TerraSense AI: Soil Quality Classification Using Image Processing, Deep Learning and an AI-Based Soil Advisory System

Knowing the soil type of a field is one of the first steps in agricultural planning, yet judging it by eye is subjective and laboratory analysis is slow. This paper presents TerraSense AI, a web-based prototype that sorts soil photographs into seven classes (Alluvial, Arid, Black, Laterite, Mountain, Red and Yellow) using a convolutional neural network (CNN) built with TensorFlow and Keras and served through Flask. The CyAUG-Dataset held 5,097 images before we audited it for exact duplicates, which turned up 54 duplicate groups and 55 extra files. We split the data only after handling the duplicates and resized every image to 224 × 224 RGB. On the held-out test split the model reached 93.42% accuracy, 92.89% weighted F1 and 83.07% macro F1, with Arid and Mountain soils noticeably weaker than the other classes. Around the classifier, the prototype adds a provisional, dataset-relative NPK rating drawn from a 257-row measurement table, a SQLite prediction history with analytics, and an AI Soil Advisor that is still under development. The CNN predicts soil type from an image; it does not measure fertility, nutrient concentration, crop suitability or overall soil health, and the NPK ratings are not validated agronomic thresholds. All results come from one dataset, so performance on other data remains untested. The contribution of this work is an integrated, transparent prototype together with an honest account of what its evaluation does and does not show