Interpretable Deep Learning Framework for Land Use and Land Cover Classification in Remote Sensing Using SHAP

Aim:

      To develop an enhanced LULC classification system using ResNet50v2 for better accuracy and LIME for explainability, while minimizing computational resource requirements.

 Abstract:

         This paper proposes an enhanced deep learning framework for Land Use and Land Cover (LULC) classification in remote sensing. Unlike traditional CNN-based approaches, our system employs the ResNet50v2 model to achieve a classification accuracy of 95%. Furthermore, to improve interpretability, we integrate Local Interpretable Model-agnostic Explanations (LIME) to provide explainability for the model’s predictions. By utilizing 20,000 satellite images of size 64×64, the proposed system significantly reduces memory usage while maintaining high accuracy. The results demonstrate the effectiveness of ResNet50v2 and LIME in improving classification in complex urban and rural areas.

Introduction:

          Land Use and Land Cover (LULC) classification plays a critical role in environmental monitoring, urban planning, and agricultural assessment. Accurate classification of satellite imagery into different land use categories is essential for decision-making processes. Existing approaches rely on Convolutional Neural Networks (CNNs) for classification. However, CNNs often lack explain ability and demand significant computational resources. Our proposed system aims to overcome these limitations by using ResNet50v2, which offers higher accuracy with fewer parameters, and LIME, which provides visual explanations of model predictions.

 Existing System:

      The existing system utilizes a compact Convolutional Neural Network (CNN) model for LULC classification and employs SHAP (Shapley Additive Explanations) to explain the classification results. The system processes Sentinel-2 satellite images containing 27,000 images of size 64×64 using three-band combinations from the available 13 spectral bands, reducing the model’s input data by 77%. Experimental results on the EuroSAT dataset achieved an overall accuracy of 94.72% with improved classification performance for urban and rural areas. However, despite the accuracy, the system requires significant memory and computational resources due to the larger input data and complex explainability mechanism.

 Disadvantages of Existing System:

1. High computational resource requirements due to large input data.
2. Limited interpretability due to reliance on SHAP, which can be computationally expensive.
3. Inability to achieve higher classification accuracy beyond 94.72%.

Proposed System:

    The proposed system introduces a ResNet50v2-based classification framework for LULC classification with enhanced accuracy and reduced resource requirements. The system also incorporates LIME (Local Interpretable Model-agnostic Explanations) to visualize and explain model predictions. Unlike the existing system, the proposed framework processes only 20,000 images of size 64×64, thereby reducing memory usage while achieving a classification accuracy of 95%.

 Algorithm:

         The deep learning process for LULC classification starts with data preprocessing, where satellite images of size 64×64 are normalized and augmented to improve generalization. Next, the ResNet50v2 model is utilized as the core architecture, leveraging pre-trained weights from ImageNet to extract deep features from the images. The use of skip connections in ResNet50v2 ensures smooth gradient flow and improves learning efficiency. During the training phase, the model is trained using cross-entropy loss and optimized using the Adam optimizer, with early stopping employed to prevent overfitting. Once trained, the prediction stage involves classifying input images into various LULC categories with high accuracy. To ensure explainability, the system employs LIME (Local Interpretable Model-agnostic Explanations) to generate visual explanations for individual predictions, highlighting the most influential regions of each image that impact classification decisions.

 Advantages:

        The proposed system offers several significant advantages over existing methods. One of the most notable improvements is the increase in classification accuracy to 95%, which surpasses traditional CNN-based models. By processing only 20,000 images instead of 27,000, the system achieves a reduction in memory usage and computational cost, while still maintaining high classification performance. This optimization ensures that the system can be implemented on hardware with limited resources, making it more accessible for real-world applications.

          The use of the ResNet50v2 model introduces several benefits. ResNet50v2 incorporates skip connections, which improve gradient flow during training and prevent issues like vanishing gradients, allowing the model to converge faster. This architecture also reduces the number of trainable parameters, making training more efficient while preserving accuracy. As a result, the system requires less computational power and training time, making it suitable for large-scale LULC classification tasks.