With the advancement of deep learning and thermal imaging technology, prediction of plant disease before the appearance of any visual symptoms gains attention. Studies showed that before the appearance of any visual symptoms, some internal changes take place in the plant that cannot be detected externally. These changes may be captured by the thermal images which will help to predict the diseases at the earlier stage. This early prediction will increase the probability and time to recovery; reduce the use of pesticide, resulting in cost effective, quantitative and qualitative production with less environmental pollution.

In this study a plant disease prediction system based on thermal images has been developed by exploring the dynamic feature extraction capability of deep learning technology. The proposed system consists of three convolutional layers to overcome the computational overhead and the over fitting problem for small dataset. The system has been tested with a very common disease Bacterial Leaf Blight, of rice plants.

The proposed model has been evaluated using several metrics like accuracy, precision, type-I error, type-II error. This novel model predicts the disease at the earliest stage (within 48 h of the inoculation) with 95% accuracy and high precision 97.5% (2.3% Type-I error and 7.7% Type-II error). A comparative study has been done with four standard deep learning models -VGG-16, VGG-19, Resnet50 and Resnet101 and also with machine learning algorithms -Linear Regression and Support Vector Machine, to establish the superiority of the proposed model.

随着深度学习和热成像技术的进步，在出现任何视觉症状之前预测植物病害受到关注。研究表明，在出现任何视觉症状之前，植物中会发生一些无法从外部检测到的内部变化。这些变化可能会被热图像捕捉到，这将有助于早期预测疾病。这种早期预测将增加恢复的可能性和时间；减少农药的使用，从而实现具有成本效益、定量和定性的生产，同时减少环境污染。

本研究通过探索深度学习技术的动态特征提取能力，开发了一种基于热图像的植物病害预测系统。所提出的系统由三个卷积层组成，以克服小数据集的计算开销和过拟合问题。该系统已经用水稻植物的一种非常常见的疾病细菌性叶枯病进行了测试。

所提出的模型已经使用多个指标进行了评估，例如准确度、精度、I 类错误、II 类错误。这种新模型以 95% 的准确度和 97.5% 的高精度（2.3% 的 I 型错误和 7.7% 的 II 型错误）在最早阶段（接种后 48 小时内）预测疾病。对四种标准深度学习模型——VGG-16、VGG-19、Resnet50和Resnet101以及机器学习算法——线性回归和支持向量机进行了比较研究，以确定所提出模型的优越性。