AgriNova — Smart Agricultural Assistant

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072

AgriNova — Smart Agricultural Assistant

Aakash Radhakrishnan, Krishna Yadav, Om Kamble, Sanket Bihare, Mrs. Nisha Karolia ***

Abstract- Most agricultural tech tools are either too expensive for small Indian farmers or impossible to use in a field where the cell signal keeps dropping. AgriNova is our MSBTE 6th-semester project and it does two things: identifies leaf diseases from phone photos and recommends crops from soil nutrient data. Frontend is React, backend is Flask. We fine-tuned an EfficientNetB0 on corn and potato images from PlantVillage at 128×128 resolution for disease detection. Crop recommendation uses a Scikit-Learn Random Forest on NPK and rainfall inputs. The vision model got 96.5% validation accuracy. Soil model got 98.2%. Training was the easy part, honestly. The hard part was stopping the 200MB Keras file from crashing our 4GB laptop every time two people used the site at once. We fixed it by loading the model once at startup instead of reloading it per-request, and response times dropped from seven seconds to 1.2. This paper covers how we trained the models, how we connected everything together, and all the things that went wrong.

Keywords Convolutional Networks, EfficientNetB0, Random Forest, Deployment Bottlenecks, Full Stack Engineering.

I. INTRODUCTION

A. The ground reality for small farms

Most potato farmers in rural Maharashtra cannot tell early blight apart from normal leaf aging. What they do is look at the brown spots, ask someone older in the family, then spray whatever chemical is already in the shed. Sometimesithelps.Mostofthetimeitdoesnot.Gettingan agricultural officer out to the village takes three to five days because each officer covers something like thirty villages by themselves. Maharashtra's agricultural extension system operates at roughly that ratio across most districts. By day four or five, the Alternaria solani spores have already spread two or three rows over and halfthecropiscompromised.

Honestly, crop selection might be an even worse problem. Farmers across the state plant the same crop every year without testing the soil. Nitrogen levels drop after a bad monsoon. Phosphorus shifts when they switch fertilizer brands. Nobody notices because government soil testkitscost300rupeesandtaketwoweekstocomeback. So the farmer skips it, yields fall 15-20%, and then everyonesaysitwastherain.

B. What we wanted to build

MSBTE 6th-semester micro-project needed to solve a real problem. That was the brief. We spent the first week just scrolling through GitHub and Kaggle looking at what already existed. There are probably 200 repositories out there that classify plant diseases. Kaggle has even more notebooks doing the same thing. But almost every single one stops at the notebook. You run the cell, get a number, close the browser tab. No one actually deploys it somewhereafarmercanuseit.

So that became the project. Get the model out of the notebook and onto a website, and make sure the whole thing loads in under three seconds on a phone browser. Once we picked that target everything else followed from it.Backendcodehadtobefast.TheAPIcouldnotbeslow. Did not matter how good the model was if the user was going to sit there staring at a loading spinner for ten seconds.

II. PRIOR WORK AND WHAT WENT WRONG

A. Why drones don't help here

Most published crop diagnostic research assumes you havea multispectral droneflyingovera 500-acresoybean field somewhere in Iowa. If you can drop money on a DJI Matrice,sure,thatworks.Butitmeansnothingforafarmer working two acres of potatoes near Pune. We were only interested in proximal sensing, which really just means holdingaregularphonecameraabout10centimetersfrom theleaf.Nodrone,noextrahardware.

Before CNNs took over this space, people were doing manual feature extraction. Around 2012 to 2014 the approach was to write code that isolated leaf edges using Cannyfilters,ortomeasuregreenchannelpixelintensities, andthenpassthosehand-builtvectorsintoanSVM.Itsort of worked under lab lighting. Take the same setup outdoors though and it falls apart. One shadow, one overexposedphotofromafternoonsunlight,andtheSVMstarts flagginghealthyleavesasinfected.

