The study, conducted by Tarza Hasan Abdullah from the Department of Computer Science at Salahaddin University-Erbil in Kurdistan, Iraq, focuses on five specific deep learning models: InceptionV3, DenseNet121, NasNetLarge, Xception, and ViT-16. These architectures represent a mix of traditional convolutional neural networks (CNNs) and newer transformer-based approaches. The author trained and tested these models using a dataset of tomato leaf images that were converted to grayscale. The choice of grayscale images suggests an attempt to reduce computational complexity and focus on structural features rather than color variations, which can sometimes be inconsistent in field conditions.

To assess the performance of each model, the study utilized four standard metrics: accuracy, precision, recall, and F1-score. These indicators provide a comprehensive view of how well each model can correctly identify disease classes without misclassifying healthy leaves or missing actual cases of disease. The experimental results revealed distinct differences in performance among the five architectures. The ViT-16 model, a vision transformer, demonstrated superior performance compared to the other four models. It achieved an accuracy of 95.34% and an F1-score of 94.33%, indicating a strong balance between precision and recall in its classification tasks.

Following the ViT-16, the DenseNet121 and Xception models also performed well, with both achieving accuracy rates exceeding 94%. These results suggest that certain convolutional architectures remain highly effective for image classification tasks in agriculture. However, the study noted that InceptionV3 and NasNetLarge achieved relatively weaker results. Specifically, these two models showed shortcomings in capturing disease-specific features, which was reflected in their lower precision and recall scores. This disparity highlights that not all deep learning models are equally suited for detecting subtle visual patterns associated with plant pathology.

The findings indicate that transformer-based models, particularly the ViT-16, hold great promise for agricultural image classification. Unlike traditional CNNs, which rely on local feature extraction through convolutional filters, transformers use self-attention mechanisms to capture global dependencies within an image. This capability may allow them to better recognize complex disease patterns that span larger areas of the leaf. The study suggests that such models could become valuable tools for farmers and agricultural professionals, providing a non-invasive and rapid method for diagnosing crop health.

The dataset used in this research is publicly available in the Kaggle repository, specifically the plant disease dataset. This accessibility allows other researchers to reproduce the study’s findings and potentially extend the work to other crops or disease types. The author also noted that all code associated with the study can be made available upon reasonable request, promoting transparency and reproducibility in the scientific community. By using a public dataset, the study ensures that the results are comparable to other research in the field, facilitating a broader understanding of model performance in agricultural applications.

While the study demonstrates the high accuracy of the ViT-16 model, it is important to consider the context of these results. The models were tested on grayscale images, which may not fully represent the variability found in real-world field conditions, such as different lighting, backgrounds, and leaf orientations. Furthermore, the study focuses on tomato diseases, and the generalizability of these findings to other crops remains to be established. Future research could explore the application of these models in multi-crop scenarios or in real-time monitoring systems using mobile devices.

The implications of this study extend beyond academic interest. As the global population grows, the demand for food production increases, making efficient crop management essential. Deep learning models that can accurately detect diseases early can help farmers apply targeted treatments, reducing the need for broad-spectrum pesticides and minimizing environmental impact. The superior performance of the ViT-16 model suggests that investing in transformer-based architectures could yield significant benefits for precision agriculture. However, practical deployment will require further optimization to ensure that these models can run efficiently on devices with limited computational resources.

In conclusion, the comparative study highlights the potential of vision transformers in agricultural image classification. The ViT-16 model’s ability to outperform established CNN architectures in detecting tomato diseases underscores the rapid evolution of deep learning techniques. As researchers continue to refine these models and test them in diverse agricultural settings, the integration of AI into farming practices may become more widespread. This study contributes to the growing body of evidence supporting the use of advanced machine learning tools to enhance food security and sustainability in agriculture.

