrode strip with a TI C2000 LaunchPad and a dedicated portable analog front end (Hasib et al., 22 Jan 2026, Kashyap et al., 2021).

The data back ends described under the AgroSense label are similarly heterogeneous but structurally comparable. One relational schema stores InsectEvent, SensorStatus, and TrapComparison tables for event-level entomological telemetry (Rydhmer et al., 2021). Another pipeline posts per-field JSON payloads to ThingSpeak and exposes live gauges, historical charts, alert logs, manual override controls, and CSV export (Hasib et al., 22 Jan 2026). The mobile soil-analysis platform stores JSON records in Cloudant and routes them through Node-RED/API Connect to dashboards and GIS applications (Silva et al., 2022). In the DTN-based rural advisory design, farmer queries, acknowledgments, and responses are represented as multimedia event objects keyed by (UID, EID) and propagated between smartphones, relay nodes, and servers (Gupta et al., 2016).

Taken together, these sources indicate that AgroSense is not a single standardized implementation but a systems pattern for precision agriculture in which sensing modalities, communications strategies, and analytics layers are selected according to deployment constraints. This suggests that the term is best understood as an integrative framework spanning field instrumentation, networking, and agronomic decision support rather than a single device class.

2. Sensing modalities and measurement principles

One major AgroSense modality is unsupervised insect monitoring using backscattered near-infrared modulation signatures. The reported sensor uses dual IR LED bands at $\lambda_1 = 808\,\mathrm{nm}$ and $\lambda_2 = 980\,\mathrm{nm}$, with total optical powers of approximately $1.6\,\mathrm{W}$ and $1.7\,\mathrm{W}$, respectively. The emitters are arranged in a checkerboard on an $82\,\mathrm{cm}^2$ board and modulated at $f_1 = 118.8\,\mathrm{kHz}$ and $f_2 = 79.2\,\mathrm{kHz}$ for synchronous lock-in detection. Backscattered light is focused by an aspheric NIR-coated lens onto a $1\,\mathrm{cm}^2$ silicon quadrant photodiode, and the beam–field-of-view overlap defines a roughly conical measurement volume whose geometry can be tuned by the angle $O_s$ between emitter and receiver. The measurement model is expressed as

$$I_{\mathrm{scat}}(r,\lambda) \approx I_{\mathrm{emit}}(\lambda)\cdot \sigma(\lambda)\cdot G(r)\cdot T_{\mathrm{opt}}/(4\pi r^2),$$

with the detection edge set where signal-to-noise ratio exceeds $\lambda_2 = 980\,\mathrm{nm}$0 for a $\lambda_2 = 980\,\mathrm{nm}$1 target. Extracted features include wing beat harmonics, melanisation, and flight direction (Rydhmer et al., 2021).

A second sensing family targets soil moisture and soil dielectric properties. The ultra-compact soil moisture sensor (UCSMS) uses a multiturn complementary spiral resonator etched into the ground plane of a microstrip transmission line on FR-4. Three variants are reported: a 3-CSR operating at approximately $\lambda_2 = 980\,\mathrm{nm}$2, a 4-CSR at approximately $\lambda_2 = 980\,\mathrm{nm}$3, and a 5-CSR at approximately $\lambda_2 = 980\,\mathrm{nm}$4. Resonance is modeled by

$\lambda_2 = 980\,\mathrm{nm}$5

and the system maps resonant-frequency shifts to permittivity and then to volumetric water content. The reported sensitivity is $\lambda_2 = 980\,\mathrm{nm}$6 per unit $\lambda_2 = 980\,\mathrm{nm}$7, the permittivity measurement span is approximately $\lambda_2 = 980\,\mathrm{nm}$8, and the sensor is described as insensitive to variations in the Volume Under Test because $\lambda_2 = 980\,\mathrm{nm}$9 varies by less than $1.6\,\mathrm{W}$0 for soil heights from $1.6\,\mathrm{W}$1 to $1.6\,\mathrm{W}$2 (Raza et al., 2024). A lower-cost irrigation implementation instead uses a capacitive soil-moisture probe and a linear calibration

$1.6\,\mathrm{W}$3

valid for $1.6\,\mathrm{W}$4 and yielding $1.6\,\mathrm{W}$5 (Hasib et al., 22 Jan 2026).

