AI almond sorting: Satake vs TOMRA vs Cimbria

Choosing an almond defect sorter

Almond processors need to distinguish saleable kernels from defective product and foreign material without discarding too much good output. Satake BELTUZA SPECTRA, TOMRA 4C and Cimbria SEA.XL combine AI with different sensors and handling designs. The right shortlist begins with the defects the line actually needs to separate.

Satake BELTUZA SPECTRA

Satake BELTUZA SPECTRA makes X-ray information central to its almond inspection. X-rays reveal differences inside a kernel that an external color image may not show, and deep learning distinguishes acceptable product from categories such as insect damage, half-splits and doubles. The operator can choose which of those categories should be rejected. This is useful where the commercial problem is a hidden cavity or structural defect rather than only a discolored surface. SPECTRA also combines external inspection with RGB, shape and NIR information, so the internal view is part of a broader sorting system. Its belt-fed architecture gives it a different physical arrangement from the chute-based alternatives here. Satake's almond examples demonstrate the intended application, but incoming product condition and settings still affect results. The ability to develop other commodity models should not be read as an entitlement to every model within an almond-machine quotation.

TOMRA 4C

TOMRA 4C is a chute sorter that combines pulsed-LED inspection with LUCAi deep learning and local machine-learning controls. Its most useful distinction is that operators can use the pretrained and locally trained approaches individually or together. If a harvest introduces a defect not well represented in the existing model, the operator can identify examples in product images and adapt the local classification. This gives an almond processor a practical way to respond to a changing incoming crop without discarding the value of a broader pretrained model. The 4C also emphasizes controlled product flow, targeted air ejection and the XCD interface shared with other TOMRA equipment. Open access for cleaning is relevant to the everyday operation of the line. The product is not presented here as an X-ray substitute: its sensor and classification approach should be matched to the relevant almond defects, rather than assuming every TOMRA nut-inspection technology is included.

Cimbria SEA.XL

Cimbria SEA.XL combines full-color RGB vision with hyperspectral SWIR information and BRAIN AI. Its sensor approach considers appearance together with chemical-signature differences, helping distinguish unwanted material that may resemble a good kernel in ordinary color images. BRAIN focuses on a simpler recipe workflow: the operator defines acceptable and unacceptable elements, and the software helps set the classification parameters. This is attractive for a processor switching between product conditions or quality targets and wanting less dependence on repeated manual adjustment. SEA.XL also offers configurations from two to six feeding chutes, with layouts for multiple passes. That physical flexibility can matter as much as the model when a plant needs to separate several fractions or fit an existing processing flow. Hyperspectral infrared analysis is different from Satake's X-ray view of internal structure; it should not be assumed to reveal the same defect evidence simply because both extend beyond visible color.

Hidden structure, material signature and visible defects

BELTUZA SPECTRA is the most direct option here for defects represented by internal X-ray differences, such as the almond insect-damage examples Satake publishes. SEA.XL emphasizes RGB and hyperspectral SWIR differences between product and unwanted material. TOMRA 4C combines its optical sensors with both local and pretrained classification. These approaches may overlap on some rejected items without being equivalent across all defects. A processor concerned with internal cavities, shell fragments and surface discoloration is considering several evidence types. The value of each machine depends on which of those problems dominates the incoming almond stream.

How operators change the sorting decision

TOMRA 4C explicitly lets local machine learning work alongside LUCAi, so an operator can add defect examples while retaining the pretrained approach. Cimbria BRAIN emphasizes recipe creation and automatic parameter selection from the good-and-bad definition. Satake describes selectable defect categories within its deep-learning separation model. These are distinct forms of control. A processor experiencing new harvest-specific appearances may value TOMRA's local adaptation. A team managing repeated recipes may prefer Cimbria's parameter workflow. A buyer focused on particular X-ray defect classes may find Satake's category selection the clearest match.

Line architecture and the commercial comparison

Satake uses a belt-fed inspection arrangement, while 4C and SEA.XL are chute-based propositions. TOMRA emphasizes its feed flow, ejection precision and shared interface; Cimbria offers several chute counts and multi-pass configurations. The relevant cost comparison therefore includes material handling and the required sorting fractions, not just AI licensing. Numeric equipment prices and billing intervals were not confirmed for these offers. Model entitlements, installation, compressed air, utilities and support need to be specified with the machine. Published detection, false-rejection or yield figures remain vendor claims rather than results from one controlled three-machine comparison.

Which almond sorter fits the main defect?

Choose SPECTRA for X-ray-visible internal defects, 4C for flexible local learning alongside LUCAi, and SEA.XL for hyperspectral sorting with configurable chutes and BRAIN recipes. Let the defect evidence and processing flow determine the shortlist before comparing commercial packages.