B. What changed with deep learning

PlantVillage changed everything. That dataset hit Kaggle and suddenly everyone was fine-tuning AlexNet or VGG16 and publishing papers showing 99% accuracy. The catch: PlantVillage images are shot against clean backgrounds with consistent lighting. Actual farm photos

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072

havedirt,fingers,overlappingleaves,terriblecompression. Try running those 99% models on real field images and accuracydropstomaybe80%,sometimeslower.

We started with ResNet50. Took 4.8 seconds per image on our laptop. Completely unusable for anything real-time. Then we tried MobileNet which was fast but keptconfusingearlyblightandlateblightonpotatoleaves. Those two need different treatments so getting them mixed up is genuinely dangerous. Ended up going with EfficientNet because it was fast enough and accurate enough. Not the best at either, but the best balance we couldfindwithoutbuyingabetterGPU.

For the soil recommendation side of things, our first try was actually a small feedforward neural network. It scored worse than a basic Random Forest on the exact same NPK dataset. Tree-based models just handle messy tabular inputs better. pH at 6.5 sitting next to rainfall at 1200mm, the forest splits wherever it needs to split. No normalizationstepneeded.

C. Why existing solutions don't translate to India

We read a lot of papers before starting. Almost every one optimizes for Western farms. Big acreage, stable broadband,GPUserversinadatacentersomewhere.None of those assumptions hold for a 2-acre potato farm in MaharashtrawheretheonlyinternetisaJio4Gconnection that drops every few minutes. Any system that needs constant server contact is broken in that context. Just broken.

Offline capability is treated as a feature request in most of these papers. For us it is the end goal. We kept inference on the server for now because converting to TFLite would have taken time we did not have, but we knowthatiswherethishastogoeventually.

And then there is the language problem that basically noonetalksabout.Everyplantdiseaseinterfacewelooked at shows the raw disease name right on the screen. "Cercospora zeae-maydis" on a results card. That is meaningless to a farmer who never studied plant pathology.Wewroteplain-languagedescriptionsforevery disease class that our model can detect. What to look for, what to spray, in normal words. Small thing, but small things decide whether someone actually opens the app a secondtime.

Barbedowroteaboutdatasetbiasbackin2018[3]and people cite that paper constantly. But the recommended fixes are always about collecting more controlled images. Real photos from cheap Android phones have JPEG compression artifacts and inconsistent white balance and halfthetimeanotherleafisblockingtheoneyouaretrying to photograph. Our augmentation pipeline was designed

around exactly this kind of noise, not around making accuracynumberslookbetteronpaper.

III. HOW WE BUILT IT

Fig. 1 shows the overall layout. Two prediction pipelinesrunninginsideoneFlaskprocess.Theimageside works like this: user uploads a JPEG, React reads it as a binary blob, packs it into FormData, and Axios fires a multipartPOST.Ontheserver,PILopensthefile,resizesto 128×128, converts to a Float32 array, and hands it off to EfficientNetB0. The response is a JSON object with probabilities for all six disease classes. The soil pipeline is morestraightforward.UsertypesinNPK,pH,temperature, humidity, and rainfall. Flask passes those through a StandardScaler, the Random Forest does its thing, and a crop name comes back as JSON. No state is held between requests.EveryPOSTcontainseverythingthemodelneeds.

MobileNetV2 sits right at the entrance of the image pipeline. If it recognizes a car or a face or a building at above 70% confidence, the request gets killed before the disease model ever loads the pixels. Both models live in memory from the moment Flask starts up, which eats about800MBof RAM,but it meansno requesthastowait for a model to load. On one laptop running everything locally,thatseemedlikeafairtrade.

Web App User

HTTPRequests | FormDataPayloads

React.js Client

TabularSoilData ←→ PixelArrayBuffer Flask WSGI Server

Scikit Random Forest

EfficientNetB0 Subsystem

StandardScaler | Weights KerasModelHierarchy

Fig.1. End-To-EndSystemArchitectureofMobile AgriNovaPlatform.

A. The frontend

React 19 with Vite for bundling. We started with Create-React-App but hot reload was taking 8 to 10

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072

seconds every time we saved a file. When you are sitting there fixing CSS bugs for three hours straight, that wait getspainful.SwitchingtoVitecutthereloadtimetounder a second. Seems like a minor thing but over a long work sessionitmadearealdifference.

