AI-Driven Precision Farming: Optimizing Pest

Detection, Soil Moisture, and Fertilization

Chiranth M Selar Vikas M K Darshan K Seervi

Dept. of Computer Science Dept. of Computer Science Dept. of Computer Science
Amrita School of Computing Amrita School of Computing Amrita School of Computing
Mysuru, India Mysuru, India Mysuru, India
[email protected] [email protected]

Abstract—This project suggests a precision farming system land types, including ”Dolomite gypsum” and ”Fallow (left
that integrates intelligent irrigation, fertiliser optimisation, and bare),” are identified through the use of image analysis and
pest detection to increase agricultural productivity and sustain- pre-trained deep learning models for soil pattern classifi-
ability. A deep learning-based pest detection model is developed
using a custom dataset of crop leaf photos tainted by pests such cation. As a result,particular fertiliser recommendations are
as beetles, aphids, and other common agricultural hazards. By provided,such as NPK 10:26:26,and application rates,such as
facilitating early detection and focused treatment, this lessens 40 kg/ha. This method reduces the negative effects of excessive
the overuse of pesticides. The use of UAV-captured soil images fertilisation on the environment while also ensuring precise nu-
to classify land types, such as fallow land and dolomite gyp- trient delivery. The system incorporates an intelligent irrigation
sum, enables data-driven fertiliser recommendations tailored to
specific soil conditions. Additionally, operations research (OR) scheduling model that makes use of image-based soil moisture
optimisation techniques are combined with image-based moisture detection to improve the efficiency of water resources.The
detection to determine efficient irrigation schedules that ensure irrigation process is improved to use less water while meeting
crops receive adequate hydration while conserving water. Crop crop hydration requirements by using operations research tech-
yield is increased by the combined approach and decreasing re- niques like linear programming and goal programming,as well
source waste,and also further supporting sustainable agriculture
activities. as by using drone imagery to assess soil texture and moisture
Index Terms—IoT, Smart Irrigation, Image Processing, Soil levels.This study aims to increase agricultural productivity,
Moisture, Machine Learning, Precision Agriculture reduce input waste,and promote sustainable farming practices
by combining these three elements—pest detection,fertiliser
I. I NTRODUCTION optimisation,and smart irrigation—into a unified AI-driven
Although agriculture is vital to human survival, it faces framework.
many challenges, such as pest issues, insufficient fertiliser
use, and subpar irrigation techniques. In addition to hav- II. R ELATED W ORKS
ing an effect on crop yield, these problems cause long- The development of intelligent irrigation systems with op-
term environmental damage.Precision agriculture, fuelled by timum utilization of water and improved crop health has been
advancements in computer vision,artificial intelligence (AI), hugely facilitated by the Internet of Things’ (IoT) adoption
and optimisation techniques,has become a ground-breaking in agriculture. To effectively control and monitor irrigation,
method as the demand for intelligent and sustainable farming such systems employ various technologies such as machine
methods grows. This study offers a holistic approach that learning, sensors, and image processing. One of these sys-
applies image-based analysis and Operations Research (OR) tems uses thermal vision to examine soil surface status and
techniques to address three core elements of modern precision image processing to identify water leaks and evaluate crop
farming.To identify plant diseases and pests directly from leaf condition based on soil moisture, temperature, and ultrasonic
photos,a deep learning-based pest detection system is first sensors linked to a Raspberry Pi [1]. While machine learning
developed. The system can classify diseases like early blight models like Gradient Boosting Trees (GBT) forecast the
and late blight or healthy conditions in crops like potatoes health of leaves from a Leaf Color Chart (LCC), techniques
by using convolutional neural networks (CNNs) that have like segmentation, noise cleaning, and distortion correction
been trained on labelled datasets. By enabling timely action, improve image quality. Although promising, the Raspberry
this real-time image-driven pest detection reduces pesticide Pi’s processing limit prevents real-time performance, and
overuse and improves crop health overall. Second, the study high initial expense and limited dataset availability limit
offers a module for optimising fertiliser distribution and rec- scalability. External factors like soil condition and weather
ommendations using UAV-captured farmland images.Various conditions also reduce accuracy. A water irrigation system
using Arduino, a Raspberry Pi for collecting data, and soil for rare pests, and the high computational costs for complex
moisture sensors is utilized in another study aiming at smart models, even though the accuracy was greater than 95%. [10].
