Aloe Vera Plant Leaf Disease Detection and Prediction using IoT and Machine Learning

Abstract

Aloe vera (Aloe barbadensis Miller) plays a vital role in healthcare and cosmetics. It is widely used in Ayurvedic medicine, skincare, nutraceuticals, and health supplements. Traditional manual inspection is time-consuming, labor-intensive, and less accurate, making it unsuitable for large-scale farming. The survey identified three major Aloe vera leaf diseases: leaf spot, aloe rust, and soft rust. This study proposes an Edge AI IoT architecture using deep learning for real-time, reliable disease detection. A comprehensive dataset was created by combining on-edge image acquisition via Edge Impulse Studio with public datasets like Kaggle for greater diversity and reliability. The model was quantised to support real-time low-latency disease detection on edge devices with a significant decrease in model size and computational complexity, with no loss in accuracy. The optimised TensorFlow Lite model was run in real time on a Raspberry Pi 4 B with a camera module for continuous image acquisition with local inference. The platform also enables the data to be sent in real time directly to Google Sheets for remote monitoring and structured data logging, which facilitates timely interventions and supports precision agriculture for Aloe Vera cultivation. The deep learning models AlexNet, VGG16, DenseNet121, InceptionV2, MobileNetV2, ResNet50, and Xception were utilised for the classification task, with additions of operations (such as batch normalisation and dropout) incorporated to improve training stability and prevent overfitting. The improved ResNet50 and Xception models outperformed the originals, while their fusion achieved the highest accuracy of 99.72%, demonstrating excellent generalization and stability. The model achieved strong metrics: precision: 0.9948, recall: 0.9942, and F1-score: 0.9944, demonstrating high reliability. When deployed on Raspberry Pi 4 B, the quantized model delivered real-time inference latency of ~80 90 ms, improving over the original 74.5 ms unoptimized model. newline