Slide 1 — Title Slide

Script
“Good morning respected session chair, dignitaries, and everyone present here.
My name is Ayush Kumar Singh, and on behalf of my co-authors, I am presenting our research
work titled Smart Farming Advisory System Using Deep Learning for Crop Disease
Detection, Yield Prediction, and Farmer-Friendly Recommendations.

The main idea behind this work is to develop an intelligent agricultural support system that can
help farmers make better decisions using artificial intelligence. In agriculture, farmers often face
multiple problems at the same time, such as crop diseases, uncertain weather, low soil fertility,
and lack of timely guidance. Instead of solving only one problem, our research aims to integrate
disease detection, yield prediction, and recommendation generation into one unified system.

Through this work, we try to make farming decisions more accurate, more explainable, and more
practical for real-world use.”

Slide 2 — Introduction
Script
“Agriculture continues to be one of the most important sectors in India. It supports the livelihood
of nearly 55 percent of the population and contributes significantly to the economy. However,
despite its importance, agriculture is still highly vulnerable to many challenges.

Crop diseases remain a major cause of reduced productivity. At the same time, climate
variability creates uncertainty in rainfall, temperature, and humidity, which directly affects crop
growth. Soil degradation and nutrient imbalance further reduce yield quality and quantity.
According to the Food and Agriculture Organization, up to 40 percent of global crop production
is lost annually due to pests and diseases, which shows the scale of the problem.

These realities indicate that traditional farming methods, which mostly depend on manual
inspection and past experience, are no longer enough by themselves. Farmers now need
intelligent, data-driven, and reliable support systems. This is where smart farming becomes
highly relevant, because it allows agricultural decisions to be supported by machine learning,
data analysis, and modern predictive models.”
Slide 3 — Problem Statement
Script
“Although many smart agriculture solutions have already been proposed, most of them are
limited in scope. They usually focus on only one task, such as disease detection, crop
recommendation, or yield prediction. But in the real farming environment, these problems are
interconnected, so farmers need a complete decision-support system rather than isolated
predictions.

Another major issue is related to the datasets used for training. Many disease detection models
are trained on laboratory datasets, which are generally clean and controlled. In actual field
conditions, however, images can contain different backgrounds, lighting variations, and
overlapping symptoms, which reduces model generalization.

Similarly, many yield prediction systems use only weather data or basic NPK values, but ignore
other important factors such as micronutrients and vegetation indices. Also, many AI systems
function like black boxes, where the user gets a prediction but does not understand why it was
made. This lack of transparency reduces trust.

So, our problem statement is clear: agriculture needs a unified, explainable, multimodal, and
practical smart farming advisory system.”

Slide 4 — Objectives
Script
“To address these issues, we defined a clear set of objectives for our research.

The first objective is to develop a system that can detect crop diseases from leaf images using
deep learning techniques. Early and accurate disease detection is very important because it helps
reduce crop damage and supports timely treatment.

The second objective is to predict crop yield by combining multiple important agricultural
factors such as weather patterns, soil nutrients, micronutrients, and vegetation indices. This gives
a more comprehensive understanding than using only a single type of data.

The third objective is to generate farmer-friendly recommendations that are not only technically
accurate, but also practical and agronomically valid.
In addition, we aim to improve transparency through explainable AI methods, so that the model’s
outputs can be understood and trusted. Finally, the system is designed in a way that supports use
in rural and resource-constrained environments.”

Slide 5 — Methodology
Script
“Our methodology is based on a hybrid multimodal framework, where different types of
agricultural data are processed together in a coordinated way.

For crop disease detection, we use EfficientNet-B4, which is a deep learning model designed for
image classification. It takes a leaf image as input and predicts the disease category along with a
confidence score.

For yield prediction, we use two different models. First, an LSTM network is used to process
weather time-series data, because weather patterns change over time and must be studied
sequentially. Second, XGBoost is used to process structured tabular data such as soil nutrients,
micronutrients, and vegetation indices.

