Publication

Abstract : In this study, we present a strong and intelligent structure aimed at increasing the accuracy of the prediction of crop yield by effectively integrating several sources of data. Our approach combines high-resolution multisial and RGB satellite imagery with the parameters of the necessary agricultural-killing, which creates a broad, data-manual system. Unlike traditional models, which mainly depend on spectral data analysis, our method takes advantage of a deep learning-based, multi-sensor fusion strategy that not only explains the spectral pattern, but also captures hidden agricultural symptoms such as firmian structure, vegetative indices and partial vegetation cover. To achieve this, we designed a hybrid model architecture, which includes the Satellite images for crop growth dynamics over time and a temporary feature included the Convolutional Neural Network (CNN’s) to extract spatial features from a temporary feature aggregator (TFA). This hierarchical pipeline systematically learns both visual and temporary characteristics of crops, resulting in a highly accurate yield estimate process. Our model demonstrated excellent performance, achieved the R2 Results score of 0.95 and exceeded the accuracy of existing traditional models to reduce the average absolute percentage error (MAPE) by 5.9%. This research forms a meaningful intersection between computer vision, remote sensing and agricultural science, which contributes to the development of durable, scalable and accurate crop forecasting systems that can benefit farmers, researchers and policy makers equally.