<p>Soil erosion, exacerbated by climate change and urbanization, poses a significant environmental threat, leading to land degradation, water pollution, and ecosystem disruption. Accurate soil erosion prediction is critical for developing effective control and prevention methods. While existing deep-learning approaches have shown promise, they are often limited by high error rates, reduced accuracy, high time complexity, overfitting, and misclassification. To address these limitations, this paper proposes a novel Self-Channel Attention-based chaotic atrous convolutional biLSTM Network (SCA-CAConvBiLSTM) for accurate soil erosion prediction. The proposed multi-phase approach involves data pre-processing, feature extraction, feature selection, prediction, and optimization. To improve data quality, the pre-processing stage uses missing value imputation and Z-score normalization. To minimize dimensionality issues, features are extracted using the ResNet-152 model, and feature selection occurs using the improved Mantis Search Optimization (IMSO) algorithm. The proposed SCA-CAConvBiLSTM model categorizes soil erosion into four categories: landslides, gullies, bad land, and no gully/bad land. It uses self-channel attention to focus on relevant data aspects and improve prediction accuracy. The Circle Chaotic Crayfish Optimization algorithm&#xa0;is used to fine-tune the model parameters and reduce the loss function. The results show that on the ESA EO4SoilProtection 2024 Predicting Erosion CAT dataset, the proposed method outperforms existing soil erosion classification models with an accuracy of 99.157%.</p>