The outputs from these models are then combined using a fusion layer. After that, a rule-based
recommendation engine checks whether the outputs are agronomically valid. This helps prevent
impractical or biologically unsuitable recommendations.

To make the system explainable, we use Grad-CAM for visual explanation in disease detection
and SHAP for feature-level explanation in yield prediction.”

Slide 6 — Dataset and Preprocessing

Script
“To make the system robust and realistic, we use multiple datasets and multiple input types.

For the disease detection module, we use PlantVillage as a benchmark dataset, along with
Cassava Disease Dataset and CottonWeedDet12, which introduce more realistic field-level
variability. This combination helps improve generalization beyond clean laboratory conditions.

For yield prediction, the system uses weather data such as temperature, rainfall, humidity, and
solar radiation. It also uses soil properties like nitrogen, phosphorus, potassium, and pH, as well
as micronutrients such as magnesium, sulfur, and zinc. In addition, vegetation indices like NDVI
and EVI are included to capture crop health from satellite-based information.

Before model training, preprocessing is applied. Images are resized to a fixed input size,
normalized, and augmented using techniques such as rotation, flipping, and brightness
adjustment. For tabular data, missing values are handled, features are scaled, and weather data is
converted into time windows for the LSTM model. These preprocessing steps improve
consistency, training efficiency, and model performance.”

Slide 7 — Challenges
Script
“There are several practical and technical challenges in building this kind of smart farming
advisory system.

The first challenge is generalization. A model trained on clean laboratory images may not work
equally well in the field, where leaves may appear under different lighting conditions, noisy
backgrounds, or early-stage disease symptoms.

The second challenge is multimodal data integration. Since we combine image data, weather
time-series, soil data, micronutrients, and vegetation indices, it becomes difficult to align
everything properly in terms of timing, scale, and importance.

Another challenge is explainability. While explainable AI techniques like Grad-CAM and SHAP
improve transparency, they also add computational cost.

Finally, rural deployment itself is a challenge. The system must not only be accurate, but also
lightweight, efficient, and practical in environments where internet connectivity and hardware
resources may be limited. So the challenge is to balance performance with deployability.”

Slide 8 — Results
Script
“The results of our proposed framework are encouraging and show that the hybrid system
performs well across both major tasks.
For crop disease detection, the EfficientNet-B4 model achieved an overall accuracy of 98.7
percent, with precision of 98.5 percent, recall of 98.4 percent, and F1-score of 98.4 percent.
These values indicate that the model can classify crop diseases with high reliability.

For yield prediction, the hybrid LSTM and XGBoost framework achieved a Root Mean Squared
Error of 0.18, a Mean Absolute Error of 0.13, and an R-squared value of 0.92. This suggests
strong agreement between predicted and actual crop yield.

In addition, Grad-CAM visualizations showed that the model was focusing on meaningful
disease-affected regions of the leaf, which supports interpretability. SHAP analysis helped
identify the importance of different soil and environmental factors.

We also conducted ablation studies, and they showed that removing NDVI, micronutrients, or
the rule-based engine caused noticeable performance drops. This confirms that each component
contributes meaningfully to the final framework.”

Slide 9 — Future Scope

Script
“Although the proposed system performs well, there are several valuable directions for future
improvement.

One promising direction is the integration of Vision Transformers, which may improve image
understanding by capturing more global relationships in leaf images. Another important step is to
expand the system using larger and more diverse real-world datasets collected from different
crops, regions, and farming conditions.

The framework can also be extended by incorporating IoT sensors for real-time monitoring of
soil moisture, temperature, and humidity. This would make the advisory system more dynamic
and responsive.

In addition, future versions of the system can provide region-specific recommendations, since
agricultural conditions vary across locations. Integration with real-time APIs, mobile
applications, and edge devices can further improve practical usability and support wider adoption
among farmers.

