Real-time & on-field crop imagery has illumination variations, specular reflections and shadows from neighbouring objects. To efficiently segment such images, a wide variety of deep learning models are proposed by researchers. But these models are either highly complex, or do not take into consideration multiple noise effects during segmentation operations. Moreover, the classification efficiency of such models is limited by their feature representation and classification performance, which limits their scalability when applied to real-time scenarios. To overcome these issues, this text proposes design of a Generative Adversarial Network (GAN) based segmentation model for efficient disease identification & severity estimation in apple crops. The model initially collects large datasets tagged with disease types & their severity levels. These datasets are used to train the Generator Network that assists in identification of noise types. These noise types are removed by the Discriminator Network via loss reduction processes. Segmented images are given to a Long-Short-Term Memory (LSTM) & Gated Recurrent Unit (GRU) based feature representation model, which assists in estimation of highly variant feature sets. These feature sets are used to train a Recurrent Neural Network (RNN) for estimation of various disease types. Due to which diseases like bacterial blight, scab, fungus, etc. can be efficiently estimated under real-time scenarios. The identified images are further processed via a GoogLeNet based Convolutional Neural Network (CNN), which assists in severity estimation for each of the disease types. Multiple severity-level networks are trained in order to improve the estimated performance for different disease types. Due to which the proposed model is able to improve disease classification accuracy by 8.5%, while improving classification precision by 4.3% under different scenarios. The model was also able to improve severity estimation accuracy by 9.4%, and severity estimation recall by 8.3% when compared with state-of-the-art deep learning techniques. Due to which, the proposed model is capable of deployment for real-time scenarios.