Effect of Preprocessing on Modelling Soil Images Captured Using Smartphone

Abstract

Knowing soil characteristics is one important step in agricultural process. Soil characteristics such as NPK and pH values could differ the production quantity and quality of a farm. To know soil characteristics, various methods could be implemented including the use of tools such as Soil Test Kit (STK), and Rapid Soil Testing (RST), among others. For an extreme case, soil laboratory work is sometimes conducted. However, such a process is considered taking time and expensive to realize. Nowadays, the use of smartphones is getting common. Smartphones can capture images, in this case soil images, in no time. However, recognizing soil characteristics based on images needs more processes. Various artificial intelligence (AI) methods exist and could be used for the purpose, including artificial neural networks, convolutional neural networks, random forest, and gradient boosting, among others. This paper tries to experiment how the soil images captured using smartphone could be used to predict soil characteristics. Various image preprocessing methods are chosen to produce images which could be modelled using various AI methods. The results show that random forest performed the best compared to other methods with overall lowest mean squared error. Predicting pH values based on soil images produced better accuracies compared to NPK values. Image preprocessing does not influence largely on the prediction accuracies. For some cases, modelling of images without preprocessing even resulted in better accuracies.