Spatiotemporal Patterns of Spider Diversity in Paddy Fields: A Deep Learning Approach from Kalahandi District

In rice agroecosystems, spiders serve as vital natural predators, contributing to ecological balance and sustainable pest management. Traditional methods of biodiversity assessment are time-consuming, require expert taxonomic knowledge, and are often limited in scale. This study presents an integrated machine learning (ML) and deep learning (DL) framework for automating spider species identification and analyzing ecological diversity across cropping seasons. High-resolution field images of spiders were used to train convolutional neural networks (CNNs) through transfer learning with models including ResNet50, EfficientNetB0, and MobileNetV2. EfficientNet-B0 achieved the highest classification accuracy (94.1%) and F 1 -score (0.93). To model ecological diversity, supervised ML algorithms-Random Forest and XGBoost-were applied using environmental parameters such as temperature, humidity, rainfall, and seasonality to predict Shannon and Simpson diversity indices, with XGBoost achieving an R² of 0.90. Unsupervised clustering using K-Means and DBSCAN revealed distinct spider community patterns linked to microhabitats, while PCA and t-SNE facilitated dimensionality reduction and visualization. Temporal trends in spider populations were forecast using LSTM and Prophet models, with LSTM yielding the lowest mean absolute error (MAE =0.08). This hybrid computational ecology approach demonstrates the potential of AI in enhancing biodiversity monitoring and agroecological research. The pipeline offers scalable, interpretable, and efficient tools for ecological field studies, contributing to precision agriculture and conservation planning.