In remote sensing (RS), deep learning (DL) has demonstrated exceptional performance. The field of RS is finding that DL is becoming increasingly useful. The proliferation of cloud platforms that grant access to massive amounts of processing power has made the deployment of DL systems as a service feasible. This, however, has given rise to fresh issues with data privacy and security. Several levels of confidentiality are necessary for the RS data required to teach DL algorithms. High-resolution satellite photos relating to public safety, for example, must be treated with the utmost secrecy. The owner of a satellite picture may also refuse to allow anybody else to use it on the grounds that it is copyright protected. Consequently, privacy-preserving deep learning (PPDL) methods might be the answer. To train DL without exposing the original plaintext, PPDL allows for the use of encrypted data. In order to classify objects in very high-resolution satellite photos, a hybrid PPDL method is proposed in this research. For privacy-preserving to be designed, existing deep learning models and algorithms must be rethought and redesigned from the ground up to meet the stringent technological and algorithmic restrictions of HE. This study presents a primer on the topic of privacy-preserving convolutional neural networks (CNNs) and a technique for their creation. High-resolution photos captured by the SPOT6 and SPOT7 satellites were used in the experiments. Using the suggested encryption approach resulted in a loss of classification accuracy of 2-3.5 percent.