Diseases from Ginkgo biloba have brought great losses to medicine and the economy. Therefore, if the degree of disease can be automatically identified in Ginkgo biloba leaves, people will take appropriate measures to avoid losses in advance. Deep learning has made great achievements in plant disease identification and classification. For this paper, the convolution neural network model was used to classify the different degrees of ginkgo leaf disease. This study used the VGGNet-16 and Inception V3 models. After preprocessing and training 1322 original images under laboratory conditions and 2408 original images under field conditions, 98.44% accuracy was achieved under laboratory conditions and 92.19% under field conditions with the VGG model. The Inception V3 model achieved 92.3% accuracy under laboratory conditions and 93.2% under field conditions. Thus, the Inception V3 model structure was more suitable for field conditions. To our knowledge, there is very little research on the classification of different degrees of the same plant disease. The success of this study will have a significant impact on the prediction and early prevention of ginkgo leaf blight.

In order to solve the problem of data loss in sensor data collection, this paper took the stem moisture data of plants as the object, and compared the filling value of missing data in the same data segment with different data filling methods to verify the validity and accuracy of the stem water filling data of the LSTM (Long Short-Term Memory) model. This paper compared the accuracy of missing stem water data for plants under different data filling methods to solve the problem of data loss in sensor data collection. Original stem moisture data was selected from Lagerstroemia Indica which was planted in the Haidian District of Beijing in June 2017. Part of the data which treated as missing data was manually deleted. Interpolation methods, time series statistical methods, the RNN (Recurrent Neural Network), and LSTM neural network were used to fill in the missing part and the filling results were compared with the original data. The result shows that the LSTM has more accurate performance than the RNN. The error values of the bidirectional LSTM model are the smallest among several models. The error values of the bidirectional LSTM are much lower than other methods. The MAPE (mean absolute percent error) of the bidirectional LSTM model is 1.813%. After increasing the length of the training data, the results further proved the effectiveness of the model. Further, in order to solve the problem of one-dimensional filling error accumulation, the LSTM model is used to conduct the multi-dimensional filling experiment with environmental data. After comparing the filling results of different environmental parameters, three environmental parameters of air humidity, photosynthetic active radiation, and soil temperature were selected as input. The results show that the multi-dimensional filling can greatly extend the sequence length while maintaining the accuracy, and make up for the defect that the one-dimensional filling accumulates errors with the increase of the sequence. The minimum MAPE of multidimensional filling is 1.499%. In conclusion, the data filling method based on LSTM neural network has a great advantage in filling the long-lost time series data which would provide a new idea for data filling.

Diverse sensing techniques have been developed and combined with machine learning method for forest fire detection, but none of them referred to identifying smoldering and flaming combustion phases. This study attempts to real-time identify different combustion phases using a developed wireless sensor network (WSN)-based multi-sensor system and artificial neural network (ANN). Sensors (CO, CO₂, smoke, air temperature and relative humidity) were integrated into one node of WSN. An experiment was conducted using burning materials from residual of forest to test responses of each node under no, smoldering-dominated and flaming-dominated combustion conditions. The results showed that the five sensors have reasonable responses to artificial forest fire. To reduce cost of the nodes, smoke, CO₂ and temperature sensors were chiefly selected through correlation analysis. For achieving higher identification rate, an ANN model was built and trained with inputs of four sensor groups: smoke; smoke and CO₂; smoke and temperature; smoke, CO₂ and temperature. The model test results showed that multi-sensor input yielded higher predicting accuracy (≥82.5%) than single-sensor input (50.9%–92.5%). Based on these, it is possible to reduce the cost with a relatively high fire identification rate and potential application of the system can be tested in future under real forest condition.

The detection and recognition of unstructured roads in forest environments are critical for smart forestry technology. Forest roads lack effective reference objects and manual signs and have high degrees of nonlinearity and uncertainty, which pose severe challenges to forest engineering vehicles. This research aims to improve the automation and intelligence of forestry engineering and proposes an unstructured road detection and recognition method based on a combination of image processing and 2D lidar detection. This method uses the “improved SEEDS + Support Vector Machine (SVM)” strategy to quickly classify and recognize the road area in the image. Combined with the remapping of 2D lidar point cloud data on the image, the actual navigation requirements of forest unmanned navigation vehicles were fully considered, and road model construction based on the vehicle coordinate system was achieved. The algorithm was transplanted to a self-built intelligent navigation platform to verify its feasibility and effectiveness. The experimental results show that under low-speed conditions, the system can meet the real-time requirements of processing data at an average of 10 frames/s. For the centerline of the road model, the matching error between the image and lidar is no more than 0.119 m. The algorithm can provide effective support for the identification of unstructured roads in forest areas. This technology has important application value for forestry engineering vehicles in autonomous inspection and spraying, nursery stock harvesting, skidding, and transportation.

Based on Internet-of-Things and multi-sensor technology, an intelligent wireless monitoring system was developed to obtain field ecological parameters and provide forest fire warning in real-time. The GPRS and China’s Beidou satellite communication were selectively used for date transmission in the field with weak cell phone signals. This monitoring system is mainly composed of several field ecological monitoring stations, a cloud server, and online system software. Atmosphere, soil, sunlight and plant parameters of different regions are obtained real-time by sensors stably and reliably. This system has functions such as field ecological data storage, dynamic query, report generation, and data analysis. As an example of typical application, the forest fire weather grade, which was supplemented with the litter layer soil humidity, was calculated to realize the early warning of the local forest fire in this system through continuous experiments at Beijing Jiufeng National Forest Park from March to May 2017 and Inner Mongolia from March to June 2018. The success ratios of data transmission through Beidou satellite were 98.57%, 99.43%, 99.59%, and 98.85%, respectively, in Beijing, and through GPRS were 99.89% and 99.90% in Inner Mongolia. Long-term real-time field ecological monitoring and forest fire warning were successfully realized. This system can be widely used for big data field acquisition and analysis in forest and agriculture regions.

This article presents a non-destructive methodology to determine the modulus of elasticity (MOE) in static bending of wood through the use of near-infrared (NIR) spectroscopy. Wood specimens were obtained from Quercus mongolica growing in Northeast of China. The NIR spectra of specimens were acquired by using a one-chip NIR fiber optic spectrometer whose spectral range was 900~1900 nm. The raw spectra of specimens were pretreated by multiplication scatter correlation and Savitzky-Golay smoothing and differentiation filter. To reduce the dimensions of data and complexity of modeling, the synergy interval partial least squares and successive projections algorithm were applied to extract the characteristic wavelengths, which had closing relevance with the MOE of wood, and five characteristic wavelengths were selected from full 117 variables of a spectrum. Taking the characteristic wavelengths as input values, partial least square regression (PLSR) and the propagation neural network (BPNN) were implemented to establish calibration models. The predictive ability of the models was estimated by the coefficient of determination (r_p) and the root mean square error of prediction (RMSEP) and in the prediction set. In comparison with the predicted results of the models, BPNN performed better results with the higher r_p of 0.91 and lower RMSEP of 0.76. The results indicate that it is feasible to accurately determine the MOE of wood by using the NIR spectroscopy technique.