Context. Astrophysical observations must be corrected for their imperfections of instrumental origin. When charge-coupled devices (CCDs) are used, their dark signal is one such hindrance. In their pristine state, most CCD pixels are cool, that is, they exhibit a low quasi-uniform dark current, which can be estimated and corrected for. In space, after having been hit by an energetic particle, pixels can turn hot, viz. they start delivering excessive, less predictable, dark current. The hot pixels therefore need to be flagged so that a subsequent analysis may ignore them. Aims. The image data of the PICARD-SODISM solar telescope require dark signal correction and hot pixel identification. Its E2V 42-80 CCD operates at -7.2 °C and has a frame-transfer architecture. Both image and memory zones thus accumulate dark current during integration and readout time, respectively. These two components must be separated in order to estimate the dark signal for any given observation. This is the main purpose of the dark signal model presented in this paper. Methods. The dark signal time-series of every pixel was processed by the unbalanced Haar technique to timestamp when its dark signal changed significantly. In-between these instants, the two components were assumed to be constant, and a robust linear regression, with respect to integration time, provides first estimates and a quality coefficient. The latter serves to assign definitive estimates for this pixel and that period. Results. Our model is part of the SODISM Level 1 data production scheme. To confirm its reliability, we verified on dark frames that it leaves a negligible residual bias (5 e -) and generates a small rms error (25 e- rms). We also examined the distribution of the image zone dark current. The cool pixel level is found to be 4.0 e- pxl-1 s-1, in agreement with the predicted value. The emergence rate of hot pixels was investigated as well. It yields a threshold criterion at 50 e- pxl-1 s-1. The growth rate is found to be on average ~500 new hot pixels per day, that is, 4.2% of the image zone area per year. Conclusions. A new method for dark signal correction of a frame-transfer CCD operating near -10 °C is described and applied. It allows making recommendations about the implementation and scientific usage of such CCDs. Moreover, aspects of the method (adaptation of the unbalanced Haar technique, dedicated robust linear regression) have a generic interest. © ESO, 2013.