Recent technological advances in light-sheet microscopy make it possible to perform whole-brain functional imaging at the cellular level with the use of Ca2+ indicators. The outstanding spatial extent and resolution of this type of data open unique opportunities for understanding the complex organization of neuronal circuits across the brain. However, the analysis of this data remains challenging because the observed variations in fluorescence are, in fact, noisy indirect measures of the neuronal activity. Moreover, measuring over large field-of-view negatively impact temporal resolution and signal-to-noise ratio, which further impedes conventional spike inference. Here we argue that meaningful information can be extracted from large-scale functional imaging data by deconvolving with the calcium response and by modeling moments of sustained neuronal activity instead of individual spikes. Specifically, we characterize the calcium response by a linear system of which the inverse is a differential operator. This operator is then included in a regularization term promoting sparsity of activity transients through generalized total variation. Our results illustrate the numerical performance of the algorithm on simulated signals; i.e., we show the firing rate phase transition at which our model outperforms spike inference. Finally, we apply the proposed algorithm to experimental data from zebrafish larvae. In particular, we show that, when applied to a specific group of neurons, the algorithm retrieves neural activation that matches the locomotor behavior unknown to the method.