Soil Image Classification Using Transfer Learning Approach: MobileNetV2 with CNN

Abstract

This paper presents a novel study on soil image classification, leveraging the synergistic potential of transfer learning and convolutional neural networks (CNNs). The proposed approach combines the strengths of the MobileNetV2 architecture with a customized CNN model for accurate and efficient soil type recognition. The pre-trained MobileNetV2 is used to capture generic features before fine-tuning it with a dedicated soil image dataset comprising four distinct classes: red, clay, black, and yellow soils. To enhance the model’s capacity for discerning intricate soil textures, a specially designed CNN architecture is incorporated. The model’s performance is rigorously evaluated on a dataset of 108 images, each sized at 256 × 256 pixels, achieving an exceptional accuracy rate of 100% on the test dataset. The promising results demonstrate the efficacy of the proposed methodology in soil image classification tasks, offering potential applications in precision agriculture, environmental monitoring, and land management. While these findings showcase remarkable accuracy, further investigations are recommended to assess the model’s generalization across diverse environmental conditions and an expanded range of soil image datasets.