Hybrid deep learning model for automated soil classification and crop recommendation to enhance agricultural productivity

Accurate soil classification is essential for crop selection and precision-agriculture decision support; however, conventional visual assessment is subjective, labor-intensive, and difficult to scale. This study aimed to develop and evaluate a dual-backbone deep-learning framework that combines complementary residual and lightweight visual features for automated soil-image classification and preliminary crop recommendation. Experiments used the publicly available Soil_Data_V3 dataset containing 5,254 RGB images across eight soil categories. The original images were divided using an image-level stratified 80:10:10 split into 4,203 training, 525 validation, and 526 independent test images. Images were resized to 224 × 224 pixels and normalized, while augmentation was restricted to the training partition to reduce data-leakage risk. ImageNet-pretrained ResNet50 and MobileNetV2 models were evaluated individually and combined in a hybrid architecture. Global-average-pooled features of 2,048 and 1,280 dimensions were projected into two 512-dimensional representations, concatenated into a 1,024-dimensional fused vector, and classified through dropout-regularized dense layers. A rule-based soil–crop suitability table subsequently mapped the predicted soil class to preliminary crop recommendations. On the independent test set, the hybrid model achieved an accuracy of 0.7365, loss of 0.7426, macro-precision of 0.7088, macro-recall of 0.7108, macro-F1 of 0.7097, and weighted-F1 of 0.7365. It substantially outperformed standalone MobileNetV2 and ResNet50, which achieved test accuracies of 0.1959 and 0.1680, respectively. These findings indicate that projection-based fusion of complementary CNN features improves soil-image classification. Nevertheless, external-dataset validation, stronger imbalance handling, and integration of measured soil chemistry, climate, irrigation, and regional information are required before field deployment or agronomic recommendation.