AgroSense also encompasses chemical and biochemical crop-state measurements. The salicylic-acid platform is bio-agent-free and relies on direct electro-oxidation of salicylic acid on a commercial screen-printed three-electrode strip with planar carbon working electrode, carbon counter electrode, and Ag/AgCl reference electrode. The simplified reaction is

$1.6\,\mathrm{W}$6

Differential-Pulse Voltammetry is applied with $1.6\,\mathrm{W}$7 pulses of width $1.6\,\mathrm{W}$8 on a staircase of $1.6\,\mathrm{W}$9 per step and a $1.7\,\mathrm{W}$0 period, producing a peak near $1.7\,\mathrm{W}$1 vs Ag/AgCl. Concentration is estimated through

$1.7\,\mathrm{W}$2

with experimentally reported values $1.7\,\mathrm{W}$3 and $1.7\,\mathrm{W}$4, and the limit of detection is given by $1.7\,\mathrm{W}$5 (Kashyap et al., 2021).

Soil chemical analysis is represented by the smartphone-read microfluidic paper analytical device for pH mapping. This two-layer vertical $1.7\,\mathrm{W}$6PAD is fabricated on Whatman CHR1 chromatography paper, loaded with Bromocresol Green and Bromocresol Purple, and packaged with five reference color patches and a QR code carrying lot, device ID, and indicator type. The platform classifies soil pH into low $1.7\,\mathrm{W}$7–$1.7\,\mathrm{W}$8, medium $1.7\,\mathrm{W}$9–$82\,\mathrm{cm}^2$0, and high $82\,\mathrm{cm}^2$1–$82\,\mathrm{cm}^2$2 classes. Illumination correction is performed as

$82\,\mathrm{cm}^2$3

after which corrected color features are used for classification (Silva et al., 2022).

Finally, the multimedia WMSN implementation shows that AgroSense is not limited to scalar probes. MiLive combines four Watermark soil-moisture probes, three Decagon soil-moisture probes, air temperature, soil temperature, air humidity, light sensing, and a USB or CSI CCD camera with optional microphone, thereby treating visual data as a first-class agronomic signal rather than a peripheral add-on (Shi et al., 2018).

3. Embedded processing, feature extraction, and inference

In the insect-monitoring design, signal processing is performed in a multi-stage embedded chain. Each quadrant photodiode current passes through a transimpedance amplifier with bandwidth $82\,\mathrm{cm}^2$4–$82\,\mathrm{cm}^2$5 and gain $82\,\mathrm{cm}^2$6. Four 14-bit ADCs sample at $82\,\mathrm{cm}^2$7, and an FPGA implements eight digital lock-in amplifiers, $82\,\mathrm{cm}^2$8 low-pass filters, and down-sampling to a $82\,\mathrm{cm}^2$9, 16-bit stream. Event detection operates on 10 min blocks with a rolling 2 s median baseline and rolling standard deviation $f_1 = 118.8\,\mathrm{kHz}$0, flags samples satisfying $f_1 = 118.8\,\mathrm{kHz}$1, applies morphological erosion of $f_1 = 118.8\,\mathrm{kHz}$2 and dilation of $f_1 = 118.8\,\mathrm{kHz}$3, and forms event candidates by taking the union across all eight channels. For each event, the MCU computes a Welch power spectral density; the wingbeat fundamental $f_1 = 118.8\,\mathrm{kHz}$4 is the largest peak in the $f_1 = 118.8\,\mathrm{kHz}$5–$f_1 = 118.8\,\mathrm{kHz}$6 band, harmonics are evaluated at $f_1 = 118.8\,\mathrm{kHz}$7 for $f_1 = 118.8\,\mathrm{kHz}$8, and additional descriptors include the harmonic index

