Agriculture is the main source of income for Indian farmers in the developing world. The current agricultural system relies heavily on technology to produce more and earn profits, but agriculture still depends on seasonal aspects. It is very difficult to identify the type and amount of pesticide needed for plant growth. Farmers face several serious challenges, including: (i) adapting to climatic conditions due to soil erosion and pollution; (ii) soil deficiencies, such as lack of potassium, phosphorus, nitrogen and other minerals, can have a negative impact on crop yields; and (iii) given the diversity of soils in the world, farmers make serious mistakes about which crops to grow in which type of soil, which is a very important factor to understand.

As the amount of data keeps adding at an exponential rate, Big Data Analytics is an increasingly critical field that needs such advanced machine learning-based data mining methods to efficiently find patterns. In this study, deep learning architectures, namely Convolutional Neural Networks (CNN) and Fully Connected Neural Networks (FCNN), are evaluated and compared regarding high dimensional feature extractions and classification with traditional Support Vector Machine (SVM) techniques. The implementation of the above-proposed framework was presented by training validated models on a high-dimensional dataset in TensorFlow and PyTorch. Classification effectiveness was assessed using performance metrics of accuracy, precision, recall, and F1-score. A PCA-based visualization was performed to analyze whether each model would extract the features well. Also CNN model has the highest accuracy i.e 93.5% compared to the accuracy of FCNN i.e 89.1 and SVM i.e 85.2 which proves its better hierarchical feature learning. It was also found that CNNs converged faster with 25 epochs, with SVM taking too long to converge and offering bad separability of the features, thus CNN towards FCNN models proved to be more effective for complex pattern recognition tasks for Big Data Analytics. Nevertheless, more research is needed to create computationally viable XAI and hybrid models for their real-world use.