irrigation for small-sized nurseries. Notifications are received For farmers in Burkina Faso who grow maize, tomatoes, and
via an Android application utilizing Firebase [2]. But its onions, PlanteSaine (2024) released a mobile AI app that uses
dependence on thermal vision, absence of temperature and ANN and EfficientNetB3 for classification. Its performance
humidity sensors, and Android compatibility restrict its wider was limited to three crops and suffered in different lighting
use.Although the omission of post-2019 studies and foreign- conditions, even though it handled over 29,000 images with
language articles brings about bias and necessitates regular DA. Additional difficulties were brought on by reliance on the
updating, a wide-ranging survey of 178 IoT-based irrigation internet and smartphone quality for updates. [11]. AI-Powered
studies classifies them into sensors, actuators, wireless tech- Intelligent Agriculture for Potato Illnesses (2024) examined
nology, and visualization methods.It emphasizes significant ML (SVM, RF) and DL (VGG, ResNet, EfficientNet) tech-
parameters like water quantity and quality, soil state, and niques in detail, pointing out that 70.8% of studies relied heav-
climate [3]. Though there is an influence on precision by ily on imagery data and that only 25% integrated climate data.
spatial rainfall variation due to the limited number of weather The main limitations for small farms were scalability con-
stations, there are other studies that investigate smartphone- cerns and early detection challenges. [12]. YOLOv8 (73.8%
based irrigation scheduling involving meteorological data and mAP) and GPT-4o were combined for pest analysis (94%
FAO Penman-Monteith equation for estimating evapotranspira- accuracy) in Advanced Multi-Modal Pest Detection (2024).
tion and adjusting schedule with crop coefficients [4]. Utilizing The system encountered issues with hardware limitations on
methods including background removal, Otsu thresholding, mobile devices, high API costs, and internet dependency for
and anisotropic diffusion for feature extraction as well as clas- LLM use while utilising large datasets (PlantVillage, IP102).
sification, yet another intelligent agriculture system combines [13]. The ML (SVM, RF) and DL (LeNet, ResNet) models
IoT and image processing to identify leaf diseases and measure tested in Tomato Pest Identification (2024) showed optimiser
soil moisture [5]. Although with promising results, further sensitivity (e.g., AdaDelta underperforming) and difficulties
research is needed to completely understand scalability and with visually similar pests. Computational demands and gaps
adaptation to environmental changes. Moreover, an automated in field testing were unresolved issues. [14]. Wireless Sensor
irrigation system utilizes an image processing unit based on Networks (WSNs) for agricultural pest detection and man-
Raspberry Pi to suggest appropriate crops based on soil pH agement are examined in this 2015 review. In addition to
values. Although this method incorporates soil pH monitoring, automated chemical and biological control techniques, the
some limits in terms of system accuracy are not yet explored study examines environmental, acoustic, and image sensors
[6]. Fisher’s Linear Discriminant Analysis (LDA) is applied for real-time monitoring. High expenses, limited multi-pest de-
for color-based moisture detection and a Transductive Support tection, problems with sensor durability, and ecological trade-
Vector Machine (TSVM) for classification in a cloud-based offs are some of the main obstacles. The study emphasises
IoT irrigation system utilizing wireless sensor networks and the need for **more scalable, farmer-adaptable solutions to
cloud computing. ZigBee and ESP8266 modules are utilized close technology gaps and enhance sustainability, even though
for data transmission, and the system operates on solar- WSNs show promise for precision agriculture. [15]. Using
powered batteries, ensuring sustainability. Scalability and real- high-resolution images, this 2025 study uses deep learning
world implementation problems are not entirely solved [7].An (CNNs and transfer learning) to detect agricultural pests with
IoT platform, Smart&Green, offers a multi-layer architecture high accuracy (99.9%). However, difficulties include identi-
for data monitoring, fusion, and soil moisture forecasting. It fying rare species, environmental variability, similar-looking
combines outlier removal techniques and soil moisture estima- pests, and poor image quality. [16].With the focus on soil
tion algorithms to improve irrigation precision. However, code sensors and pest detection models, recent work highlights
reuse, customization issues, and variability in data quality limit IoT, blockchain, and deep learning to be employed in smart
performance, necessitating further improvements [8]. Four agriculture. Infrared imaging is utilized for crop monitoring
deep learning models—AlexNet, ResNet-50, EfficientNet-B4, in some systems, while blockchain enhances data security
and Vision Transformer (ViT)—were compared with different but comes with a cost. One of the weaknesses of current
data augmentation (DA) methods in the study Classification approaches is still the centralized processing of data. A new
of Agricultural Pests from Digital Images with Deep Learning system is attaining high accuracy in soil analysis and fertilizer
(2025). Grid search was used to adjust the hyperparameters, optimization for sustainable agriculture by combining deep
but the study was constrained by a small dataset (5,494 photos learning with decentralized storage [17].
