Agriculture is the backbone for sustainability of any country and plants has a vital role to play in biodiversity sustenance. Crop yield is highly correlated with plant health. Early detection of diseased plant can reduce the adverse effect on healthy plant. Plant leaf is the primary component to identify the abnormality in a plant. Plant leaf images captured by advanced digital cameras can be passed to an advanced computer added system for automated detection of diseased plant at very early stage for fast response. Performing the same process by human being for individual plant is an inefficient and time consuming process which may lead to spreading the disease in whole crop field. Availability of high resolution and GPS enabled digital cameras and advancement in image processing techniques can be utilized to overcome this challenge of early detection of plant disease through plant leaf. The features extracted from image processing tool will be passed a pre-trained deep learning Convolutional Neural Network (CNN) based models for recognizing and classifying the plant disease. Transfer Learning approach is used to increase the efficiency and accuracy of the proposed system. Data augmentation and data balancing techniques are also employed to overcome the overfitting issue. Additionally, the performance of transfer learning approach has been improved in significant manner after adopting efficient pooling and optimization technique. Total 17820 images of different plant leaves (healthy and un-healthy) are used to train and validate pre-trained CNN models. ROC (Receiver Operating Characteristic) curve and other statistical parameters, including specificity, sensitivity, recall, precision and accuracy was applied to compare the performance of various pre-trained model used in transfer learning. The results are clearly indicates that AUC (Area Under Curve) values for implemented models are high and approaching to 0.942. Inclusion of efficient pooling strategy and optimization technique has increased the accuracy by 4-5%. Initially the was ranging from 92% to 95% but after adapting pooling and optimizer the accuracy enhanced to 95%-98%.