Image uploads were surprisingly painful. The flow sounds simple when you describe it: user picks photo, React reads the blob, renders a thumbnail, wraps it in FormData,sendsitvia Axios.Shouldhavetakenmaybe an afternoon. Instead it took us two full days. Turns out certain Samsung phones embed EXIF orientation data in their JPEGs, so the image would arrive on the server rotated 90 degrees. We had no idea this was even a thing until we started Googling why the uploaded image was arrivingsideways.

Validation on forms was another thing we did not think about until it bit us. Users would hit submit with blankinputfieldsandFlaskwouldcrashonthenullvalues. We added checks in the React components so the submit buttonstays greyuntil everyrequired fieldhassomething in it. Server error rate went from roughly 40% of all requests down to basically zero after that. Felt disproportionatelysatisfyingforwhatwasreallyjustafew if-statements.

B. The backend and the crashes

WewentwithFlask,notDjango.DjangohasafullORM and an admin panel built in and we needed none of that. WeneededsomethingthatcouldacceptaJPEGandreturn JSON.Flaskdoesthatin30lines.

First real problem was CORS. React's dev server runs on port 5173, Flask listens on 5000, and browsers block cross-origin requests by default. Every API call just died silently. We spent about four hours going through Stack Overflow threads on Flask-CORS before getting the right headerconfiguration.Lookingbackitseemsobviousbutat thetimewehadnoideawhatwashappening.

The biggerdisasterwasmodel loading.Akashhadput the load_model('model.keras') call inside the /predict route itself. Seemed reasonable. But that Keras file is 200MB and Flask was reloading it from disk on every single request, which meant a seven-second wait just to processoneimage.Wedidnotrealizehowbaditwasuntil AkashandOmbothhadtheappopeninseparatebrowser tabsatthesametime.Thelaptophas4GBofGPUmemory. Two loads running in parallel exceeded that and Python justcrashed.Memoryallocationerror,nogracefulfallback, theprocesswasgone.

Oncewefiguredoutwhatwasactuallyhappening,the fix was not complicated. Move load_model() to the top of thescript,outsideanyroute.Flaskloadsitonceonstartup

andthenmodel.predict()runsagainsttheweightsthatare alreadyinmemory.Sevensecondswentdownto1.2.

IV. ENGINEERING THE DEEP LEARNING MODEL

A. The training images

We pulled about 15,000 corn and potato leaf images from PlantVillage on Kaggle. CNNs need all inputs at the same resolution, so Keras ImageDataGenerator resized everyimageto128by128pixelsbeforefeedingitin.

Farm photos look nothing like dataset photos. Leaves are off-center, lighting varies from one image to the next, and sometimes there is literally a thumb covering part of the frame. We used data augmentation to make the model robust to this. Random rotation up to twenty degrees, horizontal flips, slight shearing at the edges. We deliberately left out vertical flips because leaves do not grow upside down in practice and flipping the lighting

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072

gradient just confuses the convolutional filters. Whole pointofaugmentationherewastoteachthemodelwhata rust spot actually looks like, not just that a rust spot happened to always appear in the same position in the trainingset.

B. The CNN setup

Pre-trained EfficientNetB0 as the backbone. EfficientNetusescompoundscaling,meaningitadjuststhe network's depth and width and input resolution all at the same time rather than just adding more layers on top. Keeps inference speed manageable on consumer-grade hardware.

We chopped off the original classification head and replaced it witha Denselayer at256 neurons followed by Dropout at 0.5. Dropout randomly zeroes out half the connections on each forward pass during training. Accuracy takes a hit in the short term but the model can not just memorize images anymore, which was the problem we kept running into. Whole thing ran for 10 epochsusingCategoricalCrossentropyasthelossfunction. Tookabout38minutesonourRTX2050.Weexpecteditto takelongerhonestly.

