Reversible image steganography is a technique for hiding covert communications under image coverings with the goal of restoring the hidden messages' original status. It is commonly used for secret communication and data storage. However, present approaches fall short due to a trade-off between security and payload capacity. This study developed a Hybrid Deep Learning Model (HDLM) for reversible image steganography. The HDLM was developed using three models: cover selection, encoding, and decoding models. The Microsoft Common Objects in Context image dataset was used to train the HDLM. Precision, F-score, and recall were used to evaluate the performance of the cover selection model. The payload capacity, Peak Signal-to-Noise-Ratio (PSNR), and Structural Similarity Index Measure (SSIM) were used to evaluate the performance of the encoding model. Mean Squared Error (MSE), PSNR, and SSIM metrics were used to evaluate the performance of the decoding model. The security of the developed HDLM was tested using the StegoExpose steganalysis tool. The results showed that the precision, F-score, and recall values for the cover selection model were 98.11% ± 0.45%, 97.19% ± 0.47%, and 96.29% ± 0.52% respectively. The performance values for the encoding model produced payload capacity of 24.85 bpp ± 0.32, PSNR of 41.44 dB ± 0.78, and SSIM of 0.97 ± 0.02, while the performance values for the decoding model yielded PSNR of 43.98 dB ± 0.65, SSIM of 0.92 ± 0.03, and MSE of 0.025 ± 0.005. The security test showed that StegoExpose can only detect correctly that 142 ± 25 out of 10,000 stego-images are suspicious. The HDLM has the highest payload capacity, PSNR, and SSIM with values of 24.85 bpp ± 0.32, 41.44 dB ± 0.78, and 0.97 ± 0.02. The study concluded that the developed HDLM performs better than the previous models regarding payload capacity and security.