A fusion transfer learning framework for intelligent pest recognition in sustainable agriculture.

With the fast growth in the population of the world, there is a constantly increasing requirement for sustainable food supplies. Agriculture is the backbone of the global food supply, with vegetables and fruits being essential for a balanced intake. Still, in recent years, the global distribution of malignant plant pests and illnesses has given rise to significant losses in the quality and yield of crops. Thus, automated detection of plant diseases is vital in monitoring huge areas of crops and automatically recognizing disease and insect pest marks immediately after they develop on plant leaves. Smart agriculture is a new field that uses artificial intelligence (AI) methods, wireless communication, and information technologies, such as the Internet of Things (IoT), to improve farming practices and achieve accurate control of fertilization, crop illnesses, and plant pests in an agricultural field. This study proposes an Insect Pest Recognition Model for Enhancing Food Production Using a Heuristic Optimiser and Fusion Transfer Learning (IPRMEFP-HOFTL) model in smart farming solutions. The aim is to provide effective automatic detection of plant diseases, which will help monitor vast fields of crops on plant leaves. Initially, the Wiener filtering (WF) method is utilized for image pre-processing to enhance image quality by removing noise, and then data augmentation is done. Furthermore, the feature extraction process is performed by the fusion models, namely CapsNet and Xception. For the classification process, the denoising autoencoder-long short-term memory (DAE-LSTM) method is implemented. Finally, the multi-objective remora optimization algorithm (MOROA)-based hyperparameter selection model is carried out for optimizing the detection outcomes of the DAE-LSTM method. Wide-ranging experiments were conducted to prove the performance of the IPRMEFP-HOFTL approach under the IP102-dataset. The comparison study of the IPRMEFP-HOFTL approach illustrated a superior accuracy value of 98.22% over existing techniques.