$f_1 = 118.8\,\mathrm{kHz}$9

the body-to-wing ratio $f_2 = 79.2\,\mathrm{kHz}$0, the melanisation contrast $f_2 = 79.2\,\mathrm{kHz}$1, and a flight-direction angle $f_2 = 79.2\,\mathrm{kHz}$2 derived from quadrant intensities (Rydhmer et al., 2021).

In the MiLive platform, local multimedia analytics are intentionally lightweight. The reported image-processing pipeline consists of white-balance and demosaicing, color-histogram segmentation with thresholds on the R/G channels to highlight diseased areas, Sobel edge detection using

$f_2 = 79.2\,\mathrm{kHz}$3

blob extraction with area and circularity descriptors, and final classification by lightweight SVM or rule-based logic that returns a confidence score $f_2 = 79.2\,\mathrm{kHz}$4. High-power image capture is triggered only when scalar thresholds such as $f_2 = 79.2\,\mathrm{kHz}$5light, $f_2 = 79.2\,\mathrm{kHz}$6, or $f_2 = 79.2\,\mathrm{kHz}$7humidity exceed predefined values (Shi et al., 2018).

The mobile soil-analysis platform uses classical machine learning rather than deep CNNs. On-device preprocessing in OpenCV locates the $f_2 = 79.2\,\mathrm{kHz}$8PAD region via fiducials, segments circular regions of interest with Hough-circle detection, extracts mean RGB values, and corrects them using onboard reference patches. Two binary classifiers—one for Bromocresol Green and one for Bromocresol Purple—are trained with 80% of the calibration dataset $f_2 = 79.2\,\mathrm{kHz}$9 and 5-fold cross-validation. Logistic regression and linear SVM are used, hyperparameters are selected by grid search, mean validation accuracy exceeds 99% in lab, and inference runs in approximately 100 ms on the phone CPU. Model parameters are retrieved at runtime as JSON from a lightweight CouchDB API (Silva et al., 2022).

By contrast, the crop-recommendation system titled AgroSense is an explicitly multimodal deep-learning framework. Its Soil Classification Module evaluates ResNet-18, EfficientNet-B0, and Vision Transformer (ViT-Base/16) on $1\,\mathrm{cm}^2$0 images, while the Crop Recommendation Module consumes normalized nutrient vectors such as $1\,\mathrm{cm}^2$1 using MLP, XGBoost, LightGBM, and TabNet. Fusion is implemented through

$1\,\mathrm{cm}^2$2

after which the fused representation is passed to a selected classifier. The same source reports that EfficientNet-B0 and ResNet-18 support quantization or pruning for resource-constrained inference and that the fusion layer with LightGBM incurs inference latency below $1\,\mathrm{cm}^2$3 on modern smartphones (Pandey et al., 1 Sep 2025).

4. Communications, cloud infrastructure, and operation under constrained connectivity

AgroSense communications architectures vary with bandwidth, energy, and coverage assumptions. The insect-monitoring modules package up to hundreds of insect mini-records per 10 min, append metadata, and push the resulting payloads via GSM modem over AMQP to a central broker. Modules self-register on the network, report heartbeats, and buffer data in local flash when connectivity is lost. The irrigation implementation uses Wi-Fi on the ESP32 to send HTTP POST messages to ThingSpeak every 60 s while respecting the 15 s per-field rate limit, and the salicylic-acid deployment plan describes BLE links to smartphones or LoRaWAN/NB-IoT gateways for long-range uplink (Rydhmer et al., 2021, Hasib et al., 22 Jan 2026, Kashyap et al., 2021).

