CNN-Driven Crop Yield Prediction and Dynamic Crop Calendar Optimization Using Agro-Climatic Data

CNN-Driven Crop Yield Prediction and Dynamic Crop Calendar Optimization Using Agro-Climatic Data

Nireeksha D

Student, St. Joseph Engineering College, Mangalore

Murari B K

Assistant Professor, St. Joseph Engineering College, Mangalore

Abstract Agriculture today faces climate change, unexpected weather and increasing challenges from ineffective crop management practices. This study presents a smart, data-driven approach that merges agricultural- agent data with crop yield forecasts. It also adjusts the crop calendar using machine learning. The proposed system is built on a TensorFlow-based deep nerve network, especially a multi-layer feedforward model. The model is developed to analyse the complex relationships among environmental characteristics, agricultural inputs, and seasonal effects on production The model uses essential inputssuch as crop type, location, rainfall, temperature, soil type, and past yieldsto generate accurate yield predictions. The yield predictions feature a dynamic crop calendar, adjusting planting and harvesting schedules based on real-time and forecasted farming conditions. The model with real -time seasonal and soil figures continuously detect the best combinations to promote system productivity and reduce risks. Display assessment, which uses a matrix like RMSE and R, shows how effective this approach is. This study suggests that adaptive schedules, intensive learning, and smart cultivation methods suited to the climate can support farmers. These practices help improve yields and stabilize agriculture.

Index Terms TensorFlow, Convolutional Neural Network (CNN), Crop Calendar Optimization

1. INTRODUCTION

   Good nutrition from agriculture is essential not just for food security but also for supporting millions of people, especially those living in rural farming communities. Today, farmers are facing growing challengeslike unpredictable weather, irregular rainfall, and shifting growing seasonsthat make it harder to grow crops successfully. These issues often lead to smaller harvests, financial stress, and inefficient use of land and resources. Traditional farming methods, such as sticking to fixed planting schedules, are no longer reliable in the face of climate change and erratic weather patterns. What farmers need now are smarter systemstools that use both real-time and historical weather data to guide them. These systems can

   help identify the best times to plant and harvest by analysing expected weather trends and past climate conditions. The goal is to reduce crop losses, improve yields, and encourage more sustainable farming. Instead of using one-size-fits-all planting calendars, this approach provides tailored recommendations that adapt to the specific needs of each region and season. Its about helping farmers make better decisionswhen to sow, when to harvestbased on solid data, not guesswork.

   By understanding how weather impacts crops and adjusting farming practices accordingly, farmers can reduce risks and make better use of their resources. This research explores how traditional farming challenges affect crop production and shows how machine learning can offer real, practical solutions. Ultimately, it proposes a smart, data-driven approach that blends age-old farming wisdom with modern technologyempowering farmers, advisors, and policymakers to build a more resilient and productive agricultural future.

2. LITERATURE REVIEW

   Agricultural productivity is highly sensitive to climate change which requires adaptive crop scheduling. Wang Atal. [1] In rice systems of India and Bangladesh, crop calendar optimization was checked. He demonstrated that adjusting the time of sowing and harvesting can reverse the yield deficit inspired by increase in temperature and reduce the use of water, which reveals the importance of the dynamic crop calendar combined with monsoon patterns and heat stress mitigation.

   Weather forecasts and crop calendars have also been incorporated into deep learning models. Kunha and Silva [2] In a pre -season and season crop yield prediction structure, deep learning, crop calendar data and accessible remote sensing/climate information using. The model avoids complex NDVI processing, making it suitable for scalable and developing areas. It clearly uses crop calendar to frame the predicted windows.

   Rajasthan, focusing on India, study in Professorial Computer Science by Jha Atal. [3] Many machines compare Learning algorithms-such as random forest and gradients boosting-agro ITS to predict crop yield using climatic features (rain,

   temperature, weather) and crop-specific data. Results describe strong forecast performances and validate the viability of climate AST -based yield forecast in Indian contexts.

   A comprehensive perspective 2024 comes from Helian Review [4], synthesizing recent applications of AI techniques (ML and DL) for predicting yield in diverse crops and geographical areas. It confirms that agro climatic variables remain primary prophets and explaining these models is an open challenge.

   Extension for global modeling, research based on seasonal climate information [5] shows how can the diversity of the year are predicted from the year are using seasonal climate forecasts. Study crop calendar underlines possible improvements in time when forecasts are integrated into planning systems.

   Finally, early research by Australia implemented the bioecian generative models to connect the monthly temperature and the yield of wheat, effectively capturing non -revolutionary growth reactions during the growing season. This approach can be adapted to estimate optimal planting windows in calendar-based systems [6].

3. DATA AND METHODOLOGY

   1. Data Description

      The Ministry of Agriculture and Farmers Welfare, Government of India, and Agmarkate database are some of the public sources used for data in this study. This study looks at data from the years 2000 to 2020, covering two full decades. It focuses on key staple crops in Indiarice, wheat, and corn and includes yearly information on how much was produced (in tonnes per hectare) along with the average wholesale market prices (in rupees per quintal). To get a better understanding of the bigger picture, climate data such as average temperatures and annual rainfall was also gathered from the Indian Meteorological Department (IMD).

