A careful analysis of an image processing-based disease identification method is required for detecting, identifying, and treating pumpkin plant disease before it causes damage to the entire plant, Because diseases in pumpkin plants cause significant production and economic losses, as well as a reduction in the quality and quantity of agricultural products. Using a laboratory and naked eyes of observation, plant pathologists can detect pumpkin disease. This may take a long time and result in a loss of extra outflow. This paper describes an automatic method for identifying pumpkin leaf diseases that affect the leaf part, as well as segmentation and hyper parameter value selection techniques for identifying the two types of pumpkin leaf diseases from normal. Different hyper parameter values are used in this paper to speed up and improve the system's accuracy. We also used two different image noise removal techniques, which are Median and Gabor. We perform an experiment for each of the hyper parameter values, noise removal techniques and find the best one. Our system is designed by incorporating four mechanisms: preprocessing (image resizing, noise removal, and JPG conversion). We normalize the image into (256,256), (227,227), (224,224), and (192,192) and (128,128) pixel sizes, but (224,224) produces higher accuracy. Segmenting the region of interest from the given image by using watershed. Finally, we used deep CNN for feature extraction and a three-way SoftMax to classify pumpkins into three categories: pumpkin powdery mildew, downy mildew, and normal pumpkins. Hyper parameter values involved in our model are selected by comparing and contrasting using our pumpkinNet model. The dataset was collected from the west go jam zone, particularly the south and north of Achefer. The system was tested and evaluated using sample images, with 80 percent of the dataset being used for training and 20 percent for testing. Finally, the model achieved 98.2% training accuracy and 98.1% testing accuracy. Keywords: Pumpkin disease, Convolutional Neural Network, Feature learning, Watershed segmentation, SoftMax, hyper parameter values.