in 12 pest classes) and computational constraints (single GPU High-speed crop variability assessment using spatial inter-
environment). Remarkably, unlike CNNs, ViT did not get polation techniques such as IDW and TIN is the central theme
better with DA. [9]. Using Deep Learning to Identify Pests of current advancements in UAV-based precision agriculture.
and Plant Diseases (2025) broadened the scope by incor- These methods maintain high correlations (up to r=0.93) with
porating IoT-based real-time monitoring, CNNs, ViTs, and high-resolution orthomosaics while reducing computation time
hybrid models with synthetic data generation (GANs/VAEs). by as much as 98% compared to traditional photogrammetry.
Limitations included environmental variability, the lack of data The platform facilitates site-specific management of irrigation,
soil fertility, and weed management through multisource UAV is made to operate iteratively, guaranteeing solid and precise
imagery for real-time monitoring. Despite boundary interpola- insights for agricultural operations. The process guarantees
tion limitation challenges, the approach is effective for a range an automated but adaptable system to help farmers efficiently
of crops and climates and offers a low-cost option for sustain- make decisions about pest identification, fertilizer application,
able agriculture [18]. With an emphasis on crop-fertilizer rec- irrigation, and soil analysis.
ommendation and soil fertility forecasts, the literature explores
AI and ML techniques for precision agriculture. Accuracy and A. Data Collection and Loading
efficiency are increased by methods like SVM, KNN, and For each goal, the project starts by gathering three different
neural networks. Live monitoring is made possible by sensor- datasets: Images of various pests, such as the Amphid, Beetle,
based and Internet of Things devices. It is anticipated that these Anyworm, Test Beetle, Test Mite, Swafly Test, Stem Fly Test,
technologies will optimize resource use and agricultural pro- and others, are included in the Pest Detection Dataset.The pest
ductivity [19]. Recent innovations use the Internet of Things detection model will be trained using these photos, which have
(IoT) in conjunction with microcontrollers and soil sensors been labelled according to the type of pest. Dataset for Soil
to optimize and automate drip irrigation scheduling. Accurate Analysis (UAV-based photos): UAV photos are gathered for
irrigation control guarantees higher crop development, while soil analysis in order to forecast the state of the soil, which
real-time monitoring improves water use efficiency. Remote is then connected to the requirement for the best possible
operation is made easier with wireless data upload to cloud fertiliser and irrigation practices. Drone-captured images of
platforms. The technique outperforms traditional methods in soil moisture levels are included in the moisture level detection
terms of water savings, plant height, and leaf size. [20]. Pre- dataset, which will help with irrigation system optimisation.
cision agriculture has been transformed by Unmanned Aerial
Vehicles (UAVs), which provide high-resolution imaging for B. Preprocessing and Target Definition
crop monitoring. UAVs use sensors like RGB, multispectral, Data Augmentation and Cleaning: To make sure the model
and thermal to help with weed identification, yield estimation, generalises well, data augmentation techniques such as rota-
and irrigation control. Photogrammetry and machine learning- tion, scaling, and flipping are used after the images have been
based data processing can lead to actionable insights. UAVs cleaned to remove noise. Feature extraction: Key characteris-
are a crucial component of smart farming because of their cost tics, particularly for the pest detection dataset, are taken out
and adaptability terms of water savings, plant height, and leaf of the photos, including edges, texture, and colour histograms.
