Issuu

International Research Journal of Engineering and Technology (IRJET)

e-ISSN: 2395-0056

Volume: 12 Issue: 10 | Oct 2025

p-ISSN: 2395-0072

www.irjet.net

Deep AgroLens: A Multimodal Edge-AI Framework for Early Detection, Prognosis, and Intelligent Management of Orange Plant Diseases Mr. Dilkhush Tikale¹,Mr. Sahil Raut², Mr. Aditya Warudkar³, Mr. Aniket Suryawanshi⁴, Mr. Shubham Shile⁵, Dr. Sushma Telrandhe⁶ ¹²³´µ¶ Department of Computer Science and Engineering, Guru Nanak Institute of Engineering and Technology, Nagpur, India ---------------------------------------------------------------------***------------------------------------------------------------------Abstract - Plant diseases severely compromise the productivity and economic value of orange cultivation, often relying on timeconsuming manual diagnosis. To address the limitations of conventional, single-modality methods, we propose Deep AgroLens, a novel Multimodal Edge-AI framework for the early detection and proactive management of citrus diseases. This system integrates multimodal data— combining high-resolution visual analysis with crucial environmental sensor inputs (VOC emissions, soil pH, temperature, and humidity). A hybrid deep learning architecture, leveraging a Vision Transformer (ViT) coupled with a CNN, extracts robust features, which are then fused at a decision level to enable accurate detection and precise severity staging (early, medium, severe). Furthermore, a time-series forecasting module uses LSTM networks to predict future disease risk based on environmental factors, shifting management from reactive to proactive. The complete system is optimized for edge deployment using pruning and quantization, ensuring reliable, low-latency, and offline functionality in rural environments. Transparency is achieved via Explainable AI (XAI) using Grad-CAM visualization. Deep AgroLens provides a scalable and sustainable solution, delivering real-time treatment recommendations through a multilingual farmer advisory interface. Key Words: Deep Learning, Edge-AI, Multimodal Fusion, Vision Transformer (ViT), IoT, Orange Disease, Explainable AI (XAI), Precision Farming

1.INTRODUCTION The agricultural sector serves as a fundamental pillar of global economic stability and food security. Within this domain, citrus cultivation, particularly oranges, holds significant economic value worldwide. However, orange crops are highly vulnerable to various destructive pathogens, including Citrus Canker, Greening (HLB), Melanose, and Black Spot. When left unchecked, these diseases can lead to catastrophic yield losses, impacting farmer livelihoods and entire supply chains. The conventional method of disease management—relying on manual inspection by agricultural experts—is inherently subjective, slow, and nonscalable, resulting in delayed diagnosis and ineffective containment strategies [1]. The integration of Artificial Intelligence (AI) and the Internet of Things (IoT) presents a paradigm shift towards Precision Farming. Existing AI-based disease detection systems predominantly utilize image processing and Convolutional Neural Networks (CNNs) to classify visual symptoms. While effective, these single-modality systems suffer from crucial limitations: they are reactive (only detecting visible symptoms), they often fail to capture the subtle early-stage chemical markers of infection (like Volatile Organic Compounds or VOCs), and they typically require powerful cloud-based resources for operation [2, 3]. This research introduces Deep AgroLens, a novel Multimodal Edge-AI framework designed to provide comprehensive, proactive, and intelligent management of orange plant diseases. This work makes the following significant contributions to the field: 1.

Multimodal Data Fusion: Development of a robust fusion strategy that combines deep visual features (extracted via a hybrid CNN-Vision Transformer (ViT) architecture) with non-visual environmental and biochemical data (VOC, pH, Temp/Humidity sensors) for superior early-stage detection.

Proactive Prognosis: Integration of a time-series forecasting module utilizing Long Short-Term Memory (LSTM) networks to predict future disease risk based on historical and real-time environmental trends, enabling true preventive intervention [4].

Edge-AI Optimization: Implementation of model compression techniques, specifically pruning and quantization, to ensure the complete system operates with high efficiency, low latency, and reliability on resource-constrained Edge devices (e.g., Raspberry Pi or mobile platforms) for offline use in rural settings [5].

Interpretability: Inclusion of an Explainable AI (XAI) module, using Grad-CAM, to provide visual evidence of the model's decision-making process, thereby building farmer trust and improving the system's overall utility.

Impact Factor value: 8.315

ISO 9001:2008 Certified Journal

Page 560