An artificial-intelligence model identified Bangladeshi mango varieties with 97.36% accuracy in a reported test of 303 held-out images, misclassifying eight. The model, EfficientNetB0, was also placed in a public web application that accepts image uploads and returns a predicted variety.
That result is best read as a benchmark within the study's own image collection. The reported test set was held back from training and hyperparameter tuning, but no uncertainty interval was reported, and the study did not report whether farmers, traders or extension workers made better decisions with the app.
Photographs from orchards and markets
The project set out to build a nine-class image dataset, compare three modern convolutional neural networks and put the strongest model into a lightweight web tool. The models were software systems designed to identify visual patterns in photographs.
Researchers collected 2,013 smartphone photographs at 3024 by 4032 pixels in orchards in Rajshahi and Chapai Nawabganj and in wholesale markets in Faridpur and Dhaka. Sessions ran from May to July 2024. The images were taken in natural daylight against varied backgrounds without artificial staging, and the initial data were organized into 10 labelled folders.
The final setup had nine classes because BARI-4 and BARI-7 were combined into one Bari class. The study says the two have highly similar visual traits and are marketed collectively. Its reported class counts were Amrapali 252, Bari 411, Fazlee 156, Harivanga 202, Kanchon Langra 210, Katimon 163, Langra 202, Mollika 221 and Nilambori 195. The nine listed counts do not quite reconcile with the reported overall total of 2,013, an inconsistency that makes the dataset accounting worth checking.
A controlled model comparison
For evaluation, the images were divided by stratified random sampling into 1,409 training images, 301 for validation and 303 for the final test, a 70-15-15 split. The researchers fixed the split after one draw and kept the test images unseen during training and hyperparameter tuning.
Training augmentation used a 0.5 probability for horizontal flips, rotations of up to 15 degrees, brightness, contrast and saturation settings of 0.2, hue of 0.1, and random crops covering 0.8 to 1.0 of the image before resizing to 224 by 224 pixels. Validation and test images were only resized and normalized.
ResNet18, ResNet50 and EfficientNetB0 all began with ImageNet-pretrained weights and were fine-tuned for nine output classes. Every layer was unfrozen, and the dropout probability was 0.3. All three models followed the same training recipe: 15 epochs, Adam optimization, an initial learning rate of 0.001, batches of 32, cross-entropy loss, learning-rate reduction when progress stalled and early stopping.
EfficientNetB0 led the benchmark
EfficientNetB0 had the highest reported validation accuracy, at 98.01%, compared with 86.47% for ResNet18 and 78.55% for ResNet50. These were descriptive results from the reported split. The paper did not report confidence intervals or significance tests.
On the held-out test set, EfficientNetB0 scored 97.36% accuracy. The paper reports this as a 2.64% misclassification rate, with eight of the 303 images assigned to the wrong variety. No uncertainty interval was reported for this estimate.
Accuracy was not the only measure. F1-score, a single measure that balances precision and recall, was 0.98 both as a macro average across classes and as a weighted average that reflects class size. The report listed class-wise F1 scores of 0.93 for Harivanga, 0.99 for Amrapali and Kanchon Langra, and 0.97 for Bari. Harivanga's reported precision was 0.88 and recall was 0.95.
The detailed class-wise results carry a caveat: the supplied analysis notes inconsistent precision and recall values between the paper's prose and its table. Those detailed figures should therefore be treated cautiously, even though the reported macro and weighted F1 scores were both 0.98.
A public app, with a narrow test
Most mistakes clustered around Bari and Nilambori, and around Bari and Harivanga. The paper notes that similar appearance, viewpoint and lighting can mask differences between varieties.
EfficientNetB0 was described as having approximately 4 million parameters, and a single prediction was reported to complete in under 1.2 seconds on a standard cloud instance. In the Streamlit app, users can upload JPEG or PNG images; the system resizes and normalizes them to 224 by 224 pixels, then returns a variety label, probabilities for all nine classes and a top-1 confidence score. The app was publicly deployed at mangoclassifier.streamlit.app.
Those details show that the system was made accessible through a web interface, but they do not establish field reliability. The study reports no external, prospective or cross-region validation, no user study with farmers, traders or extension workers, and no agronomic or economic outcome. Its collection was limited to specified Bangladeshi locations and sessions from May to July 2024, so performance on additional regions, seasons, devices or post-harvest stages remains an open question.
The web app requires an active internet connection, and the model does not distinguish BARI-4 from BARI-7 because those classes were merged. The report also does not demonstrate offline or on-device mobile performance, ripeness prediction or disease detection; those are proposed future extensions rather than tested results.
The front matter identifies the document as arXiv version 1 dated 28 Aug 2026. Funding information was not reported. At this stage, the work is an initial benchmark and proof of concept for mango-variety image classification and public deployment. The next test is whether it generalizes to independent images and provides practical benefit in the field.
Paper data and sources
Original title: Empowering Local Agriculture: A Deep Learning-Powered Web System for Identifying Bangladeshi Mango Varieties
Authors: Monowar Islam, Safaruzzaman Shovo
Journal/Repository: Journal of Bangladesh Academy of Sciences, vol. 50, Supplement 1, p. 114, 2026
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text