C. The soil model

Scikit-Learn Random Forest for the crop recommendations. Not a neural network. Random Forests build hundreds of individual decision trees and then averagetheirvotestogether.Onetreemightsaynotorice because rainfall is below 100mm. A different tree says no to wheat because nitrogen is too low. Averaging across all of them handles edge cases that any single tree would get wrong.

First version overfit badly. Training accuracy was 100%, test set was at 89%. Classic overfitting. We ran GridSearchCVoverabout300parametercombinationsand found that setting max_depth to 15 fixed it. Test accuracy wentfrom89%to98.2%withthatonechange.Pickledthe final model alongwith itsStandardScaler weightssoFlask couldloadbothfilesonstartup.

V. RESULTS

A. What the numbers say

EfficientNet landed at 96.5% validation accuracy. Higher than we expected going in. The confusion matrix showedoneproblemthatkeptcomingup:earlyblightand normalage-relatedbrowningonpotatoleaveslookalmost identical. Brown patches, concentric ring patterns, same generalshape.Anagronomistcantellthemapartbyfeeling the leaf texture but the model only sees pixels, so it gets them mixed up. The soil model came in at 98.2% after we ran grid search. Tabular classification is just a simpler

problemthanimagerecognitionsothatnumberwasnota surprise.

TableIIhastheper-classbreakdown.Weakestclasses are Corn Northern Leaf Blight at 94.1% and Potato Late Blightat93.5%.Makessensewhenyoulookattheimages. Both produce these elongated brown lesion patterns that look confusingly similar to each other, especially in early stages. The confusion matrix confirmed it, most misclassifications are between those two classes. Healthy leafdetectionisclosetoperfectacrossbothcropsbecause a clean green leaf looks nothing like a diseased one. That partwaseasyforthemodel.

TABLE II. PER-CLASS VALIDATION ACCURACY

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072

Table III has the latency comparison. MobileNetV2 is fastest at 8.7ms per inference but we did not reject it because of speed. The problem was that it kept swapping earlyblightandlateblightonpotatoes.Thosetwodiseases require completely different fungicide treatments, so a wrong prediction there is worse than being slow. EfficientNetB0at14.2mswasthemiddleoption.Notasfast as MobileNet, nowhere near as slow as ResNet50, and it doesnotconfusethetwoblighttypesnearlyasoften.

TABLE III. MODEL INFERENCE LATENCY COMPARISON

The right panel has a "Top Predictions" card that shows every class and its probability, not just the winner. Cercospora leaf spot 93.6%, Northern Leaf Blight 5.8%, Potato Early Blight 0.6%, and so on. We did this intentionally.Ifthemodelgives52%tooneclassand47% to another, the farmer can see that and know to get a second opinion rather than just trusting whichever class happenedtoedgeahead.

Bottom right has two info cards next to each other. One for symptoms, one for treatment. Both are written in simpleEnglish.Wewrote"Smallrectangularbrowntogray lesionsrunningparalleltoveins"insteadofsomethinglike "Cercospora zeae-maydis infection presenting as necrotic striations." Nobody reading this screen in a field has a pathologybackground.Theywanttoknowwhattolookfor on the leaf and what chemical to buy. So that is what we tellthem.

VII. CONCLUSION

B. The gatekeeper problem

We found a real problem during testing. Upload a photo of a car dashboard and EfficientNet still classifies it as some crop disease. The softmax layer just picks whicheverclasshasthehighestscore,evenifthatscoreis 12%.Todeal withthisweputMobileNetV2infrontofthe disease model as a gatekeeper. MobileNetV2 was already trainedonImageNetsoitcantellthedifferencebetweena plant and a car. If it tags a non-plant object at over 70% confidence, Flask rejects the image and tells the user to upload an actual leaf photo. Does not catch everything, some weird photos still get through, but it blocks the obviouscases.

VI. SYSTEM INTERFACE

Detection pagehastwopanels.Ontheleftthereisthe imageuploadareawithagreen"AnalyzeCrop"buttonand a "Clear" button belowit.Afterthemodel returnsa result, a card pops up showing the confidence as a big number, the disease name, crop type, and a badge that says HIGH, MEDIUM, or LOW confidence. Green progress bar fills proportionally. So for a Gray Leaf Spot detection at 94%, the card shows "Corn (Maize)" with "Diseased Leaf" and a HIGHCONFIDENCEbadgenexttoit.

