Deep Learning for Plant Disease Detection and Crop Yield Prediction based on NPP-WPF Analysis in Smart Agriculture

• Main Authors: S, Amudha, G, Bindu, Sankar, Sasi Rekha, Poongothai, E, K, VijayaLakshmi, Sarveshwaran, Velliangiri
• Format: Conference Proceeding
• Language: English
• Subjects: and Potassium-K) Sensor; Arduino Uno; Cameras; Crop Yield; DHT (Digital Humidity and Temperature sensor) 11 Sensor; Flask Frame Work; Humidity; Node MCU (micro-controller unit); NPK (Nitrogen-N; Phosphorus-P; Plant diseases; Smart agriculture; Sociology; Soil; Temperature sensors
• ISSN: 2768-0673
• DOI: 10.1109/I-SMAC58438.2023.10290468

Despite a sharp rise in population, agriculture still provides food for all people and acts as a backbone of the country. Nowadays, no proper guidance given to farmers and they are not able to analyze the nature of soil, disease of plant and its requirements. This leads to the depletion of farming in future. Crop diseases must be identified and prevented in order to increase the productivity. To overcome this, the nature of the soil is analyzed through various parameters like Nitrogen-Potassium-Phosphorus Content in soil, Humidity, Temperature, pH value of the soil, Rainfall using sensors, plants disease using cameras in real-time environment. These parameters are obtained using various IoT Sensors, cameras and existing Data, are analyzed through ML Algorithms. The proposed system consists of two modules in crop yield prediction system using the features NPP-WPF whereas NPP stands for Nitrogen-Potassium-Phosphorus and WPF stands for Water Content-Pesticide-Fertilizer, the Farmers can be guided with the details of the crop that is suitable for the soil and about Fertilizer, Pesticide, Water requirements. The second module called as automatic disease detection system uses CNN, ResNet, VGGNet, and Mobile net models to recognize and diagnose plant diseases from their leaves at the early stage. The proposed model is implemented with the dataset of 14 different species of plants and with the categorical classes of 38 classes. The result analysis shows that the suggested model provides accuracy of 98.2% when compared to the existing state-of-the-art methodologies.