MiLive adopts a dual-radio strategy with explicit energy-aware power gating. IEEE 802.15.4 carries low-rate scalar data and wake-up signaling, while IEEE 802.11b/g is activated only when multimedia data are requested. Babel provides mesh routing over 802.11 using a combination of proactive and reactive distance-vector techniques designed to avoid loops and black holes. PMU modes include Sleep with NanoRisc only at $1\,\mathrm{cm}^2$4, SWSN for scalar-only operation, WMSN for multimedia-only operation, and SWMSN when both subsystems are active. The aggregate energy model is written

$1\,\mathrm{cm}^2$5

and battery life is estimated by

$1\,\mathrm{cm}^2$6

An example scalar-only duty cycle with sampling every 10 min yields a projected lifetime exceeding 5 years on two AA cells (Shi et al., 2018).

Where conventional connectivity is unreliable, AgroSense includes a delay-tolerant networking layer. In the DTN-RuralSense design, the end-to-end chain is Farmer SH ↔ Relay Node ↔ Expert System, with farmer smartphones storing multimedia event objects in a local SQLite queue until acknowledged by the server. Relay nodes create WPA2-PSK Wi-Fi hotspots or Bluetooth beacons, aggregate events from up to 10 nearby smartphones, store them in a FIFO buffer of capacity approximately 100 events, and batch-upload them once cellular or Wi-Fi coverage becomes available. End-to-end delay is modeled as

$1\,\mathrm{cm}^2$7

with $1\,\mathrm{cm}^2$8 hops, and expected delivery probability within deadline $1\,\mathrm{cm}^2$9 is approximated by

$O_s$0

This design targets a tolerable advisory delay of 24 h in no-network and poor-network zones (Gupta et al., 2016).

Cloud integration in the mobile soil-analysis platform emphasizes offline robustness and remote calibration. QR-scan and image capture produce a JSON record containing raw image data, RGB values, classes, and geolocation; records are stored locally or posted via HTTPS to Cloudant; Node-RED/API Connect routes data to storage and GIS dashboards; and over-the-air updates are implemented by updating the model JSON in Cloudant. The architecture explicitly supports offline operation through queue-and-sync behavior (Silva et al., 2022).

5. Reported performance and validation results

Field validation across AgroSense implementations emphasizes agreement with conventional methods, high temporal or spatial resolution, and usable network performance. Insect monitoring validated against six yellow water traps over four weeks yielded a Spearman’s rank correlation coefficient of $O_s$1 with $O_s$2; sensors recorded approximately 19 times more insect observations, specifically $O_s$3 events·sensor$O_s$4·day$O_s$5 versus $O_s$6 insects·trap$O_s$7·day$O_s$8, and resolved intra-day and multi-day flight peaks such as a Brassicogethes aeneus swarm from May 7–11 that were not visible in once-daily trap collection (Rydhmer et al., 2021). In a 10-node MiLive deployment over a 1 ha vineyard with 25–40 m spacing, Babel mesh routing maintained at least 95% packet delivery ratio; plant-disease detection reached 94% correct with 3% false positives and 6% false negatives; insect-motion detection reached 92% correct with 4% false positives and 8% false negatives (Shi et al., 2018). In DTN-RuralSense laboratory and early field trials, mean end-to-end delay was approximately 3.2 h with standard deviation approximately 1.5 h, the 90th percentile delay was approximately 8 h, and delivery ratio within a 24 h deadline was approximately 0.99, whereas direct cellular alone yielded less than 30% successful uploads in 24 h (Gupta et al., 2016).

