Some stored-grain beetles are distinctive at a glance. Others form pairs that are separated on fine characters. Tribolium castaneum and T. confusum differ in the form of the antennal club and in the distance between the eyes on the underside of the head. Sitophilus oryzae and S. zeamais overlap in external appearance and are reliably separated only by the genitalia, and the same is true of several Cryptolestes species (Rees, 2004). The khapra beetle, Trogoderma granarium, a quarantine pest in many countries, belongs to a genus whose species are difficult even for specialists. A classifier fed dorsal photographs cannot be expected to resolve what a taxonomist resolves by dissection, and the class list has to respect that limit.

The networks that transformed image recognition were trained on ImageNet, a collection of more than a million labelled photographs (Deng et al., 2009). No grain laboratory will assemble a set of that size. It does not need to, because the early layers of a network trained on general images encode edges, textures and shapes that transfer to other tasks (Yosinski et al., 2014). Fine-tuning a pretrained network on a few hundred images per class is now the standard route for species identification (Wäldchen and Mäder, 2018). Residual networks (He et al., 2016) are the usual baseline. Compact architectures such as MobileNetV2 (Sandler et al., 2018) and the EfficientNet family (Tan and Le, 2019) are attractive where the model must run on a phone or an embedded board.

The first decision is what the classifier is for. Screening intake samples for any live insect, naming the species in a trap catch, and flagging a possible quarantine pest are different tasks with different costs of error. The class list follows from the task and from what images can show. Species that cannot be separated externally are merged into a single class named for the pair or the genus. A class for other insects and debris is included from the outset, since a deployed system will meet mites, psocids, moth larvae, fragments and weed seeds.

Labels are only as good as the identifications behind them. Each specimen should be identified by a competent person using standard keys, with dissection where the species requires it, and should carry an identifier that links it to its images. Specimens should come from several laboratory cultures and field collections for each class. Beetles from one culture share age, diet and condition, and a network can learn the culture in place of the species.

Images should be taken with the kind of device, magnification, background and lighting that will be used in service, and those conditions should be varied deliberately across all classes. If every Rhyzopertha is photographed on white card and every Oryzaephilus on grey, the background becomes the easiest feature to learn. Several views of each specimen, dorsal, lateral and ventral, add information, and the ventral view matters for the flour beetles. Dead, damaged and dusty individuals should be included because they are what real samples contain.

Dividing images at random between training and test sets is the commonest cause of inflated accuracy. Several images of one beetle are nearly duplicates, and if they fall on both sides of the split the test measures memory. All images of a specimen must go to one partition. Better still, whole sources, a culture, a collection site or an imaging session, are held out, following the argument for blocked validation of structured ecological data made by Roberts et al. (2017). The test set is then set aside and consulted once.

Two or three pretrained architectures of different size are fine-tuned on the training partition, with a validation partition used to choose the learning rate and the stopping point. Augmentation by rotation, flipping, cropping and changes of brightness and contrast enlarges the effective sample and suits specimens that have no natural orientation (Shorten and Khoshgoftaar, 2019). Colour should be perturbed only mildly, because the reddish brown of Tribolium against the darker Sitophilus is itself diagnostic. Classes with few images are over-sampled or weighted in the loss.

Overall accuracy conceals what matters when classes are unequal in size and in importance. The report should give the full confusion matrix, recall and precision for each class, and an average that weights classes equally (Sokolova and Lapalme, 2009). For a regulated species the figure of interest is the proportion of true specimens that the classifier misses. Saliency maps such as Grad-CAM (Selvaraju et al., 2017) should be inspected for a sample of images to confirm that the network attends to the insect and, ideally, to the parts a taxonomist would examine.

Modern networks are often over-confident, and their output scores should not be read as probabilities without adjustment. Temperature scaling on the validation set is a simple correction (Guo et al., 2017). With calibrated scores a threshold is set below which the system returns no identification and refers the specimen to a person. The same mechanism gives a first defence against species outside the training set, which tend to receive lower maximum scores (Hendrycks and Gimpel, 2017), although it is far from complete. The threshold is chosen from the cost of a wrong answer against the cost of a referral, and the share of images declined is reported with the accuracy on the remainder.

A classifier in service meets new cameras, new grain types and new populations. A fixed fraction of its identifications, and all referrals, should be checked by a trained person, and the agreement tracked over time. Images on which it failed are added to the training pool with verified labels, and the model is retrained at stated intervals against the same held-out test set, so that successive versions can be compared.