size. [21]. Data Splitting: To fine-tune the model, the datasets are divided
into training and testing sets along with the relevant validation
III. P ROPOSED M ETHODOLOGY sets. Definition of the Target Variable: The target for pest
detection is the type of pest (e.g., test mite, beetle, etc.). The
goal of soil analysis is to classify the soil’s condition (e.g.,
Healthy, Depleted). The goal of moisture level detection is to
ascertain the soil’s moisture content, which is used to calculate
irrigation requirements.

C. Training Machine Learning Models

Pest Detection Model: To identify pests in photos, deep
learning models like Convolutional Neural Networks (CNNs)
are employed. Soil Analysis Model: To extract pertinent fea-
tures from UAV photos, deep learning models such as CNNs
or Transfer Learning (using pre-trained models like ResNet or
VGG16) are used. Moisture Level Prediction Model: Recurrent
neural networks (RNNs) or CNNs, if time series data is avail-
able, are examples of similar models that are used to forecast
the soil’s moisture content. To guarantee optimal performance
on all datasets, the models undergo iterative training that
includes cross-validation and hyperparameter tuning.
Fig. 1. Proposed Methodology Workflow
D. Prediction and Decision Support
A methodical workflow for evaluating agricultural data Pest Detection: Farmers can take prompt action by using the
and improving farm management choices is described in the model to predict the type of pest from recently taken photos.
suggested precision farming system.Data collection, prepro- Soil Analysis: Using real-time UAV imagery, the soil analysis
cessing, target definition, machine learning model training, model forecasts the health of the soil and suggests fertilisers.
model refinement, and farm management prediction are the six Moisture Level Detection: By forecasting the soil’s moisture
crucial stages that make up the system architecture.Every stage content, the model helps optimise irrigation schedules and
avoid water waste or over-irrigation. To evaluate prediction re- TABLE II
liability, confidence scores are included with every prediction. P ERFORMANCE G AP A NALYSIS
Decision Support Interface: Farmers can quickly understand Metric Train-Test Difference
the model’s output and take appropriate action based on AI- Accuracy 7.12%
Loss 0.1660
powered insights thanks to the results’ display on an intuitive Precision 0.04
dashboard. Recall 0.05
F1-score 0.05
IV. RESULT1
A. Model Performance
The pest classification model achieved an overall test ac- • Performance Interpretation:
curacy of 78.22% with a test loss of 0.6581.The confusion • Key Observations:
matrix (Figure 2) reveals varying performance across different – The model shows no severe overfitting (differences
pest classes: ¡10% for all metrics)
– Precision remains strongest metric (0.83 test), sug-
TABLE I gesting reliable positive predictions
C LASSIFICATION R EPORT – Recall drop (0.78 test) indicates some missed detec-
Pest Class Precision Recall F1-Score Support tions in unseen data
aphid test 0.66 0.77 0.71 100 – Balanced F1-scores confirm stable overall perfor-
armyworm test 0.83 0.90 0.87 50 mance
beetle test 0.91 0.98 0.94 50
mite test 0.70 0.78 0.74 100 The metrics collectively indicate a well-functioning pest
sawfly test 0.90 0.86 0.88 50 classification system with room for targeted improve-
stem borer test 0.91 0.78 0.84 50
stem fly test 0.91 0.42 0.58 50 ments in recall and generalization.

C. Confusion Matrix Interpretation

Key observations from the classification report (Table IV)
include:
• Excellent performance on beetle identification (F1-score:
0.94)
• Strong results for armyworm (F1-score: 0.87) and sawfly
(F1-score: 0.88)
• Moderate performance for aphids (F1-score: 0.71) and
mites (F1-score: 0.74)
• Poor recall for stem fly (0.42), suggesting difficulty in
detecting this pest
This demonstrates the system’s capability to not only clas-
sify pests but also provide actionable treatment recommenda-
tions with precise dosage calculations.
