Discriminating Visually Similar Cotton Leaf Diseases: A Comparative Study of Handcrafted Colour–Texture Features and Deep Learning

Abstract

Automated diagnosis of cotton foliar disease is complicated less by detecting that a leaf is diseased than by deciding which disease it carries. The principal cotton pathogens express a small shared vocabulary of symptoms — chlorosis, marginal necrosis and interveinal green retention — and differ chiefly in the proportion and arrangement of those primitives rather than in their presence. Class boundaries are therefore intrinsically weak, and a classifier can post a respectable aggregate accuracy while systematically confusing the pairs that matter to a treatment decision.

This paper reports a controlled comparison of twenty configurations on a four-disease cotton leaf dataset of 570 images, evaluated on a common 114-image test partition. Three handcrafted representations — an HSV colour histogram, a uniform Local Binary Pattern (LBP) texture histogram [1], and their concatenation — are each paired with four classical classifiers [2]–[4], and six deep variants spanning custom convolutional networks and ImageNet transfer learning [5]–[7] are evaluated on the same split.

Four findings are reported. (i) Colour dominates texture: the best colour configuration (Random Forest, accuracy 0.8684, macro F1 0.7933) exceeds the best texture configuration (Random Forest, 0.6228 / 0.5395) by 24.6 accuracy points, and the ordering holds for every classifier. (ii) Concatenating texture with colour does not improve on colour alone (0.8596 vs 0.8684), contradicting the complementarity assumption common in this literature. (iii) We identify and measure a defect in the conventional uniform-LBP pipeline — histogramming a P = 8 uniform LBP into 58 bins leaves 48 bins (82.8%) identically zero for every image — which partially explains the texture arm's weakness. (iv) Three of six deep variants returned accuracy of exactly 0.2544 = 29/114, the majority-class rate: these models collapsed to constant prediction rather than underperforming gradually. We trace the collapse to BGR/RGB channel transposition and to omitted backbone-specific input normalisation, and argue that reported transfer-learning failures in agricultural imaging should be presumed preprocessing artefacts until preprocessing is shown correct.

We further show that accuracy is unsafe as a primary metric here: across all twelve handcrafted configurations accuracy exceeds macro F1 by 0.075–0.134 (mean 0.1016), the quantitative signature of errors concentrated in particular classes.