ModularSqueezeNet: A Modified Lightweight Deep Learning Model for Plant Disease Detection

Plant diseases make a significant contribution to production losses, which can be mitigated through constant monitoring. Manual crop disease monitoring is commonly time-consuming, costly, and erroneous. Since the performance of convolutional neural network in obtaining accurate classification results has been phenomenal, state-of-the-art network resulted in many parameters and a high computational cost. We propose a Modular SqueezeNet model which reduces the number of parameters, lowering the model's computational burden. The model has been trained on a small open dataset containing images of 14 plants with 38 disease classes including healthy and unhealthy plants to solve a multiclass classification problem. In the proposed model we have modified the state of art SqueezeNet model in terms of fire modules, filter size, and convolution filters used in squeeze and expand phases. Accuracy of the proposed Modular SqueezeNet model is good as on training accuracy as 94.3%, validation accuracy as 89.3%, and test accuracy as 88.5%. We also developed the ensemble models with ML baselines to evaluate the performance of our model in ensemble modeling. ModularSqueezeNet with support vector machine exhibit the best performance with training accuracy of 96.5%, validation accuracy of 93.2
^%
, and test accuracy of 91.2%.