Crop Disease Detection and Severity Estimation Using Machine Learning

Plant disease outbreaks destroy crops around the world, cutting into food supplies and hurting farm incomes.Although state-of-the-art deep learning models perform well on clean lab images, they often fail on real farms for three main reasons: they only provide a single disease label without assessing the severity of infection, outdoor lighting and soil backgrounds confuse them, and they misclassify leaves among different plant species.To overcome this, we suggested a practical multi-task model with an EfficientNet-B0 backbone.The network learns shared visual features and feeds them to two parallel heads: one predicting 38 disease classes, and the other estimating severity levels (Early (≤ 15%), Moderate (15-50%), Severe (> 50%)).We proposed a Species-Constrained Logit Masking step, which constrained the predictions to the selected crop in testing, decreasing cross-species false alarms from 18.4% to 0.0%.We also added Grad-CAM heatmaps for the final conv layer (features.8)as a visual verification.We observed the network detect initial lesion edges and color shifts in the first convolution (112x112) and early down sampling layers (56x56) prior to abstract feature grouping by looking at feature maps.For the performance test, we constructed a dataset consisting of 90,446 photos, mixing the laboratory images (87,848 from PlantVillage) and real field images (2,598 from PlantDoc).The model scored 99.4% accuracy on lab images and 96.2% on field photos, with a 20.8 MB file size and a 28.4 MS processing speed.