So, while the current framework is already promising, it also opens several strong opportunities
for future research and deployment.”
Slide 10 — Conclusion
Script
“To conclude, our research presents a hybrid smart farming advisory system that combines deep
learning, machine learning, rule-based reasoning, and explainable AI in a single integrated
framework.

The system is capable of detecting crop diseases from leaf images, predicting crop yield using
multimodal environmental and soil data, and generating farmer-friendly recommendations that
are practical and agronomically meaningful.

By integrating EfficientNet-B4, LSTM, XGBoost, and a rule-based engine, our framework goes
beyond isolated prediction tasks and moves toward a more complete agricultural decision-
support system. The use of Grad-CAM and SHAP further improves transparency and addresses
the black-box issue commonly associated with AI systems.

Overall, we believe this work represents a meaningful step toward scalable, explainable, and
deployable precision agriculture solutions.
1. Why did you choose this topic?
Answer
“We chose this topic because agriculture is a highly important sector, especially in countries like
India, but farmers still face serious issues such as crop disease, uncertain yield, and lack of
timely advisory support. We observed that many existing systems solve only one part of the
problem. Our goal was to build a more integrated and practical framework that could combine
disease detection, yield prediction, and farmer-friendly recommendations in a single system.”

2. What is the novelty of your work?

Answer
“The novelty of our work lies in integration. Instead of building only a disease detection model
or only a yield prediction model, we designed a unified smart farming advisory system. We
combine EfficientNet-B4 for image-based disease detection, LSTM for weather time-series
analysis, XGBoost for yield prediction, and a rule-based engine for agronomic validation. We
also include explainable AI through Grad-CAM and SHAP, which improves transparency and
trust.”

3. Why did you use EfficientNet-B4?

Answer
“We selected EfficientNet-B4 because it offers a strong balance between classification accuracy
and computational efficiency. In agricultural disease detection, image quality and feature
extraction are both important, and EfficientNet-B4 is known to perform well while remaining
lighter than some deeper alternatives. This makes it suitable not only for research evaluation but
also for future deployment in practical environments.”

4. Why did you use LSTM for weather data?

Answer
“We used LSTM because weather data is sequential in nature. Parameters like rainfall, humidity,
and temperature change over time, and their temporal relationships directly affect crop growth
and yield. LSTM is particularly suitable for such time-series data because it can learn both short-
term and long-term dependencies more effectively than traditional models.”

5. Why did you use XGBoost for yield

prediction?
Answer
“We used XGBoost because it performs very well on structured tabular data such as soil
nutrients, micronutrients, and vegetation indices. It is robust, efficient, and handles feature
interactions effectively. Since yield prediction depends on multiple static environmental and soil
features, XGBoost was a strong choice for that part of the system.”

6. Why not use only one model for

everything?
Answer
“A single model would not be ideal because the data types in our system are very different. Leaf
images require image-based feature extraction, weather data is temporal, and soil or nutrient data
is structured tabular data. Each of these data types is better handled by a specialized model. That
is why we adopted a hybrid architecture, so each module can process the type of data it is best
suited for.”

7. What is the role of the rule-based engine?

Answer
“The rule-based engine is used to validate machine learning outputs using agronomic conditions.
Sometimes a model may produce a mathematically valid output, but that output may not be
practical from an agricultural point of view. So the rule-based engine acts like a validation layer,
checking conditions such as nutrient sufficiency, rainfall suitability, or disease severity before
generating final recommendations.”

8. What is meant by explainable AI in your

system?
Answer
“In our system, explainable AI means helping the user understand why the model made a
particular prediction. For disease detection, we use Grad-CAM, which highlights the infected
regions of the leaf image that influenced the decision. For yield prediction, we use SHAP, which
shows the contribution of different features such as soil nutrients, weather, and NDVI. This
makes the system more transparent and trustworthy.”

9. What are the main limitations of your

work?
Answer
“One limitation is that real-world agricultural variability is very high, so generalization remains a
challenge. Another limitation is that integrating multiple data sources can increase complexity.
Also, explainability methods can add computational overhead. Finally, although the framework
is designed with deployment in mind, full real-world implementation would still require larger
field datasets, sensor integration, and regional customization.”

