Betel Leaf Disease Detection Using Deep Learning

Abstract

Betel leaf is a crop of economic importance in South Asia and field diagnosis is slow, subjective and constrained by the availability of expert opinion of the foliar disease. This thesis constructs and analyze an effective image based system that can determine the most common conditions of betel leaves through deep learning. The classifier divisions four categories of healthy classifications as of Leaf Burn Disease, Leaf spot disease, and Pest Damage, directly out of single photos taken in the real world environments. A selected collection of 2,149 original photographs was compiled and divided into training (1,522), validation (215), and testing (412) groups to make it possible to provide rigorous testing. The methodology is based on a transfer-learning plan and six popular ImageNet-pretrained convolutional neural nets, which would be MobileNetV2, VGG16, ResNet50, InceptionV3, EfficientNetB0, and DenseNet121. Both backbones are fed by a single classification head that consists of global average pooling, 512-unit ReLU input with dropout, and a four-output softmax. Standard data augmentation (rotation, translation, zoom, and horizontal flip), label smoothing to enhance calibration, early stopping to avoid overfitting, and adaptive learning-rate scheduling to stabilize training are all used. The last implementation was implemented into Google Colab with reference to the local CPU/GPU resources, which made it reproducible with relatively modest hardware requirements. Out-of-sample experimental performance indicates good performance of various models. VGG16 attained 95 percent accuracy, DenseNet121 94, and InceptionV3 92. MobileNetV2 was third with 71 percent and ResNet50 with 64 percent and EfficientNetB0 followed with 37 percent. Analysis by classes demonstrates a steady high recall of Healthy and Pest Damage, with the Leaf Burn recall being moderately lower as an artistic phenomenon of not being able to see the necrotic spots and pest scarring even though there were variable field conditions. Visualizations of grad-cam ensure the models also prioritize lesion characteristics and auxiliary borders, not the clutter of the background, and hence interpretability and practical trust. A methodological audit shows that performing a single decision on all the backbones with a rescaling can conflict with model specific preprocessing which may have contributed to the EfficientNetB0 collapse which should be further ameliorated by adding simple adjustments, such as per-backbone preprocessing, to model specific constructions, and by adding 299x299 inputs InceptionV3. This work has tripled contributions: crop-specific, condition of field status, betel leaf datasets; a consolidated, open data pipeline comparing six robust backbones; and a statement of realistic suggestions that resourcefully impact the performance and philosophy in bedside. It is appropriate in the pilot tool the advisory tool and extension services of farmed systems, where it can be deployed using TensorFlow Lite, ONNX, and optional leaf segmentation, ensuring there is minimal impact of the backgrounds. On the whole, this thesis shows that a well-considered transfer learning, combined with human-readable results, can provide a dependable and cost-effective answer to solving the problem of detecting betel leaf diseases and is a step along the way to apply the method to other crops and areas.