B. Model Accuracy Metrics Analysis
The performance metrics reveal several important charac-
teristics of the pest classification model:
• Training-Test Consistency:
– The model maintains relatively stable performance
between training (85.34% accuracy) and test sets
(78.22% accuracy) Fig. 2. confusion matrix for the seven-class pest classification model.
– The 7.12% accuracy drop suggests moderate but
acceptable generalization The confusion matrix in Figure 3 reveals several key
– Loss values show a similar pattern (0.4921 train vs insights about model performance:
0.6581 test) – Correct Classifications (True Positives):
• Metric Correlations: ∗ Beetle: 49/50 (98% accuracy) - nearly perfect
– High precision (0.87 train, 0.83 test) indicates few recognition
false positives ∗ Armyworm: 45/50 (90% accuracy) - strong perfor-
– Recall (0.83 train, 0.78 test) shows good detection mance
capability ∗ Aphid: 77/100 (77% accuracy) - moderate perfor-
– Balanced F1-scores (0.84 train, 0.79 test) demon- mance
strate harmonic mean consistency – Major Confusion Patterns:
 ∗ Stem Fly frequently misclassified: The optimization results demonstrate:
· 14/50 as Aphid (28% error) ∗ Practical dosage recommendation (2.50 liters) for
· 8/50 as Mite (16% error) field application
∗ Aphid-Mite confusion: ∗ Quantified cost-effectiveness metric (5.00) for
· 17/100 Aphids predicted as Mites comparison
· 3/100 Mites predicted as Aphids ∗ Computational efficiency suitable for real-time
decision support
D. Optimization Explanation
V. RESULT2
TABLE III A. Second Objective Model Performance
P ESTICIDE O PTIMIZATION R ESULTS FOR A PHID T REATMENT The soil treatment classification model achieved an
Parameter Value overall accuracy of 95% with perfect classification
Detected Pest aphid for rice husk charcoal:
Recommended Pesticide Imidacloprid
Optimal Amount 2.50 liters
Cost Function Value 5.00 TABLE IV
Optimization Method Golden Section Search C LASSIFICATION R EPORT
Search Range [0.5, 5] liters
Tolerance 0.001 Class Precision Recall F1-Score Support
Dolomite gypsum 0.93 0.98 0.95 607
Fallow (left bare) 0.88 0.88 0.88 357
The pesticide optimization employs a Golden Sec- Indian mustard 1.00 0.94 0.97 509
tion Search algorithm to determine the optimal pes- Rice husk charcoal 1.00 1.00 1.00 454
ticide quantity. The mathematical formulation con-
sists of:
∗ Top Performers:
Minimize f (x) = (x − 2.5)2 + 5 (1)
· Rice husk charcoal: Perfect classification
where: (F1=1.00)
∗ x represents the pesticide amount (in liters) · Indian mustard: Near-perfect precision (1.00)
∗ f (x) is the cost-effectiveness function with 94% recall
∗ The constant 2.5 represents the ideal dosage point ∗ Moderate Performers:
∗ The +5 term establishes a baseline cost · Dolomite gypsum: Balanced precision-recall
[h] Golden Section Search Implementa- (0.93-0.98)
tion · Fallow land: Consistent but lower performance
√
1: Initialize ϕ ← ( 5 − 1)/2 {Golden ratio} (F1=0.88)
2: Set search bounds a ← 0.5, b ← 5
B. Confusion Matrix Analysis
3: Calculate initial points c ← b − ϕ(b − a), d ←
a + ϕ(b − a)
4: while |b − a| > tolerance do
5: if f (c) < f (d) then
6: b ← d {Narrow search range}
7: else
8: a←c
9: end if
10: Update c ← b − ϕ(b − a), d ← a + ϕ(b − a)
11: end while
12: return (b + a)/2 {Optimal solution}
Key features of the implementation:
∗ Unimodal Assumption: Requires the cost func-
tion to have a single minimum within the search
interval
∗ Bracket Reduction: Systematically narrows the
search space by ϕ ≈ 0.618 each iteration
∗ Convergence: Guaranteed to find the optimum
within specified tolerance (0.001 in this case) Fig. 3. Confusion matrix visualization for soil treatment classes.
