Introduction
Precision Agriculture (PA) leverages data-driven insights to optimize farm management and resource usage. However, traditional PA systems often rely on stable internet connectivity and cloud computing, making them impractical for developing regions with low internet penetration. This paper proposes an end-to-end, offline-capable edge computing framework designed to monitor crop health and optimize irrigation without requiring a persistent cloud connection.
System Architecture and Methodology
The proposed system utilizes a multi-layered approach that integrates ground-based IoT sensors with aerial drone imagery, all managed by local edge hardware.
Hardware Integration
A Raspberry Pi 3 serves as the central "parent node," coordinating several NodeMCU units. This edge module is equipped with sensors to monitor critical environmental and soil parameters, including moisture levels, pH, nutrient content, and temperature.
Irrigation and Yield Prediction
To optimize water usage, the authors implement Partial Least Square Regression (PLSR) to predict crop water requirements based on a customized evapotranspiration equation ($ET_c = K_c \cdot ET_0$). This allows for precise moisture delivery tailored to specific growth stages.
For aerial monitoring, drones equipped with RGB sensors perform land mapping. The system employs a fine-tuned YOLOv4 model for crop counting and introduces the Stratified Pixel Analysis Classification (SPAC) algorithm. SPAC segments crops from drone imagery to identify harvest-ready regions, facilitating localized intervention and reducing the need for heavy, non-targeted machinery that can cause soil compaction.
Model Optimization for the Edge
Running complex deep learning models on low-power hardware like a Raspberry Pi requires significant optimization. The authors utilized post-training quantization and explored the Lottery Ticket Hypothesis. This technique identifies smaller "winning ticket" subnetworks within larger models that maintain high performance while significantly reducing computational latency and memory footprint.
Datasets and Validation
The study combined public datasets (historical Indian crop yield and rainfall data, and the Global Wheat Head dataset for pre-training) with custom data. A specific dataset of cabbage crops from farm sites in Nagalpur, Gujarat, was used to fine-tune the YOLOv4 algorithm. Real-time validation was conducted using live feeds from the deployed IoT sensor network.
Key Results
The system demonstrated strong performance in real-world agricultural settings:
- Yield Prediction: Among the tested models, the Long Short Term Memory (LSTM) network achieved the best results with a Mean Absolute Error (MAE) of 22.17, outperforming standard RNNs.
- Crop Counting: The fine-tuned YOLOv4 model achieved a precision of 0.8241 and an F1-score of 0.8042 for identifying ripe cabbage crops.
- Efficiency: The combination of pruning and quantization enabled real-time processing and irrigation control on a standard Raspberry Pi 3.
Conclusion
This research highlights the potential for high-performance AI to operate in resource-constrained environments. By prioritizing an "offline-first" architecture and employing model compression techniques, the authors provide a scalable solution for precision agriculture in regions where cloud dependency remains a barrier to technological adoption. Future work will focus on generalizing the system for diverse crop types and optimizing the power autonomy of the edge nodes.
Summary updated