Climate-related and sustainability risks have transformed environmental drivers of risk into risk factors for the banks through physical shocks, transition policies, and exposure related to emissions. In this paper, we present a machine learning-based anomaly detection framework aimed to assist banks with managing environmental risk through the identification of abnormal risk patterns that could signal potential emerging environmental stress. A classification experiment was developed with banking exposure variables such as loan exposure and sectoral allocation, and environmental indicators (financed emissions, carbon intensity, physical risk, transition risk) , and risk indicators (ESG score, emissions spike ratio). Various supervised learning models, such as Logistic, SVM, KNN, RF, and GBDT, were tried. Results show that ensemble-based methods outperform baseline detection techniques in the detection of anomalous events. The Random Forest model, in particular, had the best overall performance rates ( without it).0.985), Precision = 0.990, Recall = 0.857, and F1-score = 0.919, and with an ROC-AUC of 0.948; on the other hand, Gradient Boosting had slightly higher recall (0.866), an above-mentioned ROC-AUC (with an equivalent of 0.944). These results underscore the promising role of ensemble tree-based theories to identify anomalies in environmental risk, potentially lending support to machine learning early warning systems for banking crises due to climate-related risk.