∗ Efficiency: Achieves linear convergence rate with
only one function evaluation per iteration
∗ Classification Accuracy: VI. R ESULTS 3
· Dolomite gypsum: 593/607 (97.7% correct) A. Soil Moisture Classification Performance
· Fallow: 314/357 (87.9% correct) The soil moisture classification model was trained for
· Indian mustard: 481/509 (94.5% correct) 30 epochs, progressively improving in accuracy and
· Rice husk charcoal: 454/454 (100% correct) reducing loss, as shown by the training and validation
∗ Error Analysis: performance:
· Primary confusion: 12.0% of Fallow samples ∗ Final training accuracy: 87.72%, training loss:
misclassified as Dolomite (43/357) 0.3956
· Secondary confusion: 5.5% of Mustard samples ∗ Final validation accuracy: 80.14%, validation
misclassified as Fallow (28/509) loss: 0.5913
C. Fertilizer Optimization Results
Classification Report:
The optimization was performed using gradient de-
scent to minimize the cost function:
TABLE VI
C LASSIFICATION REPORT METRICS FOR EACH SOIL MOISTURE CLASS .
TABLE V
FALLOW L AND T REATMENT O PTIMIZATION Class Precision Recall F1-Score Support
Dry 0.82 0.95 0.88 44
Parameter Value High 0.91 0.74 0.82 43
Soil Treatment Class Fallow (left bare) Low 0.81 0.93 0.87 42
Recommended Fertilizer NPK 10:26:26 Medium 0.89 0.79 0.84 43
Optimal Dose 40.00 kg/ha Accuracy 0.85
Cost Function Value 10.00 Macro Avg 0.86 0.85 0.85 172
Optimization Method Gradient Descent Weighted Avg 0.86 0.85 0.85 172

NPK Fertilizer Explanation: The recommended On the test data, the model recorded a test accuracy
NPK 10:26:26 fertilizer contains: of 78.0% and a respective test loss of 0.5020%,
∗ N (Nitrogen): 10% - Promotes vegetative growth signifying strong generalization ability among the
and leaf development four diverse soil moisture classes. This indicates
∗ P (Phosphorus as P2 O5 ): 26% - Enhances root the model’s capability to strongly learn the intrin-
development and energy transfer sic patterns in the input features and make sound
∗ K (Potassium as K2 O): 26% - Improves water predictions on new, unseen data. In spite of the
regulation and stress resistance intrinsic volatility of the soil conditions, the classifier
achieves an optimal compromise between bias and
Mathematical Formulation: The cost function was
variance and can be used in real-world applications
defined as:
in precision irrigation systems.
J(θ) = (θ − θ∗ )2 + C (2)
B. Confusion Matrix Analysis
where:
∗ θ: Fertilizer dose (kg/ha)
∗ θ∗ = 40.00 kg/ha: Optimal dose
∗ C = 10.00: Baseline cost
The gradient descent update rule:
θk+1 = θk − α · 2(θk − θ∗ ) (3)
with learning rate α = 0.1, converging in 15 itera-
tions.
∗ Technical Implementation:
· CLI processing with automatic soil class detec-
tion
· Database-driven fertilizer recommendation
· Efficient convergence (tolerance = 0.001)
∗ Agricultural Implications:
· High phosphorus content (26%) ideal for soil
rejuvenation
· Balanced 40 kg/ha application rate Fig. 4. Confusion matrix visualization for soil moisture classification.
· Optimal cost-benefit ratio (J = 10.00)
 Dry: 42/44 correct (95.5%) detection could be improved.While moisture predic-
High: 32/43 correct (74.4%) tion achieves 85% accuracy for optimal irrigation,
Low: 39/42 correct (92.9%) soil condition analysis shows an impressive 95%
Medium: 34/43 correct (79.1%) accuracy, allowing for exact fertiliser recommenda-
tions such as NPK 10:26:26 applications.Cutting-
C. Irrigation Optimization Result edge optimisation strategies, such as gradient descent
∗ Predicted class: Dry algorithms and Golden Section Search, guarantee
∗ Estimated moisture: 10% effective resource allocation while reducing environ-
∗ Water needed to reach 50%: 40 units mental impact.
