By enabling machine learning on ultra-low-power microcontrollers, Tiny Machine Learning (TinyML) is transforming real-time environmental monitoring. However, implementing models on edge devices with limited resources requires optimization strategies that strike a balance between energy consumption, accuracy, and efficiency. This study assesses how well TinyML models used for environmental monitoring tasks including air quality assessment, forest fire detection, and water quality analysis perform when model compression approaches like pruning, quantization, knowledge distillation, and low-rank factorization are used. According to experimental results, quantization maintains 87.2% accuracy while reducing model size by 75% (from 512 KB to 128 KB), which makes it ideal for applications with limited memory. Pruning is successful for low-power inference, as evidenced by its ability to reduce model parameters by up to 90% with only a little loss of accuracy ( 92%). Knowledge distillation reduces computing complexity while maintaining prediction performance by effectively transferring knowledge from large models to smaller ones. The results show that hybrid approaches can further enhance TinyML deployment in real-world environmental applications, even though each strategy has trade-offs. Future research into adaptive compression techniques and federated learning is necessary to improve the robustness and scalability of TinyML in environmental sustainability because of issues including hardware limitations, dataset variability, and real-world deployment variables.