Automated Soybean Crop Health Evaluation from UAV Images Using Patch-Level CNNs

Abstract: In this paper, we present a simple system to check soybean crop health using images taken from drones (UAVs). In normal farming, checking crop diseases takes a lot of time and effort, and sometimes small disease areas are missed. Because of this, farmers may face loss in yield. So, we tried to make an automatic system which can help in early detection of crop problems.

In our approach, UAV images are not used directly as a whole. Instead, we break each large image into many small parts called patches. This helps the model to focus more on plant areas and less on unwanted background like soil or shadows. For classification, we used a lightweight deep learning model called MobileNetV3-Small. It is not very heavy, so it works faster and can be used even on limited systems.

We also applied some basic data augmentation methods like flipping, rotation, and brightness change. This is done because in real fields, lighting and angles are not always the same. After predicting each patch as healthy or diseased, we combine all results to create a full field health map. This map helps to understand which areas are affected.

The results we got are quite good and show that the method works properly on different UAV images. The system is simple, fast, and can be useful for farmers to take quick decisions. It can also be extended to other crops in future.

Keywords: Soybean crop health, UAV images, drone-based agriculture, patch-level analysis, convolutional neural network, MobileNetV3, plant disease detection, precision agriculture, image classification, deep learning in agriculture