∗ Optimal water allocation (Gradient Descent): The study outlines specific directions for future
40.00 units development, such as the creation of hybrid AI
models to enhance performance in a range of en-
Mathematical Formulation: The cost function used vironmental circumstances, the integration of real-
for irrigation optimization is defined as: time IoT sensors, and the expansion of datasets
for under-represented pests.By effectively connecting
J(θ) = (θ − θ∗ )2 + C (4) state-of-the-art technology with real-world farming
requirements, this system provides a scalable frame-
where: work for environmentally conscious and productive
∗ θ: Current water allocation (units) sustainable agriculture. The results show that AI
∗ θ∗ : Target water needed to reach desired soil and machine learning can revolutionise conventional
moisture farming methods into accurate, data-driven opera-
∗ C: Constant baseline cost (set to 5.0) tions that maximise yields while preserving essen-
The gradient (first derivative) of the cost function is: tial resources. This is a major advancement in the
development of smart agriculture solutions.
d
∇J(θ) = [(θ − θ∗ )2 + C] = 2(θ − θ∗ ) (5) R EFERENCES
dθ
[1] S. H. Ranpise, H. Mandve, S. Charan, P. Pal, S. K. Singh,
Using the gradient descent update rule: S. Bansod, and Y. Kumar, “Iot based smart irrigation
system for crops and water leakage detection using image
processing,” in 2023 Global Conference on Information
θk+1 = θk − α · ∇J(θk ) (6) Technologies and Communications (GCITC). IEEE, 2023,
pp. 1–5.
which simplifies to: [2] I. Fernando and S. Gunasekara, “Smart irrigation system for
a small plant nursery based on soil moisture level,” 2020.
θk+1 = θk − α · 2(θk − θ∗ ) (7) [3] L. Garcı́a, L. Parra, J. M. Jimenez, J. Lloret, and P. Lorenz,
“Iot-based smart irrigation systems: An overview on the
where: recent trends on sensors and iot systems for irrigation in
precision agriculture,” Sensors, vol. 20, no. 4, p. 1042, 2020.
∗ α: Learning rate (set to 0.1) [4] K. W. Migliaccio, K. T. Morgan, G. Vellidis, L. Zotarelli,
∗ k: Iteration index C. Fraisse, D. L. Rowland, J. H. Andreis, J. H. Crane, and
B. A. Zurweller, “Smartphone apps for irrigation schedul-
The algorithm iteratively updates θ until conver- ing,” in 2015 ASABE/IA Irrigation Symposium: Emerging
gence, i.e., until |∇J(θ)| < 10−3 . Technologies for Sustainable Irrigation-A Tribute to the
Career of Terry Howell, Sr. Conference Proceedings. Amer-
ican Society of Agricultural and Biological Engineers, 2015,
Optimization Highlights: pp. 1–16.
[5] S. Bagal, T. More, A. Paranjape, and S. Yadav, “Intelligent
∗ Precise classification for Dry and Low classes (F1 agriculture mechanism using internet of things and image
score: 0.87) processing.”
∗ High class showed lower recall (0.74), indicating [6] S. Sagar, B. Debjeet, L. Advait, and N. Mishra, “Moisture
and ph detection using sensors and automatic irrigation
potential room for improvement system using raspberry pi based image processing.”
∗ Consistent macro and weighted averages affirm [7] N. Kabilan and M. S. Selvi, “Surveillance and steering of
balanced model performance irrigation system in cloud using wireless sensor network and
wi-fi module,” in 2016 International Conference on Recent
VII. C ONCLUSION Trends in Information Technology (ICRTIT). IEEE, 2016,
pp. 1–5.
