<p>The Climate changes, resource depletion, and decreased efficiency impact agriculture, making the use of smart and data-based agriculture solutions decisive. Although sensors and drones in smart farming provide real-time data, current machine learning models may not be effective at integrating such varied information. This study proposes a new hybrid framework by combining Recurrent Neural Networks (RNNs), Vision Transformers (ViTs), and a fusion transformer to effectively process multimodal data. A deep learning framework for IoT-driven farming helps farmers make accurate decisions by collecting and processing various agricultural data. It all starts by joining IoT sensor information (such as temperature, humidity, soil moisture, and pH) with pictures from drones or cameras taken just in the soil. This mixture of different information represents the actual issues farmers face daily.</p><p>It gets information by using two sources namely IoT sensors to track environmental data over time which are used to predict droughts and images raised by cameras to find any issues with crops, pests or insufficient nutrients. Through the Temporal-Spatial Feature Fusion (TSFF) architecture, the model is accomplished of discovery relationships between different aspects of time and space, supporting reliable analysis and prediction in agriculture. According to results, the model provides an accuracy of 99.02%, precision of 98.87%, recall of 98.91%, and an F1 score of 98.89% when tested with the Dry Bean and Soil Type datasets.</p>