In this paper, we investigate the problem of adversarial attacks on logistic regression classifiers employed in weather prediction. We address both poison attacks, involving the injection of malicious data into the training dataset to induce incorrect predictions, and evasion attacks, encompassing the generation of modified inputs that mislead the model's predictions. To counter these attacks, we propose defenses based on adversarial training, a technique that enhances the model's resilience. Adversarial training involves training the model on a dataset comprising both benign and adversarial examples, facilitating its ability to differentiate between the two. Our experiments demonstrate that adversarial training substantially improves the robustness of logistic regression classifiers against poison and evasion attacks. Furthermore, our defenses prove effective against various attack strategies. The results of our research suggest that adversarial training holds promise as a defense mechanism for safeguarding machine learning models used in weather prediction from adversarial attacks.