WholethingrunsononelaptopwithabudgetGPU.No cloud services, no paid APIs. Getting from a Jupyter notebook toa real ReactandFlaskapplicationwasharder than any of us expected. CORS debugging took a full day. The memory crash took two days to figure out. The EXIF rotation bug was another two days on top of that. If someone had told us at the start that connecting the frontendtothebackendwouldbeharderthantrainingthe neural network, we probably would not have believed them.Butitwas.

It works now though. Open the site on a phone browser, upload a leaf photo, and a diagnosis comes back in 1.2 seconds. Soil recommendations are faster than that. For a semester project built by four students, we are all prettysatisfiedwithwhereitendedup.

VIII. WHAT WE WOULD DO NEXT

If we had more time, first thing would be Docker. SettingupthePython environment fromscratchona new machine takes about 45 minutes because TensorFlow versions keep conflicting with NumPy versions and CUDA versions. Docker would make the whole deployment a single command. We would also add Marathi and Hindi translations for the interface because the farmers who needthismostarenotreadingEnglishconfidentlyenough toactondiseasenameslike"Septorialeafspot."

Real end goal is TFLite. The Keras model is 200MB. TensorFlow Lite quantization would compress that to maybe15or20MB,smallenoughtoshipinsideanAndroid APK. At that point the phone handles inference locally. No Flask server, no internet connection required. That is wherethisprojectneedstoendupifitisgoingtoactually helpanyoneinavillagewherecellcoverageisunreliable.

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072

We also want to build a feedback loop. Right now the modelistrainedonceandneverupdated.Ifafarmercould tap a button that says "this diagnosis was wrong" and submitacorrection,thosecorrectedphotoscouldgointoa retrainingqueue.TheMobileNetV2gatekeeperwouldkeep junk out of that queue. We did not have time to build this beforesubmissionanditwouldneedadatabaselayerthat we currently do not have, but the concept is straightforwardenoughthatafutureteamcouldaddit.

Cropcoverageisthelastthing.PlantVillagehaslabeled images for 14 crops. We trained on two, corn and potato, because of the semesterdeadline. Wheat, rice, andtomato wouldcovermostofwhatgetsgrowninMaharashtra.The codechangeswouldbeminimal,justaddtheimagefolders and bump the output layer from 6 neurons to however many classes the new dataset has. Not a hard engineering problem,justatimeproblem.

REFERENCES

[1] A. Kamilaris and F. X. Prenafeta-Boldú, "Deep learning inagriculture:Asurvey,"ComputersandElectronicsin Agriculture,vol.147,pp.70-90,2018.

[2] D. J. Mulla, "Twenty five years of remote sensing in precision agriculture: Key advances and remaining knowledge gaps," Biosystems Engineering, vol. 114, no.4,pp.358-371,2013.

[3] J. G. A. Barbedo, "A review on the main challenges in plant disease data sets for machine learning," FrontiersinPlantScience,vol.9,p.1531,2018

[4]S.P.Mohanty,D.P.Hughes,andM.Salathé,"Usingdeep learning for image-based plant disease detection," FrontiersinPlantScience,vol.7,p.1419,2016.

[5] E. You, "Vite: Next Generation Frontend Tooling," vitejs.dev.[Online].Available:https://vitejs.dev.

[6] R. Behmann, A. K. Mahlein, T. Rumpf, C. Römer, L. Plümer, "A review of advanced machine learning methods for the detection of biotic stress in precision crop protection," Precision Agriculture, vol. 16, pp. 239-260,2015.

[7] A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto, and H. Adam, "MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications," arXiv preprint arXiv:1704.04861,2017.

[8] M. Tan and Q. V. Le, "EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks," InternationalConferenceonMachineLearning(ICML), 2019,pp.6105-6114.

[9] L. Breiman, "Random Forests," Machine Learning, vol. 45,no.1,pp.5-32,2001.