Chemical and environmental sensing subsystems report similarly task-specific validation. The salicylic-acid sensor exhibits a linear range of $O_s$9 to $$I_{\mathrm{scat}}(r,\lambda) \approx I_{\mathrm{emit}}(\lambda)\cdot \sigma(\lambda)\cdot G(r)\cdot T_{\mathrm{opt}}/(4\pi r^2),$$0 with $$I_{\mathrm{scat}}(r,\lambda) \approx I_{\mathrm{emit}}(\lambda)\cdot \sigma(\lambda)\cdot G(r)\cdot T_{\mathrm{opt}}/(4\pi r^2),$$1, sensitivity of $$I_{\mathrm{scat}}(r,\lambda) \approx I_{\mathrm{emit}}(\lambda)\cdot \sigma(\lambda)\cdot G(r)\cdot T_{\mathrm{opt}}/(4\pi r^2),$$2, LOD of $$I_{\mathrm{scat}}(r,\lambda) \approx I_{\mathrm{emit}}(\lambda)\cdot \sigma(\lambda)\cdot G(r)\cdot T_{\mathrm{opt}}/(4\pi r^2),$$3, response time below 1 s per DPV scan, stability below 5% drift over 1 h for repeated measurements on the same strip, and inter-strip reproducibility of approximately 8% RSD. In plant juices, measured values were $$I_{\mathrm{scat}}(r,\lambda) \approx I_{\mathrm{emit}}(\lambda)\cdot \sigma(\lambda)\cdot G(r)\cdot T_{\mathrm{opt}}/(4\pi r^2),$$4 for orange versus $$I_{\mathrm{scat}}(r,\lambda) \approx I_{\mathrm{emit}}(\lambda)\cdot \sigma(\lambda)\cdot G(r)\cdot T_{\mathrm{opt}}/(4\pi r^2),$$5 by a commercial UV-enzyme kit, and $$I_{\mathrm{scat}}(r,\lambda) \approx I_{\mathrm{emit}}(\lambda)\cdot \sigma(\lambda)\cdot G(r)\cdot T_{\mathrm{opt}}/(4\pi r^2),$$6 for tomato versus $$I_{\mathrm{scat}}(r,\lambda) \approx I_{\mathrm{emit}}(\lambda)\cdot \sigma(\lambda)\cdot G(r)\cdot T_{\mathrm{opt}}/(4\pi r^2),$$7, with deviations of $$I_{\mathrm{scat}}(r,\lambda) \approx I_{\mathrm{emit}}(\lambda)\cdot \sigma(\lambda)\cdot G(r)\cdot T_{\mathrm{opt}}/(4\pi r^2),$$8 and $$I_{\mathrm{scat}}(r,\lambda) \approx I_{\mathrm{emit}}(\lambda)\cdot \sigma(\lambda)\cdot G(r)\cdot T_{\mathrm{opt}}/(4\pi r^2),$$9 (Kashyap et al., 2021). The mobile soil-analysis system correctly classified compound-sample soil pH in 97% of cases relative to lab analysis, reduced turnaround time from days or weeks to less than 1 h per cell, and achieved a 9-fold increase in spatial resolution by moving from 1 point/ha compound mapping to approximately 9 points/ha fine-grid testing; the same study gives a resource-savings example of 22% less lime usage if only 6 of 9 subzones require correction (Silva et al., 2022). The ESP32 irrigation platform reports RMSE $\lambda_2 = 980\,\mathrm{nm}$00 over 500 paired measurements, improved to $\lambda_2 = 980\,\mathrm{nm}$01 after soil-type tuning, corresponding to approximately 92% soil-moisture accuracy, and approximately 40% water savings in three 30-day trials, with a total implementation cost of \$45.20 (Hasib et al., 22 Jan 2026).

