<![CDATA[Computer vision offers a promising method for soil moisture assessment especially for real-time field monitoring where sensor-based measurements are limited. This study evaluates the performance of a traditional Convolutional Neural Network (CNN) and ResNet-50 architecture for classifying soil moisture levels directly from in-situ surface images. The research involved 200 field-captured images and corresponding moisture data from a rainfed agricultural area. The models were trained with datasets grouped into two, three, and four moisture categories to test performance under varying complexity. The results showed poor model performance, characterized by high instability and severe overfitting across all experiments. Model accuracy for the traditional CNN significantly decreased from 0.513 to 0.256 as the number of classification categories increased and from 0.487 to 0.205 for ResNet-50. High RMSE values from 0.433 to 0.507 further confirmed substantial prediction errors. This finding highlights the limitation of RGB-based in-situ imagery for soil moisture classification, where environmental variability dominates the visual signal. It also suggests that soil moisture-related features are not sufficiently distinguishable under uncontrolled field conditions. The study concludes that the high variability of direct field images due to factors like inconsistent lighting, illumination, and the presence of non-soil objects is a primary obstacle to accurate classification. Future studies should implement advanced pre-processing techniques such as segmentation to reduce illumination noise.]]>
Copyrights © 2026
