This paper presents a three-phase deep learning framework comprising (i) multi-modal data acquisition from drones and satellites, (ii) standardized pre-processing including interpolation for missing temporal data, and (iii) CNN-based feature extraction for real-time health classification. This framework relies on a mathematical model based on neural networks that classifies and detects the condition of agriculture, removing the reliance on manual tasks and subjective diagnosis. This paper focuses on three main aspects of our framework: data acquisition, training and prediction. Data is collected using sensors like drones, cameras, and satellite imagery and is pre-processed to filter out noise and improve quality. The training part uses CNN to learn features from the data and become more meaningful. The prediction part of the task classifies, and diagnoses crop health through the trained model using the features. The framework accuracy for crops such as maize, potato, and wheat has been tested and yielded over 90% accuracy. The novelty of this work resides in the development of a multi-modal deep learning architecture that fuses macro-scale satellite imagery with micro-scale drone and IoT sensor data to improve diagnostic reliability. The framework was validated on a multi-source agricultural dataset using a 70% training, 15% validation, and 15% testing protocol. Experimental results demonstrate an accuracy exceeding 90% for staple crops. Using this framework can increase the visibility and quality of information maintained for crop health and improve the decision-making routine of farmers in real time. Additionally, automation of this process can significantly reduce labor costs and increase productivity per crop. Implementing this framework can contribute to precision agriculture and sustainable management practices.
Pal et al. (Sat,) studied this question.