      Bringing these datasets together gave valuable insights into how changing weather conditions have influenced crop productivity over time. Using this information, a combined crop calendar was developed for each crop. This calendar takes into account both yield and climate trends to map out when planting and harvesting typically occurred in the past. More importantly, it helps identify the ideal time windows for planting and harvesting in different agroclimatic zones, making farming decisions more accurate and region-specific. By using this customized calendar to plan their farming practices, farmers can make informed decisions that will increase productivity. A path to improved yield forecasting and risk mitigation in the upcoming crop season is provided by the integration of this adaptive crop calendar with future

      stating analytics. This study highlights how farmers can make more informed decisions, minimize losses, and boost agricultural flexibility by combining TensorFlow-managed neural networks with an adaptive crop calendar.

   2. Data Preparation

      Before building the model, the raw data had to be carefully cleaned and prepared to ensure it was accurate and reliable. All the datasets were combined based on the year and crop, creating a unified structure that made it easier to work with. To handle missing data, time-eries linear projections were used, and in some cases, average values were filled in where needed. Outliersespecially in climate and yield datawere managed using the Interquartile Range (IQR) method to reduce distortion and keep the data realistic.

      Next, a detailed crop calendar was created by studying past sowing and harvesting trends alongside climate factors like rainfall, temperature, and soil moisture. This helped identify the best planting and harvesting windows for different regions. These calendar insights were then transformed into model- friendly features and used as inputs. Aligning the crop cycle with environmental suitability helped the model make more accurate, region-specific recommendations.

      Numerical values were extended between 0 and 1 using minimum-max normalization to guarantee that all the features contributed proportionately. The final dataset was divided into training (80%) and testing (20%) sets in chronological order to mimic a real-world scenario where historical trends are used to predict future results. Gap features such as rainfall and pre- year yield were also included to improve forecast performance and capture transient dependence.

   3. Models Applied

      To predict crop yield when accounting for climate and seasonal variations, the appropriate machine learning models were applied. The purpose of these models is to highlight the meaningful pattern in historical agricultural-climatic data, aligning with a customized crop calendar.

      Regression Models Used: The study used standard regression models that are effective with organized agricultural data to predict crop yield based on climate factors and time-related variables.

      Linear Regression (LR): This model shows a linear relationship between the dependent variable, yield, and several independent features like rainfall, temperature, soil type, and previous yield.

      The general form is:

      Where:

      • Y = Predicted yield or price

      • , , , Independent input variables (e.g., rainfall,

        fertilizer use, prior yield)

      • Intercept term

      • Coefficients learned by the model

      • Error term

        Random Forest Regressor (RF): A cloth-based machine Learning algorithm that builds many decisions trees and outputs the average of their predictions. It helps in catching non-lectured relationships and reduces overfitting, making it effective for complex agricultural data.

        Crop Calendar Integration: A combined crop calendar was constructed by analyzing historical sowing and harvesting dates, aligned with seasonal rainfall and temperature patterns. This calendar was converted into numerical features (e.g., day- of-year of sowing and harvesting) and included as part of the input dataset. These time-aware features allow the regression models to associate yield variations with planting schedules and seasonal trends, enabling context-aware predictions.

   4. Evaluation Metrics

      To evaluate how well each model performed, we used metrics like R-squared (R²), Mean Absolute Error (MAE), and Root Mean Square Error (RMSE). These helped us understand how accurate and reliable the predictions were. To ensure the models could handle new, unseen dataand werent just memorizing the training setwe used K-fold cross-validation. This technique gave us a more balanced view of the model's performance by testing it across different subsets of the data.

   5. Tools Used

   Python was the main tool used for handling the data and building the models. Key libraries included Stats models for time series analysis, Pandas for organizing and cleaning data, Scikit-learn for applying machine learning techniques, and Matplotlib and Seaborn for creating visualizations. This setup allowed us to compare how well modern machine learning algorithms perform compared to traditional statistical methods. The goal was to generate practical insights that can help improve crop calendars and forecast crop yields more accuratelyespecially in the context of Indian agriculture.

4. RESULT ANALYSIS

   The performance of the four forecasting modelsMultiple Linear Regression (MLR), ARIMA, Random Forest, and

   GBoostwas evaluated using our enriched dataset, which includes crop yield, market price trends, climatic variables, and the impact of crop calendar optimization from 2000 to 2020. After model training and testing.

   1. Total crop yield trends

      Figure 4.1: The line shows the yield of the total crop from 2000 to 2020 using the chart. A stable upward trend is clear, suggests gradual improvement in yield – bright due to crop management, agricultural practices and progress in calendar optimization.

   2. Yield Distribution by Crop and Soil Type

      Figure 4.2: shows the percentage share of total yield by soil type. Alluvial and black soils contribute the most, confirming their known fertility advantages.