Machine-learning and RF sensing variants extend these results to recommendation accuracy and joint communication performance. The multimodal crop-recommendation AgroSense uses 50,000 soil images, 25,000 nutrient profiles, and 10,000 paired multimodal samples, and the fused model achieves 98.0% accuracy, 97.8% precision, 97.7% recall, 96.75% F1-score, RMSE of 0.32, and MAE of 0.27. Ablation studies show that removing the image modality reduces accuracy from 98.0% to 97.6%, while removing nutrient features reduces accuracy to approximately 91.0%; paired t-tests give $\lambda_2 = 980\,\mathrm{nm}$02, and one-way ANOVA on RMSE reports $\lambda_2 = 980\,\mathrm{nm}$03 with $\lambda_2 = 980\,\mathrm{nm}$04 (Pandey et al., 1 Sep 2025). The joint sensing-and-communication node reports measured peak antenna gain of $\lambda_2 = 980\,\mathrm{nm}$05, radiation efficiency above 75%, six discrete radiation patterns obtained by varactor biasing, soil-sensor sensitivity of $\lambda_2 = 980\,\mathrm{nm}$06 per unit $\lambda_2 = 980\,\mathrm{nm}$07, and a figure of merit of 1683.9, stated to be more than 10 times higher than the nearest competitor in the paper’s comparison table (Raza et al., 2024).

6. Applications, misconceptions, and research directions

AgroSense applications in the cited works are explicitly decision-oriented. The insect-monitoring design links insect events with soil-moisture probes, weather stations, and crop-health imaging in a unified time-series database and computes multivariate risk indices such as a pest-pressure index defined as normalized insect counts multiplied by humidity and crop-growth rate. Pre-trained pest-forecasting models ingest historical insect-count time series and local microclimate to predict outbreak windows, after which the system can issue SMS or email alerts and recommend targeted biopesticides, trap deployment, or natural-enemy releases; intervention outcomes are then looped back to refine threshold models and minimize chemical use (Rydhmer et al., 2021). The irrigation implementation combines real-time environmental monitoring, automated alerts, and cloud analytics to maintain moisture within hysteresis thresholds, while the crop-recommendation system uses fused soil image and nutrient information to support real-time crop selection (Hasib et al., 22 Jan 2026, Pandey et al., 1 Sep 2025).

Several misconceptions are directly addressed by the literature. One is that scalar sensing alone is sufficient for agricultural event detection: the MiLive paper states that scalar WSN cannot meet all the requirements of ubiquitous intelligent environmental event detections because parameters such as temperature, soil humidity, air humidity, and light intensity are not rich enough to detect all environmental events such as plant diseases and presence of insects, hence the need for multimedia data (Shi et al., 2018). A second is that conventional cellular connectivity is an adequate universal transport layer for digital agriculture; DTN-RuralSense shows that in no-network and poor-network zones, direct cellular alone produced less than 30% successful uploads in 24 h, motivating store-carry-forward relay nodes (Gupta et al., 2016). A third is that AgroSense inherently measures many agronomic variables with a single transducer; in fact, the UCSMS/PRA system currently measures only permittivity and hence volumetric water content, with other parameters such as salinity, pH, and temperature proposed as future overlays (Raza et al., 2024).

The major research directions are consistently toward broader multimodality, lower deployment cost, and more resilient operation. The soil $\lambda_2 = 980\,\mathrm{nm}$08PAD platform proposes additional colorimetric indicators for $\lambda_2 = 980\,\mathrm{nm}$09, $\lambda_2 = 980\,\mathrm{nm}$10, and $\lambda_2 = 980\,\mathrm{nm}$11, with future calibration models for multi-parameter extraction chemistries and multi-label AI classifiers (Silva et al., 2022). The crop-recommendation framework identifies expansion to more soil types, agro-climatic zones, and seasonal variability, integration of real-time weather and IoT feeds, exploration of lightweight transformer variants and self-supervised pretraining, and incorporation of hyperspectral imagery or in-field NIR spectroscopy (Pandey et al., 1 Sep 2025). The joint sensing-and-communication platform proposes LoRa/NB-IoT backhaul, fully printed flexible substrates, and multi-band energy harvesting, while the DTN advisory system recommends prediction of relay-node availability so that farmers can time their submissions more effectively (Raza et al., 2024, Gupta et al., 2016). This suggests that future AgroSense systems will increasingly combine heterogeneous sensors, adaptive communications, and modality-specific inference within a unified precision-agriculture stack.