Subject of Research: Agricultural Science

Article Title: A comparative study of deep learning models for tomato disease detection

Article References: Abdullah, T. H. (2026). A comparative study of deep learning models for tomato disease detection. Neural Computing and Applications, 38(17), Article 724. https://doi.org/10.1007/s00521-026-12399-z

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12399-z

Keywords: Deep Learning, Tomato Disease, Computer Vision, Agriculture, Machine Learning, comparative, deep, learning, models, tomato, disease, detection

The target resource is water ice buried beneath Martian soil and rock. Orbital observations indicate that extensive deposits occur in the planet’s mid-latitudes, sometimes beneath only centimeters to meters of regolith. At some scarps, exposed ice sheets begin roughly one to two meters below the surface and extend more than 100 meters downward. Such deposits could eventually supply water for life-support systems and for producing rocket propellant. The quantities involved are substantial: previous mission studies have estimated that a Mars ascent vehicle could require about 150 metric tons of water, while refueling a Starship-class vehicle could require approximately 600 metric tons. Reaching ice is therefore not simply a matter of finding it from orbit. A practical system must penetrate uncertain mixtures of rock, sediment, dust, and ice, remove the debris continuously or in controlled pulses, and then establish a subsurface reservoir. RedWater combines a rotary-percussive drill with a later melting and pumping sequence based on Rodriguez Well, or Rodwell, technology developed for extracting water from terrestrial polar ice.

During the drilling phase, a bottom-hole assembly breaks the formation into small particles. A hollow metallic coiled tube deploys the assembly and carries electrical, pneumatic, and hydraulic lines from the surface. Gas released near the drill bit entrains the particles and pushes them upward through the annulus. RedWater uses direct circulation: gas travels down the drill string, exits at the bit, and returns through the surrounding borehole space with the cuttings. This arrangement is simpler than reverse circulation, in which debris travels up the center of the drill pipe through a more complicated flow path. The choice is important because the gas must do two jobs. It must first pick up newly created particles at the bottom, then sustain a dilute flow capable of transporting them through the full depth of the hole. If the gas speed is too low, the particles can settle, form dense regions, and recirculate rather than leave the borehole. Accumulating debris increases the energy required to drill and can ultimately cause the drill to stall.

The model treats that transition using empirical correlations for vertical pneumatic conveying. Its central quantity is the choking velocity, the approximate gas speed below which particle transport becomes inefficient. The calculation accounts for borehole geometry, gas density and viscosity, temperature, pressure, gravity, particle density, particle diameter, particle sphericity, and the rate at which the drill advances. Because the borehole is an annulus rather than a round pipe, the researchers use a hydraulic diameter equal to the borehole diameter minus the diameter of the RedWater assembly or coiled tubing. The model estimates the mass flux of cuttings from the borehole area, penetration rate, and density of the material being drilled. It then calculates particle free-fall speed from drag relationships and corrects that speed for the irregular shape of real drill chips. Voidage, the fraction of the conveying volume occupied by gas, is coupled to the choking velocity, so the equations are solved iteratively. A further iteration estimates gas density from bottom-hole pressure using the ideal gas law, including environmental pressure, gas-solid hydrostatic pressure, and frictional losses.

To connect the predicted velocity with hardware, the study converts the result into a required gas mass flow rate. That flow can be controlled using the upstream pressure, gas type, orifice area, and discharge coefficient. The researchers also apply a safety factor of 1.5 to the estimated choking velocity when defining a design condition. The approach was tested in three different settings. In a dedicated vacuum experiment, a 2.44-meter vertical annulus used a 38.1-millimeter inner tube and a 57.2-millimeter transparent outer tube. Sieved silica sand represented drill chips, while nitrogen entered at the bottom through a long supply line. The chamber pressure was maintained at 1.3 kilopascals, and a camera recorded particle motion at 30 frames per second. Steady flow tests used rates of 0.25, 0.4, and 1.0 grams per second. At 1.0 grams per second, the sand was immediately entrained and left the observed section in less than half a second. At the two lower rates, particles recirculated and some remained after the gas was shut off.

