Smartphone light scattering at near contact supports nondestructive assessment of internal peach damage under standardized conditions

Internal peach damage is difficult to detect early, while many nondestructive systems remain too costly or cumbersome for practical deployment. This study evaluated whether near-contact light scattering acquired with a native smartphone camera and flash could support destructively verifiable assessment under standardized indoor conditions. A measurement-grounded inverse framework combined three deterministic proxy targets, compact latent feature regularization, direct regression of firmness, color difference, browning area ratio, and damage depth, and empirical-condition assignment. The proxies were operational supervision constructs rather than biologically independent tissue states. Under controlled upstream inputs, the back-end achieved 93.4% ± 0.3% classification accuracy, with R² values of 0.921, 0.931, 0.964, and 0.947 for the four continuous indicators. From standardized near-contact images, the full analytical pipeline achieved 88.7% ± 0.6% on the regular Devices A–C test set and 89.2% ± 0.6% on the held-out Device D cohort. Capacity-controlled comparisons showed incremental effects within the tested implementations: proxy supervision increased accuracy by 1.4 and 1.3% points on the regular and Device D sets, respectively, while diffusion exceeded an equal-capacity DAE by 1.4 and 2.4 points. The KA-inspired model provided an additional capacity-matched numerical improvement. The results demonstrate controlled-condition feasibility and preliminary transfer to one held-out smartphone, but not robustness across diverse devices, cultivars, maturity states, lighting conditions, operators, storage histories, or naturally occurring bruising.