10. How is your system better than existing

systems?
Answer
“Our system is better mainly because it is integrated, multimodal, and explainable. Existing
systems often focus on only one task, whereas our framework combines disease detection, yield
prediction, and advisory generation. It also includes micronutrients and vegetation indices, which
are often ignored in simpler systems. In addition, our use of explainable AI and agronomic
validation improves both transparency and practical relevance.”

11. Can this system be used directly by

farmers?
Answer
“At the current stage, the system is more like a research prototype, but it has strong potential for
farmer use. The architecture is designed with offline-first and lightweight deployment in mind,
which is important for rural environments. With further development, such as mobile application
integration, IoT sensor support, and region-specific customization, it can become a practical tool
for real-world use.”

12. Why did you include NDVI and EVI?

Answer
“We included NDVI and EVI because they provide useful information about crop health,
greenness, and vegetation condition from satellite-derived data. These indices complement soil
and weather features, giving the model a broader understanding of the crop environment. Our
ablation results also showed that removing vegetation indices reduced performance, which
confirms their importance.”

13. What is the significance of micronutrients

in your system?
Answer
“Micronutrients such as magnesium, sulfur, and zinc are essential for proper crop growth, but
they are often ignored in many recommendation and yield prediction systems that focus only on
NPK. By including micronutrients, our framework becomes more agriculturally complete and
can produce more realistic yield predictions as well as better fertilizer-related recommendations.”
14. Are the reported results realistic?
Answer
“Yes, the reported results are realistic within the scope of the proposed architecture and datasets.
EfficientNet-based disease detection models can achieve very high accuracy on benchmark and
mixed datasets, and hybrid yield prediction models can also achieve strong regression
performance when multimodal inputs are used. However, for real-world deployment, these
results should always be validated further on larger and more diverse field datasets.”

15. What is the biggest challenge in practical

deployment?
Answer
“The biggest challenge in practical deployment is ensuring robust performance in real farming
conditions. Field environments are much more variable than controlled datasets. In addition,
collecting and aligning multimodal data such as images, weather, soil, and satellite information
can be difficult. There is also the challenge of making the system lightweight enough for rural
and low-resource settings.”

16. How does your work support precision

agriculture?
Answer
“Our work supports precision agriculture by making farming decisions more specific, data-
driven, and targeted. Instead of relying on general recommendations, the system analyzes leaf
condition, environmental data, soil properties, and crop health indicators together. This allows
recommendations to be more personalized and scientifically grounded, which is the main idea of
precision agriculture.”
17. Why is your work important for India?
Answer
“This work is especially important for India because agriculture supports a very large share of
the population, but many farmers still rely on manual observation and generalized advice. Since
India also faces issues like crop disease, climate uncertainty, and soil variability, an intelligent
advisory system can provide significant practical value. A scalable and explainable framework
like ours can contribute to more informed farming decisions.”

18. What would you do next if you continue

this research?
Answer
“If we continue this research, the next steps would be to collect larger real-world datasets,
integrate IoT sensors for real-time monitoring, and develop a mobile or edge-based deployment
version. We would also like to explore transformer-based image models and make the
recommendation logic more region-specific, so that the system becomes even more practical and
robust.”

19. Why did you choose a hybrid model

instead of a pure deep learning pipeline?
Answer
“We chose a hybrid model because different parts of the problem require different strengths.
Deep learning is very effective for image-based disease detection, while models like XGBoost
are more effective and efficient for structured tabular data. LSTM is appropriate for sequential
weather information. By combining them, we can use the best method for each data type instead
of forcing one model to handle everything.”
20. What is one line that summarizes your
research?
Answer
“Our research proposes an integrated, explainable, and practical smart farming advisory
framework that combines disease detection, yield prediction, and recommendation generation for
precision agriculture.”