Forest fires pose a significant threat to ecosystems, especially in regions like Sikkim, where climate conditions increase the frequency of forest fire. Accurate prediction and assessment of forest fires is essential to mitigate damage to forest land and wildlife. However, existing research has not adequately combined meteorological data with land imagery for forest fire prediction in this region. This study proposes a novel Kernelized Deep Convolutional Parametric Modswish Neural Network (KDCPMNN) model for forest fire prediction using satellite images and meteorological data. Initially, satellite images are segmented using the K-Means (KM) algorithm, followed by pre-processing with Ani-sotropic Exponential Diffusion Filter (AEDF) and Bayesian Contrast Limited Adaptive Histogram Equalization (B-CLAHE). The Dark Channel Prior (DCP) and Transmission Map (TM) are then used to calculate intensity values and depth information, respectively, from which features are extracted. Simultaneously, vegetation details and attributes from meteorological data are also extracted. These combined features are input into the KDCPMNN to predict the likelihood of a forest fire. In the event of a fire, the burn area is assessed using the Composite Burned Index (CBI) within a Secant Linear Membership-based Fuzzy Inference System (SLM-FIS), while the fire's potential is evaluated through confidence score. Experimental results show that the proposed KDCPMNN model achieves 98.2% accuracy, surpassing existing methods.