The global warming has serious impact on our climate. Due to this, the frequency and the intensity of forest fires is increasing. It has shown serious challenges such as the protection of resources, human and wild life, health, and property. This study focuses on developing an artificial intelligence assistive innovative solution for active fire detection in the context of smart cities and vicinities. This paper addresses spectral analysis, detection and classification of active fires and seeing the invisible through smoke and thin clouds. The appealing applications are in urban surveillance, smart cities, future industries, forests and earth observation. The idea is realizable by using an intelligent hybridization of machine/deep learning models and using multi-sensor images (aerial, satellite). For this purpose, we use hyperspectral images (Visible, Near Infra-red (NIR) and Short-Wave Infrared (SWIR)) from AVIRIS aerial and Multi-Spectral Sentinel-2 satellite images. AVIRIS images are 224 spectral bands of wavelengths with a spatial resolution of 15 meters, which varies from 366nm (nanometers) up to 2500nm. However, AVIRIS image studied for their spectral richness of wavelengths not yet completely exploited by machine and deep learning and in SWIR to detect active fires. While, Sentinel-2 image has 13 spectral bands (Visible, NIR and SWIR) with three spatial resolutions (10, 20 and 60 meters). First, we explain and describe the preparation phase of hyperspectral and multispectral image databases of forest fires. These databases contain hyperspectral and multispectral endmembers data of different sites for forest fires. Then, we conduct a spectral analysis from these endmembers to characterize the hyperspectral/multispectral reflectance of active fires to identify the distinct wavelengths for fire detection. We identify the wavelengths that can be used for an effective identification of fire and to see through fires smoke and thin clouds. Onward, the selected feature set is processed by robust machine/deep learning algorithms and their performance is compared for automated identification of fire and invisible vision amelioration. The proposed machine/deep learning method secured an overall test accuracy of 99.1%.