PLANT DISEASE AND PEST CONTROL METHOD USING SPECTRAL REMOTE SENSING AND ARTIFICIAL INTELLIGENCE

Abstract

Disclosed herein is a plant disease and pest control method using aerial photography and spectral remote sensing technology to record and analyze crop orthographic images. After building meshes from a dense point cloud or 3D depth map, orthographic images are generated. Then, the numbers of crop pest insect, crop leaf infestation area, and the ratio between leaf and its farmland area are calculated using deep learning techniques. After that, the growth curve of the pest population is established via modeling techniques and a pest and disease prediction model is established to determine the optimized timing for pesticide spraying.

Claims

1. A plant disease and pest control method comprising the following steps of: (a) providing orthographic images of a plurality of spectral image files of a collection of crop leaves and establishing an image data set of the crop leaves; (b) building polygon meshes of the image files using a dense point cloud or 3D depth map; (c) collecting spectral reflectance of a range of spectral colors from the spectral image files to establish a spectral feature analysis; (d) calculating the area of the crop leaves and the total number of pixels of the crop leaves based the orthographic images; (e) determining an infestation map of the crop leaves and the total number of pixels of infested crop leaves using a hyperspectral image detection algorithm or a machine learning technique; (f) calculating the area of crop leaves and the total number of pixels of the crop leaves from the orthographic images; and (g) dividing the total number of pixels of the infested crop leaves by the total number of pixels of the crop leaves to determine an infested area of the crop leaves.

2. The plant disease and pest control method described of claim 1, further comprising establishing a pest and disease prediction model using a deep learning based semantic segmentation model to estimate the size of a pest population and devise pest and disease control measures.

3. The plant disease and pest control method described of claim 1, wherein the orthographic images of the plurality of spectral image files are obtained using an aerial camera drone.

4. The plant disease and pest control method described of claim 1, wherein the crops include, but are not limited to, lotus leaves, mangoes, pumpkins, and lychees.

5. The plant disease and pest control method described of claim 1, wherein the pests and diseases include, but are not limited to, prodenia litura larvae, cabbage leaf moth larvae, small leaf moths, scale insects, gall gnats, thrips, oriental fruit flies, pumpkin flies, and lychee stink bugs.

6. A method for determining the size and growth of an insect population comprising the following steps of: (a) hanging several sheets of pest glue traps such that they are spread out evenly in the field; (b) periodically retrieving and taking high resolution photos of the glue traps; and (c) detecting the insects and determining the total number of the insects using a customized deep learning based objection detection model.

7. The method for determining the size and growth of an insect population of claim 6, further comprising establishing a pest and disease prediction model using a deep learning based object detection model to estimate the size of a pest population and pest and disease control timing.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0017] FIG. 1 is a flowchart illustrating the overall technique in accordance with the present disclosure;

[0018] FIG. 2 is a graph illustrating a prediction model of small yellow thrips population and prediction of the timing of damage prevention;

[0019] FIG. 3 shows the detection and the calculation of small yellow thrips from the back of a leaf; and

[0020] FIG. 4 shows the detection and the calculation of small yellow thrips from a yellow pest glue trap in the field.

DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] Prediction Model of Small Yellow Thrips

[0022] As shown in FIG. 2, x axis denotes time and y axis denotes the accumulated infested area of lotus leaves. Through deep learning based semantic segmentation model, the ratio of the area of lotus leaves in a lotus field and the ratio of the pest infestation area are detected to evaluate the growth of the lotuses. Then, a pest and disease prediction model is established using a single day as the unit.

[0023] As shown in FIG. 2, the number of small yellow thrips started to multiply on the 88.sup.th day after planting the lotuses and the growth was exponential and reached its peak on the 115.sup.th day of planting.

[0024] As calculated using the regression model shown in FIG. 2, the recommended optimized timing for pesticide application is one to two days before the exponential growth of the small yellow thrips. Thus, pesticide was applied on the 98.sup.th day after planting in the control group. The number of small yellow thrips after treatment was a quarter of that in the case where no pesticide was used.

[0025] Detection and Calculation of Small Yellow Thrips on Back of Lotus Leaves

[0026] After increasing the resolution of out-of-focused areas using a super resolution network, the number of small yellow thrips on lotus leaves is calculated using a deep learning based object detection model.

[0027] As shown in FIG. 3, an image of a whole leave is first segmented into several partial images and insects are detected using a trained deep learning model. The detection results are then combined to determine the total number of insects. The detection rate of the detection model used for small yellow thrips on the back of lotus leaves could be as high as 96.14%.

[0028] Detection and Calculation of Small Yellow Thrips from Yellow Pest Glue Traps in the Field

[0029] The numbers of small yellow thrips on yellow pest glue traps are detected and calculated using deep learning based objection detection models to estimate the size of the pest population.

[0030] As shown in FIG. 4, the detection rates of small yellow thrips on yellow pest glue traps in the field under different deep learning models could reach between 89.99% and 93.34%.