SOIL TEXTURE CLASSIFICATION FROM SMARTPHONE FIELD IMAGERY UNDER UNCONTROLLED ILLUMINATION: A BENCHMARK OF EIGHT CONVOLUTIONAL ARCHITECTURES

Abstract

Soil texture governs water retention, nutrient availability and crop suitability, yet establishing it conventionally requires laboratory sedimentation and digestion analysis with a turnaround of several days and a per-sample cost that limits sampling density. This paper evaluates whether the texture class of a field can instead be recovered from a single close-range photograph taken with an ordinary smartphone under whatever light the field offers. A dataset of 2,455 field images at 1024 x 768 pixels was curated in Fort Abbas tehsil, Bahawalpur division, Punjab, Pakistan, covering four locally recognised classes — silt (556 images), sand (561), chalky (656) and loam (682) — captured across three morning sessions spanning fog, smog and direct low-angle sunlight, with no colour correction or illumination normalisation applied. A compact convolutional neural network was designed for the task and benchmarked against eight established architectures, VGG16, ResNet-50, ResNet-101, EfficientNetV2B3, MobileNetV3-Small and MobileNetV3-Large for whole-frame classification and YOLOv5 and Detectron2 for localisation, with every model trained and evaluated on one fixed 70/15/15 partition so that architectural effect is not confounded by split variance. Training accuracy exceeded 0.98 for every backbone, but generalisation separated them sharply: test accuracy ranged from 0.63 for ResNet-101 to 0.9887 for VGG16 with transfer learning, with EfficientNetV2B3 second at 0.9865. The best configuration exceeds the best verifiable figure reported by earlier hand-engineered-feature pipelines, 95.3%, while using consumer hardware and free cloud compute, and the detection models recover soil regions from unconstrained frames without manual cropping. The results indicate that deep convolutional models trained on illumination-diverse field imagery can support a phone-based screening tool for texture classification, subject to the geographic and seasonal limits of a single-locality dataset.