   3. Adaptive Crop Calendar Optimization Impact

      Figure 4.3: Compares actual and optimized total yields. The enhanced schedule resulted in an average yield increase of 8 12%, showcasing the effectiveness of aligning sowing/harvest times with agro-climatic conditions.

   4. Mobile Application Deployment

   To ensure the practical use of the CNN model, we integrated it into a custom Android application, called the AgriSarthi. The app focused on simple and user adapted. This allows farmers to reach important agricultural information such as crop types, weather conditions, farming areas, rain, fertilizers and pesticides. After submitting the information, the app processes it using the trained CNN model and shows an estimated yield on the screen.

   Figure 4.4: AGRISARTHI – Mobile App Interface for Yield Prediction.

5. FUTURE SCOPE

   An important future of your research is contained in integrating the monitoring of the IOT-based soil and nutrients in the calendar-aware machine learning framework. Deploying field-level sensors-such as soil moisture, pH, temperature, and nitrogen/phosphorus/potassium detectors-can feed real-time inputs in their forecast models. This allows the data-powered prosperity system to update the recommendations of the yield forecasts and calendar dynamically to the middle-session, making the advisory output more accurate and adaptive, as displayed in recent studies that depict IOT-competent predictions that acquire more than 90% of accurate.

   Construction on this sensor Foundation, the reaction of soil moisture with automatic irrigation scheduling represents another promising increase. By connecting the IOT moisture data with irrigation control within its crop calendar model, the system can create a closed-loop adaptation-propagated planting dates and water applications in response to real-time climate and soil condition. Field tests by IRRI in rice

   cultivation have shown that such sensor-based irrigation systems improve water efficiency and improve yield flexibility, which confirm the value of a combination of agricultural culture monitoring with schedule logic.

   Finally, applying a digital twin of integrated form environment with reinforcement learning (RL) can raise your system to a simulation-based planner. A digital twin- a virtual replica of field dynamics- can simulate how sowing dates, soil moisture levels and irrigation strategy changes, affect the results of the crop. RL agents can then optimize this decision location to learn the best planting windows under conditions developing. Review confirms that implementing RL-based framework in form digital twins can improve irrigation, scheduling and decision making in planning in the season.

6. CONCLUSION

   This research introduces a leading closed All Loop Framework that inegrates the prediction of machine learning-power yield with dynamic crop calendar optimization, which is grounded in agricultural and climatic data and calendar-aware scheduling. Unlike pre-studies, which treat forecasts and scheduling as individual functions, your functioning creates an integrated pipeline, where sowing dates inform the prediction of the yield and in turn, refine the approximate result calendar-delient actionable, data-operated planting

   guidance.

   By field-specific model training using localized agro and climatic features and calendar variables (eg sowing weeks), your system supports granular, east-weather regional recommendations that adapt more accurately to local climate references. By incorporating the weekly temporal resolution, your work exceeds the traditional monthly model, which offers a fine crop window insight, which is more helpful to make a field decision.

   Ultimately, this research forecast paves the way for intelligent, climate-smart agricultural systems by combining modeling, dynamic scheduling and clear AI. It determines a solid base for future enhancement such as IOT-based soil sensation, real-time irrigation control, and reinforcement- education-smooth digital twins, all converts to skeletal, durable farming practices that are suited to the changing climate.

7. REFERENCES

1. Wang, iaobo, et al. "Crop calendar optimization for climate change adaptation in rice-based multiple cropping systems of India and Bangladesh." Agricultural and Forest Meteorology 315 (2022): 108830.

2. de Freitas Cunha, Renato Luiz, and Bruno Silva. "Estimating crop yields with remote sensing and deep learning." 2020 IEEE Latin American GRSS & ISPRS Remote Sensing Conference (LAGIRS). IEEE, 2020.

3. Jhajharia, Kavita, et al. "Crop yield prediction using machine learning and deep learning techniques." Procedia Computer Science 218 (2023): 406-417.

4. Jabed, Md Abu, and Masrah Azrifah Azmi Murad. "Crop yield prediction in agriculture: A comprehensive review of machine learning and deep learning approaches, with insights for future research and sustainability." Heliyon 10.24 (2024).

5. Iizumi, Toshichika, et al. "Global crop yield forecasting using seasonal climate information from a multi-model ensemble." Climate Services 11 (2018): 13-23.

6. Brown, Jaclyn N., et al. "Seasonal climate forecasts provide more definitive and accurate crop yield predictions." Agricultural and forest meteorology 260 (2018): 247-254.

7. Kalimuthu, M., P. Vaishnavi, and M. Kishore. "Crop prediction using machine learning." 2020 third international conference on smart systems and inventive technology (ICSSIT). IEEE, 2020.

8. Rao, Madhuri Shripathi, et al. "Crop prediction using machine learning." Journal of Physics: Conference Series. Vol. 2161. No. 1. IOP Publishing, 2022.

9. Elbasi, Ersin, et al. "Crop prediction model using machine learning algorithms." Applied Sciences 13.16 (2023): 9288.

10. Champaneri, M., Chachpara, D., Chandvidkar, C. and Rathod, M., 2016. Crop yield prediction using machine learning. Technology, 9(38).