Advanced Crop Health Monitoring for Smart Agriculture Using Fusion of Transfer Learning With Attentive Convolutional Recurrent Neural Network on Remote Sensing Images

Climate change and variability, highlighted by increasing temperatures, altered precipitation, and extreme weather, have significant effects on agriculture. A suitable crop assessment at an enhanced level needs observing extensive areas by a powerful method. Remote sensing (RS) skill provides this over non-destructive synoptic screening abilities. The technology of RS has revolutionized the farming sector by providing vital insights into resource allocation, environmental sustainability, and crop management. RS provides a cost-effective and non-invasive means of observing crops throughout their lifecycle. Early recognition allows for targeted interventions, decreasing crop damage and diminishing the requirement for chemical inputs. It also helps in monitoring an environment. In the era of digital data propagation, agriculture is poised for a transformative innovation driven by deep learning, utilizing RS imagery. In this manuscript, a Robust Crop Health Assessment Using Improved Fusion Neural Network with Remote Sensing Imagery (RCHA-IFNNRSI) approach is proposed. The aim is to develop an integrated crop health assessment framework using RS and leaf image analysis for enhancing precision and decision-making in smart agriculture. In the initial image pre-processing stage, the bilateral filter (BF)-based noise elimination and contrast enhancement is performed to enhance the raw images by removing noise. For feature extraction, the RCHA-IFNNRSI approach integrates Swin Transformer (ST) and ResNet50 models to create a fused backbone that effectively captures both local and global features for crop health assessment. Finally, the attentive convolutional recurrent neural network (ACRNN) technique is employed to detect and classify crops into diverse health status classes automatically. Extensive experimentation of the RCHA-IFNNRSI technique on the MH-SoyaHealthVision dataset illustrated superior performance, achieving a high accuracy of 98.97% compared to existing models.