Enhancing Pepper Leaf Disease Detection Using Deep Transfer Learning For Sustainable Agricultural Sector In Ksa

Sustainable agriculture in the Kingdom of Saudi Arabia (KSA) intends to develop farming practices that maintain productivity while promoting economic viability and preserving environmental resources. Moreover, the country invests in advanced technologies, which include vertical farming and hydroponics, to reduce environmental impact and enhance food security. Globally, the pepper crop is one of the leading agricultural products of human food security. However, it is vulnerable to various diseases such as gray leaf spots, powdery mildew symptoms on pepper leaf, blight leaf disease, common rust, fruit rot disease, etc. Usually, farmers identify the disease through visual inspection; however, this has its drawbacks as it is generally time-consuming and inaccurate. Several research workers have previously presented different pepper plant disease classification techniques, mainly deep learning (DL) and image processing approaches. This paper introduces a novel Pepper Leaf Disease Detection using the Optimal Deep Transfer Learning (PLDD-ODTL) technique for the sustainable agricultural sector in KSA. As a preliminary preprocessing stage, the PLDD-ODTL technique initially utilizes an adaptive window filtering (AWF) approach to eliminate the noise in the plant images. Next, the PLDD-ODTL approach involves a feature fusion process comprising three DL approaches: residual neural network (ResNet), VGG-19, and DenseNet models. To enhance the performance of the DL techniques in biological systems modeling, the hyperparameter selection process is done by a hybrid artificial ecosystem fractal optimizer using a chaos game optimization (ACGO) technique. Finally, a deep belief network (DBN) is applied to classify the disease. A series of experiments were conducted to illustrate the enhanced performance of the PLDD-ODTL technique. The experimental results of the PLDD-ODTL technique portrayed a superior accuracy outcome of 99.71% over existing approaches.