This study introduces a novel AI-powered precision [8] N. GS Campos, A. R. Rocha, R. Gondim, T. L. Coelho da
farming system that uses integrated insect detection, Silva, and D. G. Gomes, “Smart & green: An internet-of-
things framework for smart irrigation,” Sensors, vol. 20,
soil analysis, and intelligent irrigation to handle no. 1, p. 190, 2019.
major agricultural issues in a comprehensive man- [9] P. L. de Oliveira Costa, T. M. de Oliveira Costa, L. F. R.
ner. The system’s accuracy in classifying pests is Moreira, L. H. F. P. Silva, and J. F. Mari, “Classification of
agricultural pests through digital images using deep learn-
78.22%; it excels at identifying beetles and army- ing,” Revista de Informática Teórica e Aplicada, vol. 32,
worms,but it also identifies areas where stem fly no. 1, pp. 18–25, 2025.
[10] M. Shoaib, A. Sadeghi-Niaraki, F. Ali, I. Hussain, and
S. Khalid, “Leveraging deep learning for plant disease and
pest detection: a comprehensive review and future direc-
tions,” Frontiers in Plant Science, vol. 16, p. 1538163, 2025.
[11] O. Appiah, K. O. Hackman, B. A. A. Diallo, K. O.
Ogunjobi, S. Diakalia, O. Valentin, D. Abdoul-Karim, and
G. Dabire, “Plantesaine: An artificial intelligent empowered
mobile application for pests and disease management for
maize, tomato, and onion farmers in burkina faso,” Agricul-
ture, vol. 14, no. 8, p. 1252, 2024.
[12] A. Kaur, G. S. Randhawa, F. Abbas, M. Ali, T. J. Esau,
A. A. Farooque, and R. Singh, “Artificial intelligence driven
smart farming for accurate detection of potato diseases: A
systematic review,” IEEE Access, 2024.
[13] M. Mohammed and J. Vodnala, “Advanced plant pest de-
tection system: A multi-modal approach to agricultural pest
management.”
[14] C. H. Praharsha, A. Poulose, and C. Badgujar, “Compre-
hensive investigation of machine learning and deep learning
networks for identifying multispecies tomato insect images,”
Sensors, vol. 24, no. 23, p. 7858, 2024.
[15] S. Azfar, A. Nadeem, and A. Basit, “Pest detection and
control techniques using wireless sensor network: A review,”
Journal of Entomology and Zoology Studies, vol. 3, no. 2,
pp. 92–99, 2015.
[16] A. Rafique, M. Abbasi, N. Akram et al., “Application of
deep learning models for pest detection and identification,”
Mehran University Research Journal of Engineering and
Technology, vol. 44, no. 2, pp. 117–128, 2025.
[17] B. C. Preethi, G. Sugitha, and T. B. Sivakumar,
“Optimized fertilizer dispensing for sustainable agriculture
through secured iot-blockchain framework,” International
Journal of Advanced Computer Science and Appli-
cations, vol. 15, no. 9, 2024. [Online]. Available:
http://dx.doi.org/10.14569/IJACSA.2024.0150927
[18] S. Vélez, M. Ariza-Sentis, M. Panić, B. Ivoević,
D. Stefanović, J. Kaivosoja, and J. Valente,
“Speeding up uav-based crop variability assessment
through a data fusion approach using spatial
interpolation for site-specific management,” Smart
Agricultural Technology, 2024. [Online]. Available:
https://api.semanticscholar.org/CorpusID:270528115
[19] G. V. Reddy, M. V. K. Reddy, K. Spandana, Y. Subbarayudu,
A. Albawi, R. Chandrashekar, A. Singla, and P. Praveen,
“Precision farming practices with data-driven analysis and
machine learning-based crop and fertiliser recommendation
system,” in E3S Web of Conferences, vol. 507. EDP
Sciences, 2024, p. 01078.
[20] C. D. Singh, K. Rao, M. Kumar, Y. A. Rajwade et al.,
“Development of a smart iot-based drip irrigation system for
precision farming.” Irrigation & Drainage, vol. 72, no. 1,
2023.
[21] D. C. Tsouros, S. Bibi, and P. G. Sarigiannidis, “A review
on uav-based applications for precision agriculture,” Infor-
mation, vol. 10, no. 11, p. 349, 2019.