This paper deals with wildfire identification in the Alaska regions as a semantic segmentation task using support vector machine classifiers. Instead of colour information represented by means of BGR channels, we proceed with a normalized reflectance over 152 days so that such time series is assigned to each pixel. We compare models associated with $𝓁1$-loss and $𝓁2$-loss functions and stopping criteria based on a projected gradient and duality gap in the presented benchmarks.