The two other experiments used the RedWater system itself. In a freezer test, the drill operated in a 1.4-meter crystalline ice tower at approximately minus 10 degrees Celsius. The system reached an average penetration rate of 0.59 millimeters per second, with instantaneous rates between about 0.35 and 0.75 millimeters per second. Nitrogen was supplied at 550 kilopascals, corresponding to a calculated flow of 16.7 grams per second, but some chips continued to recirculate. After the supply was increased to 690 kilopascals and 18.5 grams per second, the chips were observed to blow out with minimal recirculation. In a separate thermal-vacuum test, the ice was cooled to roughly 210 kelvin and the chamber pressure was reduced to 1.0 kilopascal. Carbon dioxide flowed continuously at a directly measured 0.55 grams per second, and the system cleared the brittle ice chips without visible difficulty. That test also demonstrated end-to-end operation, with liquid water delivered to a container outside the vacuum chamber, although the experiment was designed primarily as a system demonstration rather than a dedicated threshold measurement.

When measured conditions were supplied to the model, its estimates broadly matched the observed transitions. For the sand experiment, the predicted minimum was 0.97 grams per second, within the observed range between inefficient transport at 0.4 grams per second and effective clearing at 1.0 grams per second. For the thermal-vacuum experiment, the model predicted 0.45 grams per second, slightly below the 0.55 grams per second that successfully cleared the chips. The freezer prediction was 16.4 to 16.9 grams per second, close to the 16.7 grams per second at which recirculation was still visible and just below the 18.5 grams per second that cleared the borehole. Because no intermediate rates were tested, the exact threshold remains uncertain. The comparison nevertheless spans different gases, pressures, geometries, and particle conditions. The researchers report that the more than 30-fold difference between the freezer and thermal-vacuum flow rates is driven mainly by ambient pressure and its effect on gas density. At near-Martian pressure, a given gas supply can produce high velocities near the drill bit, increasing drag and momentum transfer to the particles.

The results have direct implications for mission architecture, but they do not represent a final qualification of the drilling system. The model indicates that an optimized RedWater design could require less than one kilogram of gas to drill through a meter of rocky overburden under the projected Martian conditions. Gas could be transported from Earth or compressed from the Martian atmosphere, an approach made more credible by the demonstrated operation of the MOXIE instrument’s atmospheric gas compressor. The calculations also suggest that the minimum instantaneous flow rate is not strongly controlled by the drilling penetration rate, because the gas-solid mixture remains highly dilute even as more cuttings are generated. Faster drilling can still reduce total gas consumed per meter by shortening the time the flow must operate. A pulsed system could offer another efficiency benefit: chips might be allowed to accumulate briefly before a gas pulse carries them to the surface. The pulse would need to last long enough for particles to travel the increasing distance as the borehole deepens.

Important uncertainties remain. All three experiments were performed in Earth gravity, whereas Martian gravity is about 38 percent as strong, and the model predicts that lower particle weight should reduce the required gas flow. The tests also used prepared sand or relatively homogeneous ice, not fractured, porous, dusty, or mixed Martian formations. Gas could leak into surrounding rock or ice instead of returning through the annulus, raising the supply requirement. Particle size and sphericity were estimated, even though drill chips can be irregular and span a broad distribution; the largest particles may determine whether clearing succeeds. The model also stops at the borehole exit and does not address how discharged cuttings will be diverted from the surface opening. The researchers recommend full-scale tests at depths approaching 25 meters, experiments that resolve the transition between recirculation and clearing more finely, computational-fluid-dynamics simulations, reduced-gravity testing, and trials using realistic regolith and ice mixtures. Despite these limitations, the agreement between model and observations supports pneumatic chip clearing as a plausible component of future Martian water-mining systems.

Subject of Research: Pneumatic removal of drill cuttings during Martian water-ice extraction

Article Title: Deep drilling on Mars: pneumatic chip clearing model for the RedWater mining system

Article References: Stolov, L., Palmowski, J., Zacny, K., Yen, B., Mellerowicz, B., Mank, Z., Sanasarian, L., & Schultz, J. (2026). Deep drilling on Mars: pneumatic chip clearing model for the RedWater mining system. Space and Planetary Resources, 2(1), Article 7. https://doi.org/10.1007/s44461-026-00013-y

Image Credits: AI Generated

DOI: 10.1007/s44461-026-00013-y

Keywords: Mars, water ice mining, planetary drilling, pneumatic conveying, RedWater, in-situ resource utilization, coiled tubing, thermal-vacuum testing, Deep, drilling, pneumatic, chip